null  null
MIXED-SIGNAL AND DSP
DESIGN TECHNIQUES
INTRODUCTION
SECTION 1
SAMPLED DATA SYSTEMS
SECTION 2
ADCs FOR DSP APPLICATIONS
SECTION 3
DACs FOR DSP APPLICATIONS
SECTION 4
FAST FOURIER TRANSFORMS
SECTION 5
DIGITAL FILTERS
SECTION 6
DSP HARDWARE
SECTION 7
INTERFACING TO DSPs
SECTION 8
DSP APPLICATIONS
SECTION 9
HARDWARE DESIGN
TECHNIQUES
SECTION 10
INDEX
ANALOG DEVICES TECHNICAL REFERENCE BOOKS
PUBLISHED BY PRENTICE HALL
Analog-Digital Conversion Handbook
Digital Signal Processing Applications Using the ADSP-2100 Family
(Volume 1:1992, Volume 2:1994)
Digital Signal Processing in VLSI
DSP Laboratory Experiments Using the ADSP-2101
ADSP-2100 Family User's Manual
PUBLISHED BY ANALOG DEVICES
Practical Design Techniques for Sensor Signal Conditioning
Practical Design Techniques for Power and Thermal Management
High Speed Design Techniques
Practical Analog Design Techniques
Linear Design Seminar
ADSP-21000 Family Applications Handbook
System Applications Guide
Amplifier Applications Guide
Nonlinear Circuits Handbook
Transducer Interfacing Handbook
Synchro & Resolver Conversion
THE BEST OF Analog Dialogue, 1967-1991
HOW TO GET INFORMATION FROM ANALOG DEVICES
Analog Devices publishes data sheets and a host of other technical literature
supporting our products and technologies. Follow the instructions below for
worldwide access to this information.
FOR DATA SHEETS
U.S.A. and Canada
■
Fax Retrieval. Telephone number 800-446-6212. Call this number
and use a faxcode corresponding to the data sheet of your choice
for a fax-on-demand through our automated AnalogFax™ system.
Data sheets are available 7 days a week, 24 hours a day.
Product/faxcode cross reference listings are available by calling
the above number and following the prompts. There is a short
index with just part numbers, faxcodes, page count and revision
for each data sheet (Prompt # 28). There is also a longer index sorted by
product type with short descriptions (Prompt #29).
■
World Wide Web and Internet. Our address is http://www.analog.com.
Use the browser of your choice and follow the prompts. We also
provide extensive DSP literature support on an Internet FTP site. Type
ftp:// ftp.analog.com or ftp 137.71.23.11. Log in as anonymous using your
e-mail address for your password.
■
Analog Devices Literature Distribution Center. Call 800-262-5643 and
select option two from the voice prompts, or call 781-329-4700 for
direct access, or fax your request to 508-894-5114.
■
Analog Devices Southeast Asia Literature Distribution Centre. Fax
requests to 65-746-9115. Email address is
[email protected]
Europe and Israel
■
World Wide Web. Our address is http://www.analog.com. use the
browser of your choice and follow the prompts.
■
Analog Devices Sales Offices. Call your local sales office and request
a data sheet. A Worldwide Sales Directory including telephone listings
is on pp. 347-348 of the 1999 Winter Short Form Designers' Guide.
■
DSP Support Center. Fax requests to **49-89-76903-307 or e-mail
[email protected]
India
■
Call 91-80-526-3606 or fax 91-80-526-3713 and request the data
sheet of interest.
Other Locations
■
World Wide Web. Our address is http://www.analog.com. Use the
browser of your choice and follow the prompts.
■
Analog Devices Sales Offices. Call your local sales office and request
a data sheet. A Worldwide Sales Directory including telephone numbers
is listed on the back cover of the 1997 Short Form Designers' Guide.
TECHNICAL SUPPORT AND CUSTOMER SERVICE
■
In the U.S.A. and Canada, call 800-ANALOGD, (800-262-5643).
For technical support on all products, select option one, then select
the product area of interest. For price and delivery, select option three.
For literature and samples, select option two.
Non-800 Number: 781-937-1428.
MIXED-SIGNAL AND DSP
DESIGN TECHNIQUES
a
ACKNOWLEDGMENTS
Thanks are due the many technical staff members of Analog Devices in Engineering
and Marketing who provided invaluable inputs during this project. Particular credit is due
the individual authors whose names appear at the beginning of their material.
Special thanks go to Wes Freeman, Ed Grokulsky, Bill Chestnut, Dan King, Greg
Geerling, Ken Waurin, Steve Cox, and Colin Duggan for reviewing the material for
content and accuracy.
Judith Douville compiled the index.
Walt Kester
2000
Copyright  2000 by Analog Devices, Inc.
Printed in the United States of America
All rights reserved. This book, or parts thereof, must not be reproduced in any form
without permission of the copyright owner.
Information furnished by Analog Devices, Inc., is believed to be accurate and reliable.
However, no responsibility is assumed by Analog Devices, Inc., for its use.
Analog Devices, Inc., makes no representation that the interconnections of its circuits as
described herein will not infringe on existing or future patent rights, nor do the
descriptions contained herein imply the granting of licenses to make, use, or sell
equipment constructed in accordance therewith.
Specifications are subject to change without notice.
ISBN-0-916550-23-0
MIXED-SIGNAL AND DSP
DESIGN TECHNIQUES
SECTION 1
INTRODUCTION
SECTION 2
SAMPLED DATA SYSTEMS
■
Discrete Time Sampling of Analog Signals
■
ADC and DAC Static Transfer Functions and DC Errors
■
AC Errors in Data Converters
■
DAC Dynamic Performance
SECTION 3
ADCs FOR DSP APPLICATIONS
■
Successive Approximation ADCs
■
Sigma-Delta ADCs
■
Flash Converters
■
Subranging (Pipelined) ADCs
■
Bit-Per-Stage (Serial, or Ripple) ADCs
SECTION 4
DACs FOR DSP APPLICATIONS
■
DAC Structures
■
Low Distortion DAC Architectures
■
DAC Logic
■
Sigma-Delta DACs
■
Direct Digital Synthesis (DDS)
SECTION 5
FAST FOURIER TRANSFORMS
■
The Discrete Fourier Transform
■
The Fast Fourier Transform
■
FFT Hardware Implementation and Benchmarks
■
DSP Requirements for Real Time FFT Applications
■
Spectral Leakage and Windowing
SECTION 6
DIGITAL FILTERS
■
Finite Impulse Response (FIR) Filters
■
Infinite Impulse Response (IIR) Filters
■
Multirate Filters
■
Adaptive Filters
SECTION 7
DSP HARDWARE
■
Microcontrollers, Microprocessors, and Digital Signal
Processors (DSPs)
■
DSP Requirements
■
ADSP-21xx 16-Bit Fixed-Point DSP Core
■
Fixed-Point Versus Floating Point
■
ADI SHARC® Floating Point DSPs
■
ADSP-2116x Single-Instruction, Multiple Data (SIMD)
Core Architecture
■
TigerSHARC™: The ADSP-TS001 Static Superscalar
DSP
■
DSP Benchmarks
■
DSP Evaluation and Development Tools
SECTION 8
INTERFACING TO DSPs
■
Parallel Interfacing to DSP Processors: Reading Data
From Memory-Mapped Peripheral ADCs
■
Parallel Interfacing to DSP Processors: Writing Data to
Memory-Mapped DACs
■
Serial Interfacing to DSP Processors
■
Interfacing I/O Ports, Analog Front Ends, and Codecs to
DSPs
■
DSP System Interface
SECTION 9
DSP APPLICATIONS
■
High Performance Modems for Plain Old Telephone
Service (POTS)
■
Remote Access Server (RAS) Modems
■
ADSL (Assymetric Digital Subscriber Line)
■
Digital Cellular Telephones
■
GSM Handset Using SoftFone™ Baseband Processor
and Othello™ Radio
■
Analog Cellular Basestations
■
Digital Cellular Basestations
■
Motor Control
■
Codecs and DSPs in Voiceband and Audio Applications
■
A Sigma-Delta ADC with Programmable Digital Filter
SECTION 10
HARDWARE DESIGN TECHNIQUES
■
Low Voltage Interfaces
■
Grounding in Mixed Signal Systems
■
Digital Isolation Techniques
■
Power Supply Noise Reduction and Filtering
■
Dealing with High Speed Logic
INDEX
MIXED-SIGNAL AND DSP
DESIGN TECHNIQUES
INTRODUCTION
SECTION 1
SAMPLED DATA SYSTEMS
SECTION 2
ADCs FOR DSP APPLICATIONS
SECTION 3
DACs FOR DSP APPLICATIONS
SECTION 4
FAST FOURIER TRANSFORMS
SECTION 5
DIGITAL FILTERS
SECTION 6
DSP HARDWARE
SECTION 7
INTERFACING TO DSPs
SECTION 8
DSP APPLICATIONS
SECTION 9
HARDWARE DESIGN
TECHNIQUES
SECTION 10
INDEX
INTRODUCTION
SECTION 1
INTRODUCTION
1.a
INTRODUCTION
1.b
INTRODUCTION
SECTION 1
INTRODUCTION
Walt Kester
ORIGINS OF REAL-WORLD SIGNALS AND THEIR UNITS
OF MEASUREMENT
In this book, we will primarily be dealing with the processing of real-world signals
using both analog and digital techniques. Before starting, however, let's look at a
few key concepts and definitions required to lay the groundwork for things to come.
Webster's New Collegiate Dictionary defines a signal as "A detectable (or
measurable) physical quantity or impulse (as voltage, current, or magnetic field
strength) by which messages or information can be transmitted." Key to this
definition are the words: detectable, physical quantity, and information.
SIGNAL CHARACTERISTICS
■ Signal Characteristics
◆ Signals are Physical Quantities
◆ Signals are Measurable
◆ Signals Contain Information
◆ All Signals are Analog
■ Units of Measurement
◆ Temperature: °C
◆ Pressure: Newtons/m2
◆ Mass: kg
◆ Voltage: Volts
◆ Current: Amps
◆ Power: Watts
Figure 1.1
By their very nature, signals are analog , whether DC, AC, digital levels, or pulses.
It is customary, however, to differentiate between analog and digital signals in the
following manner: Analog (or real-world) variables in nature include all measurable
physical quantities. In this book, analog signals are generally limited to electrical
variables, their rates of change, and their associated energy or power levels. Sensors
are used to convert other physical quantities (temperature, pressure, etc.) to
electrical signals. The entire subject of signal conditioning deals with preparing
real-world signals for processing and includes such topics as sensors (temperature
1.1
INTRODUCTION
and pressure, for example), isolation and instrumentation amplifiers, etc. (see
Reference 1).
Some signals result in response to other signals. A good example is the returned
signal from a radar or ultrasound imaging system, both of which result from a
known transmitted signal.
On the other hand, there is another classification of signals, called digital, where
the actual signal has been conditioned and formatted into a digit. These digital
signals may or may not be related to real-world analog variables. Examples include
the data transmitted over local area networks (LANs) or other high speed networks.
In the specific case of Digital Signal Processing (DSP), the analog signal is
converted into binary form by a device known as an analog-to-digital converter
(ADC). The output of the ADC is a binary representation of the analog signal and is
manipulated arithmetically by the Digital Signal Processor. After processing, the
information obtained from the signal may be converted back into analog form using
a digital-to-analog converter (DAC).
Another key concept embodied in the definition of signal is that there is some kind
of information contained in the signal. This leads us to the key reason for processing
real-world analog signals: the extraction of information.
REASONS FOR PROCESSING REAL-WORLD SIGNALS
The primary reason for processing real-world signals is to extract information from
them. This information normally exists in the form of signal amplitude (absolute or
relative), frequency or spectral content, phase, or timing relationships with respect
to other signals. Once the desired information is extracted from the signal, it may be
used in a number of ways.
In some cases, it may be desirable to reformat the information contained in a signal.
This would be the case in the transmission of a voice signal over a frequency
division multiple access (FDMA) telephone system. In this case, analog techniques
are used to "stack" voice channels in the frequency spectrum for transmission via
microwave relay, coaxial cable, or fiber. In the case of a digital transmission link,
the analog voice information is first converted into digital using an ADC. The digital
information representing the individual voice channels is multiplexed in time (time
division multiple access, or TDMA) and transmitted over a serial digital
transmission link (as in the T-Carrier system).
Another requirement for signal processing is to compress the frequency content of
the signal (without losing significant information) then format and transmit the
information at lower data rates, thereby achieving a reduction in required channel
bandwidth. High speed modems and adaptive pulse code modulation systems
(ADPCM) make extensive use of data reduction algorithms, as do digital mobile
radio systems, MPEG recording and playback, and High Definition Television
(HDTV).
1.2
INTRODUCTION
Industrial data acquisition and control systems make use of information extracted
from sensors to develop appropriate feedback signals which in turn control the
process itself. Note that these systems require both ADCs and DACs as well as
sensors, signal conditioners, and the DSP (or microcontroller). Analog Devices offers
a family of MicroConverters™ which include precision analog conditioning circuitry,
ADCs, DACs, microcontroller, and FLASH memory all on a single chip.
In some cases, the signal containing the information is buried in noise, and the
primary objective is signal recovery. Techniques such as filtering, auto-correlation,
convolution, etc. are often used to accomplish this task in both the analog and
digital domains.
REASONS FOR SIGNAL PROCESSING
■ Extract Information About The Signal (Amplitude, Phase,
Frequency, Spectral Content, Timing Relationships)
■ Reformat the Signal (FDMA, TDMA, CDMA Telephony)
■ Compress Data (Modems, Cellular Telephone, HDTV, MPEG)
■ Generate Feedback Control Signal (Industrial Process Control)
■ Extract Signal From Noise (Filtering, Autocorrelation,
Convolution)
■ Capture and Store Signal in Digital Format for Analysis (FFT
Techniques)
Figure 1.2
GENERATION OF REAL-WORLD SIGNALS
In most of the above examples (the ones requiring DSP techniques), both ADCs and
DACs are required. In some cases, however, only DACs are required where realworld analog signals may be generated directly using DSP and DACs. Video raster
scan display systems are a good example. The digitally generated signal drives a
video or RAMDAC. Another example is artificially synthesized music and speech.
In reality, however, the real-world analog signals generated using purely digital
techniques do rely on information previously derived from the real-world equivalent
analog signals. In display systems, the data from the display must convey the
appropriate information to the operator. In synthesized audio systems, the
statistical properties of the sounds being generated have been previously derived
using extensive DSP analysis (i.e.,sound source, microphone, preamp, ADC, etc.).
1.3
INTRODUCTION
METHODS AND TECHNOLOGIES AVAILABLE FOR
PROCESSING REAL-WORLD SIGNALS
Signals may be processed using analog techniques (analog signal processing, or
ASP), digital techniques (digital signal processing, or DSP), or a combination of
analog and digital techniques (mixed signal processing, or MSP). In some cases, the
choice of techniques is clear; in others, there is no clear cut choice, and second-order
considerations may be used to make the final decision.
With respect to DSP, the factor that distinguishes it from traditional computer
analysis of data is its speed and efficiency in performing sophisticated digital
processing functions such as filtering, FFT analysis, and data compression in real
time.
The term mixed signal processing implies that both analog and digital processing is
done as part of the system. The system may be implemented in the form of a printed
circuit board, hybrid microcircuit, or a single integrated circuit chip. In the context
of this broad definition, ADCs and DACs are considered to be mixed signal
processors, since both analog and digital functions are implemented in each. Recent
advances in Very Large Scale Integration (VLSI) processing technology allow
complex digital processing as well as analog processing to be performed on the same
chip. The very nature of DSP itself implies that these functions can be performed in
real-time.
ANALOG VERSUS DIGITAL SIGNAL PROCESSING
Today's engineer faces a challenge in selecting the proper mix of analog and digital
techniques to solve the signal processing task at hand. It is impossible to process
real-world analog signals using purely digital techniques, since all sensors
(microphones, thermocouples, strain gages, microphones, piezoelectric crystals, disk
drive heads, etc.) are analog sensors. Therefore, some sort of signal conditioning
circuitry is required in order to prepare the sensor output for further signal
processing, whether it be analog or digital. Signal conditioning circuits are, in
reality, analog signal processors, performing such functions as multiplication (gain),
isolation (instrumentation amplifiers and isolation amplifiers), detection in the
presence of noise (high common-mode instrumentation amplifiers, line drivers, and
line receivers), dynamic range compression (log amps, LOGDACs, and
programmable gain amplifiers), and filtering (both passive and active).
Several methods of accomplishing signal processing are shown in Figure 1.3. The
top portion of the figure shows the purely analog approach. The latter parts of the
figure show the DSP approach. Note that once the decision has been made to use
DSP techniques, the next decision must be where to place the ADC in the signal
path.
1.4
INTRODUCTION
ANALOG AND DIGITAL SIGNAL PROCESSING OPTIONS
REAL WORLD
SENSOR
ANALOG
CONDITIONING
SENSOR
ANALOG
CONDITIONING
SENSOR
SENSOR
SIGNAL PROCESSING
ANALOG
SIGNAL
PROCESSING
ADC
DSP
DAC
ADC AND
CONDITIONING
DSP
DAC
ADC
CODEC OR AFE
DAC
(ANALOG FRONT END)
DSP
Figure 1.3
In general, as the ADC is moved closer to the actual sensor, more of the analog
signal conditioning burden is now placed on the ADC. The added ADC complexity
may take the form of increased sampling rate, wider dynamic range, higher
resolution, input noise rejection, input filtering and programmable gain amplifiers
(PGAs), on-chip voltage references, etc., all of which add functionality and simplify
the system. With today’s high-resolution/high sampling rate data converter
technology, significant progress has been made in integrating more and more of the
conditioning circuitry within the ADC/DAC itself. In the measurement area, for
instance, 24-bit ADCs are available with built-in programmable gain amplifiers
(PGAs) which allow fullscale bridge signals of 10mV to be digitized directly with no
further conditioning (e.g. AD773x-series). At voiceband and audio frequencies,
complete coder-decoders (Codecs – or Analog Front Ends) are available which have
sufficient on-chip analog circuitry to minimize the requirements for external
conditioning components (AD1819B and AD73322). At video speeds, analog front
ends are also available for such applications as CCD image processing and others
(e.g., AD9814, AD9816, and the AD984x series).
A PRACTICAL EXAMPLE
As a practical example of the power of DSP, consider the comparison between an
analog and a digital lowpass filter, each with a cutoff frequency of 1kHz. The digital
filter is implemented in a typical sampled data system shown in Figure 1.4. Note
that there are several implicit requirements in the diagram. First, it is assumed
that an ADC/DAC combination is available with sufficient sampling frequency,
1.5
INTRODUCTION
resolution, and dynamic range to accurately process the signal. Second, the DSP
must be fast enough to complete all its calculations within the sampling interval,
1/fs. Third, analog filters are still required at the ADC input and DAC output for
antialiasing and anti-imaging, but the performance demands are not as great.
Assuming these conditions have been met, the following offers a comparison
between the digital and analog filters.
DIGITAL FILTER
H(f)
1kHz
t
t
f
ANALOG
ANTIALIASING
FILTER
x(n)
ADC
DIGITAL
LOWPASS
FILTER
fs = 10kSPS
y(n)
DAC
ANALOG
ANTI-IMAGING
FILTER
y(n) MUST BE COMPUTED
DURING THE SAMPLING
INTERVAL, 1 / fs
Figure 1.4
The required cutoff frequency of both filters is 1kHz. The analog filter is realized as
a 6-pole Chebyshev Type 1 filter (ripple in passband, no ripple in stopband), and the
response is shown in Figure 1.5. In practice, this filter would probably be realized
using three 2-pole stages, each of which requires an op amp, and several resistors
and capacitors. Modern filter design CAD packages make the 6-pole design
relatively straightforward, but maintaining the 0.5dB ripple specification requires
accurate component selection and matching.
On the other hand, the 129-tap digital FIR filter shown has only 0.002dB passband
ripple, linear phase, and a much sharper roll off. In fact, it could not be realized
using analog techniques! Another obvious advantage is that the digital filter
requires no component matching, and it is not sensitive to drift since the clock
frequencies are crystal controlled. The 129-tap filter requires 129 multiplyaccumulates (MAC) in order to compute an output sample. This processing must be
completed within the sampling interval, 1/fs, in order to maintain real-time
operation. In this example, the sampling frequency is 10kSPS, therefore 100µs is
available for processing, assuming no significant additional overhead requirement.
The ADSP-21xx-family of DSPs can complete the entire multiply-accumulate
1.6
INTRODUCTION
process (and other functions necessary for the filter) in a single instruction cycle.
Therefore, a 129-tap filter requires that the instruction rate be greater than
129/100µs = 1.3 million instructions per second (MIPS). DSPs are available with
instruction rates much greater than this, so the DSP certainly is not the limiting
factor in this application. The ADSP-218x 16-bit fixed point series offers instruction
rates up to 75MIPS.
The assembly language code to implement the filter on the ADSP-21xx-family of
DSPs is shown in Figure 1.6. Note that the actual lines of operating code have been
marked with arrows; the rest are comments.
ANALOG VERSUS DIGITAL FILTER
FREQUENCY RESPONSE COMPARISON
dB
ANALOG FILTER
DIGITAL FILTER
Chebyshev Type 1
6 Pole, 0.5dB Ripple
FIR, 129-Tap, 0.002dB Ripple,
Linear Phase, fs = 10kSPS
dB
0
0
–20
–20
–40
–40
–60
–60
–80
–80
–100
–100
0
1
2
3
4
5
FREQUENCY (kHz)
0
1
2
3
4
5
FREQUENCY (kHz)
Figure 1.5
In a practical application, there are certainly many other factors to consider when
evaluating analog versus digital filters, or analog versus digital signal processing in
general. Most modern signal processing systems use a combination of analog and
digital techniques in order to accomplish the desired function and take advantage of
the best of both the analog and the digital world.
1.7
INTRODUCTION
ADSP-21XX FIR FILTER ASSEMBLY CODE
(SINGLE PRECISION)
.MODULE
{
fir_sub;
FIR Filter Subroutine
Calling Parameters
I0 --> Oldest input data value in delay line
I4 --> Beginning of filter coefficient table
L0 = Filter length (N)
L4 = Filter length (N)
M1,M5 = 1
CNTR = Filter length - 1 (N-1)
Return Values
MR1 = Sum of products (rounded and saturated)
I0 --> Oldest input data value in delay line
I4 --> Beginning of filter coefficient table
Altered Registers
MX0,MY0,MR
Computation Time
(N - 1) + 6 cycles = N + 5 cycles
All coefficients are assumed to be in 1.15 format. }
.ENTRY
fir:
fir;
MR=0, MX0=DM(I0,M1), MY0=PM(I4,M5);
CNTR = N-1;
DO convolution UNTIL CE;
MR=MR+MX0*MY0(SS), MX0=DM(I0,M1), MY0=PM(I4,M5);
MR=MR+MX0*MY0(RND);
IF MV SAT MR;
RTS;
convolution:
.ENDMOD;
Figure 1.6
REAL-TIME SIGNAL PROCESSING
■ Digital Signal Processing;
◆ ADC / DAC Sampling Frequency Limits Signal Bandwidth
● (Don't forget Nyquist!)
◆ ADC / DAC Resolution / Performance Limits Signal Dynamic
Range
◆ DSP Processor Speed Limits Amount of Digital Processing
Available, Because:
● All DSP Computations Must Be Completed During the
Sampling Interval, 1 / fs, for Real-Time Operation!
■ Don't Forget Analog Signal Processing
◆ High Frequency / RF Filtering, Modulation, Demodulation
◆ Analog Anti-Aliasing and Reconstruction Filters with ADCs
and DACs
◆ Where COMMON SENSE and Economics Dictate!
Figure 1.7
1.8
INTRODUCTION
REFERENCES
1.
Practical Design Techniques for Sensor Signal Conditioning,
Analog Devices, 1998.
2.
Daniel H. Sheingold, Editor, Transducer Interfacing Handbook,
Analog Devices, Inc., 1972.
3.
Richard J. Higgins, Digital Signal Processing in VLSI, Prentice-Hall,
1990.
1.9
INTRODUCTION
1.10
SAMPLED DATA SYSTEMS
SECTION 2
SAMPLED DATA SYSTEMS
■
Discrete Time Sampling of Analog Signals
■
ADC and DAC Static Transfer Functions and DC Errors
■
AC Errors in Data Converters
■
DAC Dynamic Performance
2.a
SAMPLED DATA SYSTEMS
2.b
SAMPLED DATA SYSTEMS
SECTION 2
SAMPLED DATA SYSTEMS
Walt Kester, James Bryant
INTRODUCTION
A block diagram of a typical sampled data DSP system is shown in Figure 2.1. Prior
to the actual analog-to-digital conversion, the analog signal usually passes through
some sort of signal conditioning circuitry which performs such functions as
amplification, attenuation, and filtering. The lowpass/bandpass filter is required to
remove unwanted signals outside the bandwidth of interest and prevent aliasing.
FUNDAMENTAL SAMPLED DATA SYSTEM
fs
fa
LPF
OR
BPF
N-BIT
ADC
AMPLITUDE
QUANTIZATION
fs
LPF
OR
BPF
N-BIT
DAC
DSP
DISCRETE
TIME SAMPLING
fa
1
ts= f
s
t
Figure 2.1
The system shown in Figure 2.1 is a real-time system, i.e., the signal to the ADC is
continuously sampled at a rate equal to fs, and the ADC presents a new sample to
the DSP at this rate. In order to maintain real-time operation, the DSP must
perform all its required computation within the sampling interval, 1/fs, and present
an output sample to the DAC before arrival of the next sample from the ADC. An
example of a typical DSP function would be a digital filter.
In the case of FFT analysis, a block of data is first transferred to the DSP memory.
The FFT is calculated at the same time a new block of data is transferred into the
memory, in order to maintain real-time operation. The DSP must calculate the FFT
during the data transfer interval so it will be ready to process the next block of data.
2.1
SAMPLED DATA SYSTEMS
Note that the DAC is required only if the DSP data must be converted back into an
analog signal (as would be the case in a voiceband or audio application, for
example). There are many applications where the signal remains entirely in digital
format after the initial A/D conversion. Similarly, there are applications where the
DSP is solely responsible for generating the signal to the DAC, such as in CD player
electronics. If a DAC is used, it must be followed by an analog anti-imaging filter to
remove the image frequencies.
There are two key concepts involved in the actual analog-to-digital and digital-toanalog conversion process: discrete time sampling and finite amplitude resolution
due to quantization. An understanding of these concepts is vital to DSP
applications.
DISCRETE TIME SAMPLING OF ANALOG SIGNALS
The concepts of discrete time sampling and quantization of an analog signal are
shown in Figure 2.1. The continuous analog data must is sampled at discrete
intervals, ts = 1/fs which must be carefully chosen to insure an accurate
representation of the original analog signal. It is clear that the more samples
taken (faster sampling rates), the more accurate the digital representation, but if
fewer samples are taken (lower sampling rates), a point is reached where critical
information about the signal is actually lost. This leads us to the statement of
Nyquist's criteria given in Figure 2.2.
NYQUIST'S CRITERIA
■ A signal with a bandwidth fa must be sampled at a rate fs > 2fa or
information about the signal will be lost.
■ Aliasing occurs whenever fs < 2fa
■ The concept of aliasing is widely used in communications
applications such as direct IF-to-digital conversion.
Figure 2.2
Simply stated, the Nyquist Criteria requires that the sampling frequency be at least
twice the signal bandwidth, or information about the signal will be lost. If the
sampling frequency is less than twice the analog signal bandwidth, a phenomena
known as aliasing will occur.
In order to understand the implications of aliasing in both the time and frequency
domain, first consider case of a time domain representation of a single tone
sinewave sampled as shown in Figure 2.3. In this example, the sampling frequency
fs is only slightly more than the analog input frequency fa, and the Nyquist criteria
is violated. Notice that the pattern of the actual samples produces an aliased
sinewave at a lower frequency equal to fs – fa.
2.2
SAMPLED DATA SYSTEMS
The corresponding frequency domain representation of this scenario is shown in
Figure 2.4B. Now consider the case of a single frequency sinewave of frequency fa
sampled at a frequency fs by an ideal impulse sampler (see Figure 2.4A). Also
assume that fs > 2fa as shown. The frequency-domain output of the sampler shows
aliases or images of the original signal around every multiple of fs, i.e. at
frequencies equal to |± Kfs ± fa|, K = 1, 2, 3, 4, .....
ALIASING IN THE TIME DOMAIN
ALIASED SIGNAL = fs – fa
INPUT = fa
1
fs
t
NOTE: fa IS SLIGHTLY LESS THAN fs
Figure 2.3
ANALOG SIGNAL fa SAMPLED @ fs USING IDEAL SAMPLER
HAS IMAGES (ALIASES) AT |±Kfs ±fa|, K = 1, 2, 3, . . .
A
fa
1.5fs
ZONE 2
ZONE 1
ZONE 3
fa
I
0.5fs
I
I
fs
0.5fs
B
I
I
ZONE 4
I
fs
2fs
I
I
1.5fs
2fs
Figure 2.4
2.3
SAMPLED DATA SYSTEMS
The Nyquist bandwidth is defined to be the frequency spectrum from DC to fs/2. The
frequency spectrum is divided into an infinite number of Nyquist zones, each having
a width equal to 0.5fs as shown. In practice, the ideal sampler is replaced by an
ADC followed by an FFT processor. The FFT processor only provides an output from
DC to fs/2, i.e., the signals or aliases which appear in the first Nyquist zone.
Now consider the case of a signal which is outside the first Nyquist zone (Figure
2.4B). The signal frequency is only slightly less than the sampling frequency,
corresponding to the condition shown in the time domain representation in Figure
2.3. Notice that even though the signal is outside the first Nyquist zone, its image
(or alias), fs–fa, falls inside. Returning to Figure 2.4A, it is clear that if an
unwanted signal appears at any of the image frequencies of fa, it will also occur at
fa, thereby producing a spurious frequency component in the first Nyquist zone.
This is similar to the analog mixing process and implies that some filtering ahead of
the sampler (or ADC) is required to remove frequency components which are outside
the Nyquist bandwidth, but whose aliased components fall inside it. The filter
performance will depend on how close the out-of-band signal is to fs/2 and the
amount of attenuation required.
Baseband Antialiasing Filters
Baseband sampling implies that the signal to be sampled lies in the first Nyquist
zone. It is important to note that with no input filtering at the input of the ideal
sampler, any frequency component (either signal or noise) that falls outside the
Nyquist bandwidth in any Nyquist zone will be aliased back into the first Nyquist
zone. For this reason, an antialiasing filter is used in almost all sampling ADC
applications to remove these unwanted signals.
Properly specifying the antialiasing filter is important. The first step is to know the
characteristics of the signal being sampled. Assume that the highest frequency of
interest is fa. The antialiasing filter passes signals from DC to fa while attenuating
signals above fa.
Assume that the corner frequency of the filter is chosen to be equal to fa. The effect
of the finite transition from minimum to maximum attenuation on system dynamic
range is illustrated in Figure 2.5A.
Assume that the input signal has fullscale components well above the maximum
frequency of interest, fa. The diagram shows how fullscale frequency components
above fs – fa are aliased back into the bandwidth DC to fa. These aliased
components are indistinguishable from actual signals and therefore limit the
dynamic range to the value on the diagram which is shown as DR.
Some texts recommend specifying the antialiasing filter with respect to the Nyquist
frequency, fs/2, but this assumes that the signal bandwidth of interest extends from
DC to fs/2 which is rarely the case. In the example shown in Figure 2.5A, the
aliased components between fa and fs/2 are not of interest and do not limit the
dynamic range.
2.4
SAMPLED DATA SYSTEMS
The antialiasing filter transition band is therefore determined by the corner
frequency fa, the stopband frequency fs – fa, and the desired stopband attenuation,
DR. The required system dynamic range is chosen based on the requirement for
signal fidelity.
OVERSAMPLING RELAXES REQUIREMENTS
ON BASEBAND ANTIALIASING FILTER
fa
B
A
fa
fs - fa
Kfs - fa
DR
fs
2
fs
Kfs
2
STOPBAND ATTENUATION = DR
TRANSITION BAND: fa to fs - fa
CORNER FREQUENCY: fa
Kfs
STOPBAND ATTENUATION = DR
TRANSITION BAND: fa to Kfs - fa
CORNER FREQUENCY: fa
Figure 2.5
Filters become more complex as the transition band becomes sharper, all other
things being equal. For instance, a Butterworth filter gives 6dB attenuation per
octave for each filter pole. Achieving 60dB attenuation in a transition region
between 1MHz and 2MHz (1 octave) requires a minimum of 10 poles - not a trivial
filter, and definitely a design challenge.
Therefore, other filter types are generally more suited to high speed applications
where the requirement is for a sharp transition band and in-band flatness coupled
with linear phase response. Elliptic filters meet these criteria and are a popular
choice. There are a number of companies which specialize in supplying custom
analog filters. TTE is an example of such a company (Reference 1).
From this discussion, we can see how the sharpness of the antialiasing transition
band can be traded off against the ADC sampling frequency. Choosing a higher
sampling rate (oversampling) reduces the requirement on transition band sharpness
(hence, the filter complexity) at the expense of using a faster ADC and processing
data at a faster rate. This is illustrated in Figure 2.5B which shows the effects of
increasing the sampling frequency by a factor of K, while maintaining the same
analog corner frequency, fa, and the same dynamic range, DR, requirement. The
wider transition band (fa to Kfs – fa) makes this filter easier to design than for the
case of Figure 2.5A.
2.5
SAMPLED DATA SYSTEMS
The antialiasing filter design process is started by choosing an initial sampling rate
of 2.5 to 4 times fa. Determine the filter specifications based on the required
dynamic range and see if such a filter is realizable within the constraints of the
system cost and performance. If not, consider a higher sampling rate which may
require using a faster ADC. It should be mentioned that sigma-delta ADCs are
inherently oversampling converters, and the resulting relaxation in the analog antialiasing filter requirements is therefore an added benefit of this architecture.
The antialiasing filter requirements can also be relaxed somewhat if it is certain
that there will never be a fullscale signal at the stopband frequency fs – fa. In many
applications, it is improbable that fullscale signals will occur at this frequency. If
the maximum signal at the frequency fs – fa will never exceed XdB below fullscale,
then the filter stopband attenuation requirement is reduced by that same amount.
The new requirement for stopband attenuation at fs – fa based on this knowledge of
the signal is now only DR – XdB. When making this type of assumption, be careful
to treat any noise signals which may occur above the maximum signal frequency fa
as unwanted signals which will also alias back into the signal bandwidth.
Undersampling (Harmonic Sampling, Bandpass Sampling, IF Sampling,
Direct IF to Digital Conversion)
Thus far we have considered the case of baseband sampling, i.e., all the signals of
interest lie within the first Nyquist zone. Figure 2.6A shows such a case, where the
band of sampled signals is limited to the first Nyquist zone, and images of the
original band of frequencies appear in each of the other Nyquist zones.
UNDERSAMPLING
A
ZONE 1
I
I
0.5fs
fs
I
I
1.5fs
I
2.5fs
2fs
3.5fs
3fs
ZONE 2
B
I
I
fs
0.5fs
I
1.5fs
I
I
2fs
I
3fs
2.5fs
3.5fs
ZONE 3
C
I
I
0.5fs
I
fs
1.5fs
I
2fs
Figure 2.6
2.6
I
I
2.5fs
I
3fs
3.5fs
SAMPLED DATA SYSTEMS
Consider the case shown in Figure 2.6B, where the sampled signal band lies entirely
within the second Nyquist zone. The process of sampling a signal outside the first
Nyquist zone is often referred to as undersampling, or harmonic sampling. Note
that the first Nyquist zone image contains all the information in the original signal,
with the exception of its original location (the order of the frequency components
within the spectrum is reversed, but this is easily corrected by re-ordering the
output of the FFT).
Figure 2.6C shows the sampled signal restricted to the third Nyquist zone. Note
that the first Nyquist zone image has no frequency reversal. In fact, the sampled
signal frequencies may lie in any unique Nyquist zone, and the first Nyquist zone
image is still an accurate representation (with the exception of the frequency
reversal which occurs when the signals are located in even Nyquist zones). At this
point we can clearly restate the Nyquist criteria:
A signal must be sampled at a rate equal to or greater than twice its bandwidth in
order to preserve all the signal information.
Notice that there is no mention of the absolute location of the band of sampled
signals within the frequency spectrum relative to the sampling frequency. The only
constraint is that the band of sampled signals be restricted to a single Nyquist zone,
i.e., the signals must not overlap any multiple of fs/2 (this, in fact, is the primary
function of the antialiasing filter).
Sampling signals above the first Nyquist zone has become popular in
communications because the process is equivalent to analog demodulation. It is
becoming common practice to sample IF signals directly and then use digital
techniques to process the signal, thereby eliminating the need for the IF
demodulator. Clearly, however, as the IF frequencies become higher, the dynamic
performance requirements on the ADC become more critical. The ADC input
bandwidth and distortion performance must be adequate at the IF frequency, rather
than only baseband. This presents a problem for most ADCs designed to process
signals in the first Nyquist zone, therefore an ADC suitable for undersampling
applications must maintain dynamic performance into the higher order Nyquist
zones.
ADC AND DAC STATIC TRANSFER FUNCTIONS AND DC ERRORS
The most important thing to remember about both DACs and ADCs is that either
the input or output is digital, and therefore the signal is quantized. That is, an N-bit
word represents one of 2N possible states, and therefore an N-bit DAC (with a fixed
reference) can have only 2N possible analog outputs, and an N-bit ADC can have
only 2N possible digital outputs. The analog signals will generally be voltages or
currents.
The resolution of data converters may be expressed in several different ways: the
weight of the Least Significant Bit (LSB), parts per million of full scale (ppm FS),
millivolts (mV), etc. Different devices (even from the same manufacturer) will be
specified differently, so converter users must learn to translate between the
different types of specifications if they are to compare devices successfully. The size
of the least significant bit for various resolutions is shown in Figure 2.7.
2.7
SAMPLED DATA SYSTEMS
QUANTIZATION:
THE SIZE OF A LEAST SIGNIFICANT BIT (LSB)
RESOLUTION
N
2N
VOLTAGE
(10V FS)
ppm FS
% FS
dB FS
2-bit
4
2.5 V
250,000
25
-12
4-bit
16
625 mV
62,500
6.25
-24
6-bit
64
156 mV
15,625
1.56
-36
8-bit
256
39.1 mV
3,906
0.39
-48
10-bit
1,024
9.77 mV (10 mV)
977
0.098
-60
12-bit
4,096
2.44 mV
244
0.024
-72
14-bit
16,384
610 µV
61
0.0061
-84
16-bit
65,536
153 µV
15
0.0015
-96
18-bit
262,144
38 µV
4
0.0004
-108
20-bit
1,048,576
9.54 µV (10 µV)
1
0.0001
-120
22-bit
4,194,304
2.38 µV
0.24
0.000024
-132
24-bit
16,777,216
596 nV*
0.06
0.000006
-144
*600nV is the Johnson Noise in a 10kHz BW of a 2.2kΩ
Ω Resistor @ 25°C
Remember: 10-bits and 10V FS yields an LSB of 10mV, 1000ppm, or 0.1%.
All other values may be calculated by powers of 2.
Figure 2.7
Before we can consider the various architectures used in data converters, it is
necessary to consider the performance to be expected, and the specifications which
are important. The following sections will consider the definition of errors and
specifications used for data converters. This is important in understanding the
strengths and weaknesses of different ADC/DAC archectures.
The first applications of data converters were in measurement and control where
the exact timing of the conversion was usually unimportant, and the data rate was
slow. In such applications, the DC specifications of converters are important, but
timing and AC specifications are not. Today many, if not most, converters are used
in sampling and reconstruction systems where AC specifications are critical (and
DC ones may not be) - these will be considered in the next part of this section.
Figure 2.8 shows the ideal transfer characteristics for a 3-bit unipolar DAC, and
Figure 2.9 a 3-bit unipolar ADC. In a DAC, both the input and the output are
quantized, and the graph consists of eight points - while it is reasonable to discuss
the line through these points, it is very important to remember that the actual
transfer characteristic is not a line, but a number of discrete points.
2.8
SAMPLED DATA SYSTEMS
TRANSFER FUNCTION FOR IDEAL 3-BIT DAC
FS
ANALOG
OUTPUT
000
001
010
011
100
101
110
111
DIGITAL INPUT
Figure 2.8
TRANSFER FUNCTION FOR IDEAL 3-BIT ADC
111
110
DIGITAL
OUTPUT
101
100
011
010
QUANTIZATION
UNCERTAINTY
001
000
ANALOG INPUT
FS
Figure 2.9
2.9
SAMPLED DATA SYSTEMS
The input to an ADC is analog and is not quantized, but its output is quantized. The
transfer characteristic therefore consists of eight horizontal steps (when considering
the offset, gain and linearity of an ADC we consider the line joining the midpoints of
these steps).
In both cases, digital full scale (all "1"s) corresponds to 1 LSB below the analog full
scale (the reference, or some multiple thereof). This is because, as mentioned above,
the digital code represents the normalized ratio of the analog signal to the
reference.
The (ideal) ADC transitions take place at ½ LSB above zero, and thereafter every
LSB, until 1½ LSB below analog full scale. Since the analog input to an ADC can
take any value, but the digital output is quantized, there may be a difference of up
to ½ LSB between the actual analog input and the exact value of the digital output.
This is known as the quantization error or quantization uncertainty as shown in
Figure 2.9. In AC (sampling) applications this quantization error gives rise to
quantization noise which will be discussed in the next section.
There are many possible digital coding schemes for data converters: binary, offset
binary, 1's complement, 2's complement, gray code, BCD and others. This section,
being devoted mainly to the analog issues surrounding data converters, will use
simple binary and offset binary in its examples and will not consider the merits and
disadvantages of these, or any other forms of digital code.
The examples in Figures 2.8 and 2.9 use unipolar converters, whose analog port has
only a single polarity. These are the simplest type, but bipolar converters are
generally more useful in real-world applications. There are two types of bipolar
converters: the simpler is merely a unipolar converter with an accurate 1 MSB of
negative offset (and many converters are arranged so that this offset may be
switched in and out so that they can be used as either unipolar or bipolar converters
at will), but the other, known as a sign-magnitude converter is more complex, and
has N bits of magnitude information and an additional bit which corresponds to the
sign of the analog signal. Sign-magnitude DACs are quite rare, and sign-magnitude
ADCs are found mostly in digital voltmeters (DVMs).
The four DC errors in a data converter are offset error, gain error, and two types of
linearity error. Offset and gain errors are analogous to offset and gain errors in
amplifiers as shown in Figure 2.10 for a bipolar input range. (Though offset error
and zero error, which are identical in amplifiers and unipolar data converters, are
not identical in bipolar converters and should be carefully distinguished.) The
transfer characteristics of both DACs and ADCs may be expressed as D = K + GA,
where D is the digital code, A is the analog signal, and K and G are constants. In a
unipolar converter, K is zero, and in an offset bipolar converter, it is –1 MSB. The
offset error is the amount by which the actual value of K differs from its ideal value.
The gain error is the amount by which G differs from its ideal value, and is
generally expressed as the percentage difference between the two, although it may
be defined as the gain error contribution (in mV or LSB) to the total error at full
scale. These errors can usually be trimmed by the data converter user. Note,
however, that amplifier offset is trimmed at zero input, and then the gain is
trimmed near to full scale. The trim algorithm for a bipolar data converter is not so
straightforward.
2.10
SAMPLED DATA SYSTEMS
CONVERTER OFFSET AND GAIN ERROR
+FS
+FS
ACTUAL
ACTUAL
IDEAL
IDEAL
0
0
ZERO ERROR
OFFSET
ERROR
ZERO ERROR
NO GAIN ERROR:
ZERO ERROR = OFFSET ERROR
–FS
–FS
WITH GAIN ERROR:
OFFSET ERROR = 0
ZERO ERROR RESULTS
FROM GAIN ERROR
Figure 2.10
The integral linearity error of a converter is also analogous to the linearity error of
an amplifier, and is defined as the maximum deviation of the actual transfer
characteristic of the converter from a straight line, and is generally expressed as a
percentage of full scale (but may be given in LSBs). There are two common ways of
choosing the straight line: end point and best straight line (see Figure 2.11).
METHOD OF MEASURING INTEGRAL LINEARITY ERRORS
(SAME CONVERTER ON BOTH GRAPHS)
END POINT METHOD
BEST STRAIGHT LINE METHOD
OUTPUT
LINEARITY
ERROR = X
INPUT
LINEARITY
ERROR ≈ X/2
INPUT
Figure 2.11
2.11
SAMPLED DATA SYSTEMS
In the end point system, the deviation is measured from the straight line through
the origin and the full scale point (after gain adjustment). This is the most useful
integral linearity measurement for measurement and control applications of data
converters (since error budgets depend on deviation from the ideal transfer
characteristic, not from some arbitrary "best fit"), and is the one normally adopted
by Analog Devices, Inc.
The best straight line, however, does give a better prediction of distortion in AC
applications, and also gives a lower value of "linearity error" on a data sheet. The
best fit straight line is drawn through the transfer characteristic of the device using
standard curve fitting techniques, and the maximum deviation is measured from
this line. In general, the integral linearity error measured in this way is only 50% of
the value measured by end point methods. This makes the method good for
producing impressive data sheets, but it is less useful for error budget analysis. For
AC applications, it is even better to specify distortion than DC linearity, so it is
rarely necessary to use the best straight line method to define converter linearity.
The other type of converter non-linearity is differential non-linearity (DNL). This
relates to the linearity of the code transitions of the converter. In the ideal case, a
change of 1 LSB in digital code corresponds to a change of exactly 1 LSB of analog
signal. In a DAC, a change of 1 LSB in digital code produces exactly 1 LSB change
of analog output, while in an ADC there should be exactly 1 LSB change of analog
input to move from one digital transition to the next.
Where the change in analog signal corresponding to 1 LSB digital change is more or
less than 1 LSB, there is said to be a DNL error. The DNL error of a converter is
normally defined as the maximum value of DNL to be found at any transition.
If the DNL of a DAC is less than –1 LSB at any transition (see Figure 2.12), the
DAC is non-monotonic i.e., its transfer characteristic contains one or more localized
maxima or minima. A DNL greater than +1 LSB does not cause non-monotonicity,
but is still undesirable. In many DAC applications (especially closed-loop systems
where non-monotonicity can change negative feedback to positive feedback), it is
critically important that DACs are monotonic. DAC monotonicity is often explicitly
specified on data sheets, although if the DNL is guaranteed to be less than 1 LSB
(i.e., |DNL| ≤ 1LSB) then the device must be monotonic, even without an explicit
guarantee.
ADCs can be non-monotonic, but a more common result of excess DNL in ADCs is
missing codes (see Figure 2.13). Missing codes (or non-monotonicity) in an ADC are
as objectionable as non-monotonicity in a DAC. Again, they result from DNL >
1 LSB.
2.12
SAMPLED DATA SYSTEMS
TRANSFER FUNCTION OF NON-IDEAL 3-BIT DAC
FS
ANALOG
OUTPUT
NON-MONOTONIC
000
001
010
011
100
101
110
111
DIGITAL INPUT
Figure 2.12
TRANSFER FUNCTION OF NON-IDEAL 3-BIT ADC
111
110
DIGITAL
OUTPUT
101
100
MISSING CODE
011
010
001
000
ANALOG INPUT
FS
Figure 2.13
2.13
SAMPLED DATA SYSTEMS
Defining missing codes is more difficult than defining non-monotonicity. All ADCs
suffer from some transition noise as shown in Figure 2.14 (think of it as the flicker
between adjacent values of the last digit of a DVM). As resolutions become higher,
the range of input over which transition noise occurs may approach, or even exceed,
1 LSB. In such a case, especially if combined with a negative DNL error, it may be
that there are some (or even all) codes where transition noise is present for the
whole range of inputs. There are therefore some codes for which there is no input
which will guarantee that code as an output, although there may be a range of
inputs which will sometimes produce that code.
COMBINED EFFECTS OF ADC
CODE TRANSITION NOISE AND DNL
CODE TRANSITION NOISE
DNL
TRANSITION NOISE
AND DNL
ADC
OUTPUT
CODE
ADC INPUT
ADC INPUT
ADC INPUT
Figure 2.14
For lower resolution ADCs, it may be reasonable to define no missing codes as a
combination of transition noise and DNL which guarantees some level (perhaps
0.2 LSB) of noise-free code for all codes. However, this is impossible to achieve at
the very high resolutions achieved by modern sigma-delta ADCs, or even at lower
resolutions in wide bandwidth sampling ADCs. In these cases, the manufacturer
must define noise levels and resolution in some other way. Which method is used is
less important, but the data sheet should contain a clear definition of the method
used and the performance to be expected.
AC ERRORS IN DATA CONVERTERS
Over the last decade, a major application of data converters is in AC sampling and
reconstruction. In very simple terms, a sampled data system is a system where the
instantaneous value of an AC waveform is sampled at regular intervals. The
resulting digital codes may be used to store the waveform (as in CDs and DATs), or
intensive computation on the samples (Digital Signal Processing, or DSP) may be
used to perform filtering, compression, and other operations. The inverse operation,
reconstruction, occurs when a series of digital codes are fed to a DAC to reconstruct
2.14
SAMPLED DATA SYSTEMS
an AC waveform - an obvious example of this is a CD or DAT player, but the
technique is very widely used indeed in telecommunications, radio, synthesizers,
and many other applications.
The data converters used in these applications must have good performance with
AC signals, but may not require good DC specifications. The first high performance
converters to be designed for such applications were often manufactured with good
AC specifications but poor, or unspecified, DC performance. Today the design
tradeoffs are better understood, and most converters will have good, and
guaranteed, AC and DC specifications. DACs for digital audio, however, which must
be extremely competitive in price, are generally sold with comparatively poor DC
specifications - not because their DC performance is poor, but because it is not
tested during manufacture.
While it is easier to discuss the DC parameters of both DACs and ADCs together,
their AC specifications are sufficiently different to deserve separate consideration.
Distortion and Noise in an Ideal N-Bit ADC
Thus far we have looked at the implications of the sampling process without
considering the effects of ADC quantization. We will now treat the ADC as an ideal
sampler, but include the effects of quantization.
The only errors (DC or AC) associated with an ideal N-bit ADC are those related to
the sampling and quantization processes. The maximum error an ideal ADC makes
when digitizing a DC input signal is ±1/2LSB. Any AC signal applied to an ideal Nbit ADC will produce quantization noise whose rms value (measured over the
Nyquist bandwidth, DC to fs/2) is approximately equal to the weight of the least
significant bit (LSB), q, divided by √12. (See Reference 2). This assumes that the
signal is at least a few LSBs in amplitude so that the ADC output always changes
state. The quantization error signal from a linear ramp input is approximated as a
sawtooth waveform with a peak-to-peak amplitude equal to q, and its rms value is
therefore q/√12 (see Figure 2.15).
It can be shown that the ratio of the rms value of a full scale sinewave to the rms
value of the quantization noise (expressed in dB) is:
SNR = 6.02N + 1.76dB,
where N is the number of bits in the ideal ADC. This equation is only valid if the
noise is measured over the entire Nyquist bandwidth from DC to fs/2 as shown in
Figure 2.16. If the signal bandwidth, BW, is less than fs/2, then the SNR within the
signal bandwidth BW is increased because the amount of quantization noise within
the signal bandwidth is smaller. The correct expression for this condition is given
by:
 fs 
SNR = 6.02N + 176
. dB + 10 log
.
 2 ⋅ BW 
2.15
SAMPLED DATA SYSTEMS
IDEAL N-BIT ADC QUANTIZATION NOISE
DIGITAL
CODE
OUTPUT
ANALOG
INPUT
q = 1LSB
ERROR
RMS ERROR = q/√ 12
fs
2•BW
SNR = 6.02N + 1.76dB + 10log
FOR FS SINEWAVE
Figure 2.15
QUANTIZATION NOISE SPECTRUM
RMS QUANTIZATION NOISE = q /
q = WEIGHT OF LSB
12,
SNR (FS RMS INPUT) = (6.02N + 1.76)dB, OVER fs/2 BANDWIDTH
WHERE N = NUMBER OF BITS
ASSUME QUANTIZATION NOISE IS UNIFORMLY
DISTRIBUTED: DC TO f s /2
(SPECIAL CASES OCCUR WHERE THIS IS NOT TRUE)
fs
DC
BW
fs
2
IF BW <
fs
2
SNR = 6.02N + 1.76dB + 10log
THEN:
fs
2•BW
Figure 2.16
2.16
FOR FS SINEWAVE
SAMPLED DATA SYSTEMS
The above equation reflects the condition called oversampling, where the sampling
frequency is higher than twice the signal bandwidth. The correction term is often
called processing gain. Notice that for a given signal bandwidth, doubling the
sampling frequency increases the SNR by 3dB.
Although the rms value of the noise is accurately approximated by q/√12, its
frequency domain content may be highly correlated to the AC input signal. For
instance, there is greater correlation for low amplitude periodic signals than for
large amplitude random signals. Quite often, the assumption is made that the
theoretical quantization noise appears as white noise, spread uniformly over the
Nyquist bandwidth DC to fs/2. Unfortunately, this is not true. In the case of strong
correlation, the quantization noise appears concentrated at the various harmonics of
the input signal, just where you don't want them.
In most applications, the input to the ADC is a band of frequencies (usually
summed with some noise), so the quantization noise tends to be random. In spectral
analysis applications (or in performing FFTs on ADCs using spectrally pure
sinewaves - see Figure 2.17), however, the correlation between the quantization
noise and the signal depends upon the ratio of the sampling frequency to the input
signal. This is demonstrated in Figure 2.18, where an ideal 12-bit ADCs output is
analyzed using a 4096-point FFT. In the left-hand FFT plot, the ratio of the
sampling frequency to the input frequency was chosen to be exactly 32, and the
worst harmonic is about 76dB below the fundamental. The right hand diagram
shows the effects of slightly offsetting the ratio, showing a relatively random noise
spectrum, where the SFDR is now about 92dBc. In both cases, the rms value of all
the noise components is q/√12, but in the first case, the noise is concentrated at
harmonics of the fundamental.
DYNAMIC PERFORMANCE ANALYSIS
OF AN IDEAL N-BIT ADC
fs
ANALOG
INPUT
fa
IDEAL
N-BIT
ADC
N
BUFFER
MEMORY
M-WORDS
M POINT
M-POINT
2
FFT
PROCESSOR SPECTRAL
OUTPUT
Figure 2.17
2.17
SAMPLED DATA SYSTEMS
EFFECT OF RATIO OF SAMPLING CLOCK TO INPUT
FREQUENCY ON SFDR FOR IDEAL 12-BIT ADC
0
-10
-20
-30
-40
-50
-60
-70
-80
-90
-100
-110
-120
fs / fa = 32
0
500
1000
SFDR = 76dBc
M = 4096
1500
2000
0
fs / fa = 32.25196850394
500
1000
SFDR = 92dBc
1500
2000
Figure 2.18
Note that this variation in the apparent harmonic distortion of the ADC is an
artifact of the sampling process and the correlation of the quantization error with
the input frequency. In a practical ADC application, the quantization error
generally appears as random noise because of the random nature of the wideband
input signal and the additional fact that there is a usually a small amount of system
noise which acts as a dither signal to further randomize the quantization error
spectrum.
It is important to understand the above point, because single-tone sinewave FFT
testing of ADCs is a universally accepted method of performance evaluation. In
order to accurately measure the harmonic distortion of an ADC, steps must be
taken to ensure that the test setup truly measures the ADC distortion, not the
artifacts due to quantization noise correlation. This is done by properly choosing the
frequency ratio and sometimes by injecting a small amount of noise (dither) with
the input signal.
Now, return to Figure 2.18, and note that the average value of the noise floor of the
FFT is approximately 100dB below full scale, but the theoretical SNR of a 12-bit
ADC is 74dB. The FFT noise floor is not the SNR of the ADC, because the FFT acts
like an analog spectrum analyzer with a bandwidth of fs/M, where M is the number
of points in the FFT. The theoretical FFT noise floor is therefore 10log10(M/2)dB
below the quantization noise floor due to the so-called processing gain of the FFT
(see Figure 2.19). In the case of an ideal 12-bit ADC with an SNR of 74dB, a 4096point FFT would result in a processing gain of 10log10(4096/2) = 33dB, thereby
2.18
SAMPLED DATA SYSTEMS
resulting in an overall FFT noise floor of 74+33=107dBc. In fact, the FFT noise floor
can be reduced even further by going to larger and larger FFTs; just as an analog
spectrum analyzer's noise floor can be reduced by narrowing the bandwidth. When
testing ADCs using FFTs, it is important to ensure that the FFT size is large
enough so that the distortion products can be distinguished from the FFT noise floor
itself.
NOISE FLOOR FOR AN IDEAL 12-BIT ADC
USING 4096-POINT FFT
(dB)
0
ADC FULLSCALE
N = 12-BITS
M = 4096
20
40
74dB = 6.02N + 1.76dB
60
RMS QUANTIZATION NOISE LEVEL
80
74dB
( M)
33dB = 10log 2
100
FFT NOISE FLOOR
120
BIN SPACING =
fs
4096
107dB
fs
2
Figure 2.19
Distortion and Noise in Practical ADCs
A practical sampling ADC (one that has an integral sample-and-hold), regardless of
architecture, has a number of noise and distortion sources as shown in Figure 2.20.
The wideband analog front-end buffer has wideband noise, non-linearity, and also
finite bandwidth. The SHA introduces further non-linearity, bandlimiting, and
aperture jitter. The actual quantizer portion of the ADC introduces quantization
noise, and both integral and differential non-linearity. In this discussion, assume
that sequential outputs of the ADC are loaded into a buffer memory of length M and
that the FFT processor provides the spectral output. Also assume that the FFT
arithmetic operations themselves introduce no significant errors relative to the
ADC. However, when examining the output noise floor, the FFT processing gain
(dependent on M) must be considered.
2.19
SAMPLED DATA SYSTEMS
ADC MODEL SHOWING NOISE AND
DISTORTION SOURCES
fs
ANALOG
INPUT
ADC
BUFFER
NOISE
DISTORTION
BAND LIMITING
TO MEMORY
SAMPLE
AND
HOLD
ENCODER
N-BITS
NOISE
DISTORTION
BAND LIMITING
APERTURE JITTER
N
TEST
SYSTEM
BUFFER
MEMORY
M- WORDS
M-POINT
FFT
PROCESSOR
QUANTIZATION NOISE
DIFFERENTIAL NON-LINEARITY
INTEGRAL NON-LINEARITY
M POINT
2
SPECTRAL
OUTPUT
PROCESSING GAIN = 10log
( M2 )
ROUND OFF ERROR (NEGLIGIBLE)
Figure 2.20
Equivalent Input Referred Noise (Thermal Noise)
The wideband ADC internal circuits produce a certain amount of wideband rms
noise due to thermal and kT/C effects. This noise is present even for DC input
signals, and accounts for the fact that the output of most wideband (or high
resolution) ADCs is a distribution of codes, centered around the nominal value of a
DC input (see Figure 2.21). To measure its value, the input of the ADC is grounded,
and a large number of output samples are collected and plotted as a histogram
(sometimes referred to as a grounded-input histogram). Since the noise is
approximately Gaussian, the standard deviation of the histogram is easily
calculated (see Reference 3), corresponding to the effective input rms noise. It is
common practice to express this rms noise in terms of LSBs, although it can be
expressed as an rms voltage.
There are various ways to characterize the AC performance of ADCs. In the early
years of ADC technology (over 30 years ago) there was little standardization with
respect to AC specifications, and measurement equipment and techniques were not
well understood or available. Over nearly a 30 year period, manufacturers and
customers have learned more about measuring the dynamic performance of
converters, and the specifications shown in Figure 2.22 represent the most popular
ones used today. Practically all the specifications represent the converter’s
performance in the frequency domain. The FFT is the heart of practically all these
measurements and is discussed in detail in Section 5 of this book.
2.20
SAMPLED DATA SYSTEMS
EFFECT OF INPUT-REFERRED NOISE
ON ADC "GROUNDED INPUT" HISTOGRAM
NUMBER OF
OCCURANCES
P-P INPUT NOISE
≈
6.6 × RMS NOISE
RMS NOISE
n–4
n–3
n–2
n–1
n
n+1
n+2
n+3
n+4
OUTPUT CODE
Figure 2.21
QUANTIFYING ADC DYNAMIC PERFORMANCE
■ Harmonic Distortion
■ Worst Harmonic
■ Total Harmonic Distortion (THD)
■ Total Harmonic Distortion Plus Noise (THD + N)
■ Signal-to-Noise-and-Distortion Ratio (SINAD, or S/N +D)
■ Effective Number of Bits (ENOB)
■ Signal-to-Noise Ratio (SNR)
■ Analog Bandwidth (Full-Power, Small-Signal)
■ Spurious Free Dynamic Range (SFDR)
■ Two-Tone Intermodulation Distortion
■ Multi-tone Intermodulation Distortion
Figure 2.22
2.21
SAMPLED DATA SYSTEMS
Integral and Differential Non-Linearity Distortion Effects
One of the first things to realize when examining the nonlinearities of data
converters is that the transfer function of a data converter has artifacts which do
not occur in conventional linear devices such as op amps or gain blocks. The overall
integral non-linearity of an ADC is due to the integral non-linearity of the front-end
and SHA as well as the overall integral non-linearity in the ADC transfer function.
However, differential non-linearity is due exclusively to the encoding process and
may vary considerably dependent on the ADC encoding architecture. Overall
integral non-linearity produces distortion products whose amplitude varies as a
function of the input signal amplitude. For instance, second-order intermodulation
products increase 2dB for every 1dB increase in signal level, and third-order
products increase 3dB for every 1dB increase in signal level.
The differential non-linearity in the ADC transfer function produces distortion
products which not only depend on the amplitude of the signal but the positioning of
the differential non-linearity along the ADC transfer function. Figure 2.23 shows
two ADC transfer functions having differential non-linearity. The left-hand diagram
shows an error which occurs at midscale. Therefore, for both large and small
signals, the signal crosses through this point producing a distortion product which
is relatively independent of the signal amplitude. The right-hand diagram shows
another ADC transfer function which has differential non-linearity errors at 1/4 and
3/4 full scale. Signals which are above 1/2 scale peak-to-peak will exercise these
codes and produce distortion, while those less than 1/2 scale peak-to-peak will not.
TYPICAL ADC / DAC DNL ERRORS
OUT
OUT
IN
IN
MIDSCALE DNL
1/4FS, 3/4FS DNL
Figure 2.23
Most high-speed ADCs are designed so that differential non-linearity is spread
across the entire ADC range. Therefore, for signals which are within a few dB of full
scale, the overall integral non-linearity of the transfer function determines the
distortion products. For lower level signals, however, the harmonic content becomes
2.22
SAMPLED DATA SYSTEMS
dominated by the differential non-linearities and does not generally decrease
proportionally with decreases in signal amplitude.
Harmonic Distortion, Worst Harmonic, Total Harmonic Distortion (THD),
Total Harmonic Distortion Plus Noise (THD + N)
There are a number of ways to quantify the distortion of an ADC. An FFT analysis
can be used to measure the amplitude of the various harmonics of a signal. The
harmonics of the input signal can be distinguished from other distortion products by
their location in the frequency spectrum. Figure 2.24 shows a 7MHz input signal
sampled at 20MSPS and the location of the first 9 harmonics. Aliased harmonics of
fa fall at frequencies equal to |±Kfs±nfa|, where n is the order of the harmonic, and
K = 0, 1, 2, 3,.... The second and third harmonics are generally the only ones
specified on a data sheet because they tend to be the largest, although some data
sheets may specify the value of the worst harmonic. Harmonic distortion is normally
specified in dBc (decibels below carrier), although at audio frequencies it may be
specified as a percentage. Harmonic distortion is generally specified with an input
signal near full scale (generally 0.5 to 1dB below full scale to prevent clipping), but
it can be specified at any level. For signals much lower than full scale, other
distortion products due to the DNL of the converter (not direct harmonics) may limit
performance.
LOCATION OF HARMONIC DISTORTION PRODUCTS:
INPUT SIGNAL = 7MHz, SAMPLING RATE = 20MSPS
fa = 7MHz
RELATIVE
AMPLITUDE
fs = 20MSPS
HARMONICS AT: |±Kfs±nfa|
n = ORDER OF HARMONIC, K = 0, 1, 2, 3, . . .
3
2
6
1
2
9
8
5
3
4
5
4
6
7
8
7
9
10
FREQUENCY (MHz)
Figure 2.24
2.23
SAMPLED DATA SYSTEMS
Total harmonic distortion (THD) is the ratio of the rms value of the fundamental
signal to the mean value of the root-sum-square of its harmonics (generally, only the
first 5 are significant). THD of an ADC is also generally specified with the input
signal close to full scale, although it can be specified at any level.
Total harmonic distortion plus noise (THD+ N) is the ratio of the rms value of the
fundamental signal to the mean value of the root-sum-square of its harmonics plus
all noise components (excluding DC). The bandwidth over which the noise is
measured must be specified. In the case of an FFT, the bandwidth is DC to fs/2. (If
the bandwidth of the measurement is DC to fs/2, THD+N is equal to SINAD - see
below).
Signal-to-Noise-and-Distortion Ratio (SINAD), Signal-to-Noise Ratio (SNR),
and Effective Number of Bits (ENOB)
SINAD and SNR deserve careful attention, because there is still some variation
between ADC manufacturers as to their precise meaning. Signal-to-noise-and
Distortion (SINAD, or S/N+D) is the ratio of the rms signal amplitude to the mean
value of the root-sum-square (RSS) of all other spectral components, including
harmonics, but excluding DC. SINAD is a good indication of the overall dynamic
performance of an ADC as a function of input frequency because it includes all
components which make up noise (including thermal noise) and distortion. It is
often plotted for various input amplitudes. SINAD is equal to THD+N if the
bandwidth for the noise measurement is the same. A typical plot for the AD9220 12bit, 10MSPS ADC is shown in Figure 2.26.
SINAD, ENOB, AND SNR
■ SINAD (Signal-to-Noise-and-Distortion Ratio):
◆ The ratio of the rms signal amplitude to the mean value of the
root-sum-squares (RSS) of all other spectral components,
including harmonics, but excluding DC.
■ ENOB (Effective Number of Bits):
ENOB =
SINAD – 1.76dB
6.02
■ SNR (Signal-to-Noise Ratio, or Signal-to-Noise Ratio Without
Harmonics:
◆ The ratio of the rms signal amplitude to the mean value of the
root-sum-squares (RSS) of all other spectral components,
excluding the first 5 harmonics and DC
Figure 2.25
2.24
SAMPLED DATA SYSTEMS
AD9220 12-BIT, 10MSPS ADC SINAD AND ENOB
FOR VARIOUS INPUT SIGNAL LEVELS
13.0
12.2
11.3
SINAD
(dB)
ENOB
10.5
9.7
8.8
8.0
7.2
6.3
ANALOG INPUT FREQUENCY (MHz)
Figure 2.26
The SINAD plot shows where the AC performance of the ADC degrades due to highfrequency distortion and is usually plotted for frequencies well above the Nyquist
frequency so that performance in undersampling applications can be evaluated.
SINAD is often converted to effective-number-of-bits (ENOB) using the relationship
for the theoretical SNR of an ideal N-bit ADC: SNR = 6.02N + 1.76dB. The equation
is solved for N, and the value of SINAD is substituted for SNR:
ENOB =
SINAD − 176
. dB
.
6.02
Signal-to-noise ratio (SNR, or SNR-without-harmonics) is calculated the same as
SINAD except that the signal harmonics are excluded from the calculation, leaving
only the noise terms. In practice, it is only necessary to exclude the first 5
harmonics since they dominate. The SNR plot will degrade at high frequencies, but
not as rapidly as SINAD because of the exclusion of the harmonic terms.
Many current ADC data sheets somewhat loosely refer to SINAD as SNR, so the
engineer must be careful when interpreting these specifications.
Analog Bandwidth
The analog bandwidth of an ADC is that frequency at which the spectral output of
the fundamental swept frequency (as determined by the FFT analysis) is reduced by
3dB. It may be specified for either a small signal (SSBW- small signal bandwidth),
2.25
SAMPLED DATA SYSTEMS
or a full scale signal (FPBW- full power bandwidth), so there can be a wide variation
in specifications between manufacturers.
Like an amplifier, the analog bandwidth specification of a converter does not imply
that the ADC maintains good distortion performance up to its bandwidth frequency.
In fact, the SINAD (or ENOB) of most ADCs will begin to degrade considerably
before the input frequency approaches the actual 3dB bandwidth frequency. Figure
2.27 shows ENOB and full scale frequency response of an ADC with a FPBW of
1MHz, however, the ENOB begins to drop rapidly above 100kHz.
ADC GAIN (BANDWIDTH) AND ENOB VERSUS FREQUENCY
SHOWS IMPORTANCE OF ENOB SPECIFICATION
FPBW = 1MHz
GAIN (FS INPUT)
GAIN
ENOB (FS INPUT)
ENOB
ENOB (-20dB INPUT)
10
100
1k
10k
100k
1M
10M
ADC INPUT FREQUENCY (Hz)
Figure 2.27
Spurious Free Dynamic Range (SFDR)
Probably the most significant specification for an ADC used in a communications
application is its spurious free dynamic range (SFDR). The SFDR specification is to
ADCs what the third order intercept specification is to mixers and LNAs. SFDR of
an ADC is defined as the ratio of the rms signal amplitude to the rms value of the
peak spurious spectral content (measured over the entire first Nyquist zone, DC to
fs/2). SFDR is generally plotted as a function of signal amplitude and may be
expressed relative to the signal amplitude (dBc) or the ADC full scale (dBFS) as
shown in Figure 2.28.
2.26
SAMPLED DATA SYSTEMS
SPURIOUS FREE DYNAMIC RANGE (SFDR)
FULL SCALE (FS)
INPUT SIGNAL LEVEL (CARRIER)
SFDR (dBFS)
dB
SFDR (dBc)
WORST SPUR LEVEL
FREQUENCY
fs
2
Figure 2.28
For a signal near full scale, the peak spectral spur is generally determined by one of
the first few harmonics of the fundamental. However, as the signal falls several dB
below full scale, other spurs generally occur which are not direct harmonics of the
input signal. This is because of the differential non-linearity of the ADC transfer
function as discussed earlier. Therefore, SFDR considers all sources of distortion,
regardless of their origin.
The AD9042 is a 12-bit, 41MSPS wideband ADC designed for communications
applications where high SFDR is important. The SFDR for a 19.5MHz input and a
sampling frequency of 41MSPS is shown in Figure 2.29. Note that a minimum of
80dBc SFDR is obtained over the entire first Nyquist zone (DC to 20MHz). The plot
also shows SFDR expressed as dBFS.
SFDR is generally much greater than the ADCs theoretical N-bit SNR (6.02N +
1.76dB). For example, the AD9042 is a 12-bit ADC with an SFDR of 80dBc and a
typical SNR of 65dBc (theoretical SNR is 74dB). This is because there is a
fundamental distinction between noise and distortion measurements. The process
gain of the FFT (33dB for a 4096-point FFT) allows frequency spurs well below the
noise floor to be observed. Adding extra resolution to an ADC may serve to increase
its SNR but may or may not increase its SFDR.
2.27
SAMPLED DATA SYSTEMS
AD9042 12-BIT, 41MSPS ADC
SFDR VS. INPUT POWER LEVEL
SAMPLING FREQUENCY = 41MSPS
INPUT FREQUENCY
= 19.5MHz
100
dBFS
90
80
SFDR
(dB)
70
60
50
40
SFDR = 80dB
REFERENCE LINE
dBc
30
20
10
0
-80
-70
-60
-50
-40
-30
-20
-10
0
ANALOG INPUT POWER LEVEL (dBFS)
Figure 2.29
Two Tone Intermodulation Distortion (IMD)
Two tone IMD is measured by applying two spectrally pure sinewaves to the ADC at
frequencies f1 and f2, usually relatively close together. The amplitude of each tone
is set slightly more than 6dB below full scale so that the ADC does not clip when the
two tones add in-phase. The location of the second and third-order products are
shown in Figure 2.30. Notice that the second-order products fall at frequencies
which can be removed by digital filters. However, the third-order products 2f2–f1
and 2f1–f2 are close to the original signals and are more difficult to filter. Unless
otherwise specified, two-tone IMD refers to these third-order products. The value of
the IMD product is expressed in dBc relative to the value of either of the two
original tones, and not to their sum.
Note, however, that if the two tones are close to fs/4, then the aliased third
harmonics of the fundamentals can make the identification of the actual 2f2–f1 and
2f1–f2 products difficult. This is because the third harmonic of fs/4 is 3fs/4, and the
alias occurs at fs – 3fs/4 = fs/4. Similarly, if the two tones are close to fs/3, the
aliased second harmonics may interfere with the measurement. The same reasoning
applies here; the second harmonic of fs/3 is 2 fs/3, and its alias occurs at fs – 2 fs/3 =
fs/3.
2.28
SAMPLED DATA SYSTEMS
SECOND AND THIRD-ORDER INTERMODULATION
PRODUCTS FOR f1 = 5MHz, f2 = 6MHz
2 = SECOND ORDER IMD PRODUCTS
f1
3
f2
= THIRD ORDER IMD PRODUCTS
NOTE: f1 = 5MHz, f2 = 6MHz
2
2
f2 - f1
f2 + f 1
2f 2
2f 1
3
3
2f1 - f2
1
4
5
2f2 - f1
6 7
10 11 12
3
3f 1
3
2f2 + f1
2f1 + f2
3f 2
15 16 17 18
FREQUENCY: MHz
Figure 2.30
The concept of second and third-order intercept points is not valid for an ADC,
because the distortion products do not vary in a predictable manner (as a function
of signal amplitude). The ADC does not gradually begin to compress signals
approaching full scale (there is no 1dB compression point); it acts as a hard limiter
as soon as the signal exceeds the ADC input range, thereby suddenly producing
extreme amounts of distortion because of clipping. On the other hand, for signals
much below full scale, the distortion floor remains relatively constant and is
independent of signal level.
Multi-tone SFDR is often measured in communications applications. The larger
number of tones more closely simulates the wideband frequency spectrum of cellular
telephone systems such as AMPS or GSM. Figure 2.31 shows the 4-tone
intermodulation performance of the AD6640 12-bit, 65MSPS ADC. High SFDR
increases the receiver’s ability to capture small signals in the presence of large ones,
and prevent the small signals from being masked by the intermodulation products
of the larger ones.
2.29
SAMPLED DATA SYSTEMS
MULTITONE TESTING:
AD6640 12-BIT, 65MSPS ADC
Figure 2.31
Noise Power Ratio (NPR)
Noise power ratio testing has been used extensively to measure the transmission
characteristics of Frequency Division Multiple Access (FDMA) communications
links (see Reference 4). In a typical FDMA system, 4kHz wide voice channels are
"stacked" in frequency bins for transmission over coaxial, microwave, or satellite
equipment. At the receiving end, the FDMA data is demultiplexed and returned to
4kHz individual baseband channels. In an FDMA system having more than
approximately 100 channels, the FDMA signal can be approximated by Gaussian
noise with the appropriate bandwidth. An individual 4kHz channel can be
measured for "quietness" using a narrow-band notch (bandstop) filter and a
specially tuned receiver which measures the noise power inside the 4kHz notch (see
Figure 2.32).
Noise Power Ratio (NPR) measurements are straightforward. With the notch filter
out, the rms noise power of the signal inside the notch is measured by the
narrowband receiver. The notch filter is then switched in, and the residual noise
inside the slot is measured. The ratio of these two readings expressed in dB is the
NPR. Several slot frequencies across the noise bandwidth (low, midband, and high)
are tested to characterize the system adequately. NPR measurements on ADCs are
made in a similar manner except the analog receiver is replaced by a buffer memory
and an FFT processor.
2.30
SAMPLED DATA SYSTEMS
NOISE POWER RATIO (NPR) MEASUREMENTS
GAUSSIAN
NOISE
SOURCE
GAUSSIAN
NOISE
SOURCE
LPF
NOTCH
FILTER
LPF
NOTCH
FILTER
TRANSMISSION
SYSTEM
NARROWBAND
RECEIVER
N
ADC
BUFFER
MEMORY
AND FFT
PROCESSOR
fs
RMS
NOISE
LEVEL
(dB)
NPR
FREQUENCY
0.5f s
Figure 2.32
NPR is usually plotted on an NPR curve. The NPR is plotted as a function of rms
noise level referred to the peak range of the system. For very low noise loading level,
the undesired noise (in non-digital systems) is primarily thermal noise and is
independent of the input noise level. Over this region of the curve, a 1dB increase in
noise loading level causes a 1dB increase in NPR. As the noise loading level is
increased, the amplifiers in the system begin to overload, creating intermodulation
products which cause the noise floor of the system to increase. As the input noise
increases further, the effects of "overload" noise predominate, and the NPR is
reduced dramatically. FDMA systems are usually operated at a noise loading level a
few dB below the point of maximum NPR.
In a digital system containing an ADC, the noise within the slot is primarily
quantization noise when low levels of noise input are applied. The NPR curve is
linear in this region. As the noise level increases, there is a one-for-one
correspondence between the noise level and the NPR. At some level, however,
"clipping" noise caused by the hard-limiting action of the ADC begins to dominate. A
theoretical curve for 10, 11, and 12-bit ADCs is shown in Figure 2.33 (see Reference
5).
In multi-channel high frequency communication systems, NPR can also be used to
simulate the distortion caused by a large number of individual channels, similar to
an FDMA system. A notch filter is placed between the noise source and the ADC,
and an FFT output is used in place of the analog receiver. The width of the notch
filter is set for several MHz as shown in Figure 2.34 for the AD9042. NPR is the
"depth" of the notch. An ideal ADC will only generate quantization noise inside the
notch, however a practical one has additional noise components due to
2.31
SAMPLED DATA SYSTEMS
intermodulation distortion caused by ADC non-linearity. Notice that the NPR is
about 60dB compared to 62.7dB theoretical.
THEORETICAL NPR FOR 10, 11, 12-BIT ADCs
ADC RANGE = ±VO
NPR
(dB)
62.7dB
V
k= O
σ
60
σ = RMS NOISE LEVEL
IT
S
57.1dB
11
-B
IT
S
12
-B
55
51.6dB
10
-B
IT
S
50
45
–30
–25
–20
–15
–10
RMS NOISE LOADING LEVEL = -20log(k) dB
Figure 2.33
AD9042 12-BIT, 41MSPS ADC NPR
MEASURES 60dB (62.7dB THEORETICAL)
POWER RELATIVE TO ADC FULL SCALE - dB
0
ENCODE = 41 MSPS
AIN = BROADBAND_NOISE
-20
-40
-60
-80
-100
-120
dc
4.1
8.2
12.3
FREQUENCY (MHz)
Figure 2.34
2.32
16.4
20.5
SAMPLED DATA SYSTEMS
Aperture Jitter and Aperture Delay
Another reason that the SNR of an ADC decreases with input frequency may be
deduced from Figure 2.35, which shows the effects of phase jitter (or aperture time
jitter) on the sampling clock of an ADC (or internal in the sample-and-hold). The
phase jitter causes a voltage error which is a function of slew rate and results in an
overall degradation in SNR as shown in Figure 2.36. This is quite serious, especially
at higher input/output frequencies. Therefore, extreme care must be taken to
minimize phase noise in the sampling/reconstruction clock of any sampled data
system. This care must extend to all aspects of the clock signal: the oscillator itself
(for example, a 555 timer is absolutely inadequate, but even a quartz crystal
oscillator can give problems if it uses an active device which shares a chip with
noisy logic); the transmission path (these clocks are very vulnerable to interference
of all sorts), and phase noise introduced in the ADC or DAC. A very common source
of phase noise in converter circuitry is aperture jitter in the integral
sample-and-hold (SHA) circuitry.
EFFECTS OF APERTURE AND SAMPLING CLOCK JITTER
dv
∆v =
dt
ANALOG
INPUT
dv
= SLOPE
dt
∆t
∆ v RMS = APERTURE JITTER ERROR
NOMINAL
HELD
OUTPUT
∆ t RMS = APERTURE JITTER
HOLD
TRACK
Figure 2.35
Two decades or so ago, sampling ADCs were built up from a separate SHA and
ADC. Interface design was difficult, and a key parameter was aperture jitter in the
SHA. Today, most sampled data systems use sampling ADCs which contain an
integral SHA. The aperture jitter of the SHA may not be specified as such, but this
2.33
SAMPLED DATA SYSTEMS
is not a cause of concern if the SNR or ENOB is clearly specified, since a guarantee
of a specific SNR is an implicit guarantee of an adequate aperture jitter
specification. However, the use of an additional high-performance SHA will
sometimes improve the high-frequency ENOB of a even the best sampling ADC by
presenting "DC" to the ADC, and may be more cost-effective than replacing the ADC
with a more expensive one.
It should be noted that there is also a fixed component which makes up the ADC
aperture time. This component, usually called effective aperture delay time, does not
produce an error. It simply results in a time offset between the time the ADC is
asked to sample and when the actual sample takes place (see Figure 2.37), and may
be positive or negative. The variation or tolerance placed on this parameter from
part to part is important in simultaneous sampling applications or other
applications such as I and Q demodulation where two ADCs are required to track
each other.
SNR DUE TO APERTURE AND SAMPLING CLOCK JITTER
tj = 1ps
100
SNR = 20log 10
16
1
π ft j
2π
14
tj = 10ps
80
SNR
(dB)
12
10
tj = 100ps
60
8
tj = 1ns
40
6
4
20
0
1
3
10
30
FULLSCALE SINEWAVE INPUT FREQUENCY (MHz)
Figure 2.36
2.34
100
ENOB
SAMPLED DATA SYSTEMS
EFFECTIVE APERTURE DELAY TIME
+FS
ZERO CROSSING
ANALOG INPUT
SINEWAVE
0V
-FS
-t e
+t e
SAMPLING
CLOCK
te
Figure 2.37
DAC DYNAMIC PERFORMANCE
The AC specifications which are most likely to be important with DACs are settling
time, glitch, distortion, and Spurious Free Dynamic Range (SFDR).
The settling time of a DAC is the time from a change of digital code to when the
output comes within and remains within some error band as shown in Figure 2.38.
With amplifiers, it is hard to make comparisons of settling time, since their error
bands may differ from amplifier to amplifier, but with DACs the error band will
almost invariably be ±1 or ±½ LSB.
The settling time of a DAC is made up of four different periods: the switching time
or dead time (during which the digital switching, but not the output, is changing),
the slewing time (during which the rate of change of output is limited by the slew
rate of the DAC output), the recovery time (when the DAC is recovering from its fast
slew and may overshoot), and the linear settling time (when the DAC output
approaches its final value in an exponential or near-exponential manner). If the
slew time is short compared to the other three (as is usually the case with current
output DACs), then the settling time will be largely independent of the output step
size. On the other hand, if the slew time is a significant part of the total, then the
larger the step, the longer the settling time.
2.35
SAMPLED DATA SYSTEMS
DAC SETTLING TIME
ERROR BAND
T=0
DEAD
TIME
SLEW
TIME
RECOVERY
TIME
LINEAR
SETTLING
Figure 2.38
Ideally, when a DAC output changes it should move from one value to its new one
monotonically. In practice, the output is likely to overshoot, undershoot, or both (see
Figure 2.39). This uncontrolled movement of the DAC output during a transition is
known as glitch. It can arise from two mechanisms: capacitive coupling of digital
transitions to the analog output, and the effects of some switches in the DAC
operating more quickly than others and producing temporary spurious outputs.
DAC TRANSITIONS (SHOWING GLITCH)
IDEAL TRANSITION
TRANSITION WITH
DOUBLET GLITCH
t
t
Figure 2.39
2.36
TRANSITION WITH
UNIPOLAR (SKEW)
GLITCH
t
SAMPLED DATA SYSTEMS
Capacitive coupling frequently produces roughly equal positive and negative spikes
(sometimes called a doublet glitch) which more or less cancel in the longer term. The
glitch produced by switch timing differences is generally unipolar, much larger and
of greater concern.
Glitches can be characterized by measuring the glitch impulse area, sometimes
inaccurately called glitch energy. The term glitch energy is a misnomer, since the
unit for glitch impulse area is Volt-seconds (or more probably µV-sec or pV-sec. The
peak glitch area is the area of the largest of the positive or negative glitch areas.
The glitch impulse area is the net area under the voltage-versus-time curve and can
be estimated by approximating the waveforms by triangles, computing the areas,
and subtracting the negative area from the positive area. The midscale glitch
produced by the transition between the codes 0111...111 and 1000...000 is usually
the worst glitch. Glitches at other code transition points (such as 1/4 and 3/4 full
scale) are generally less. Figure 2.40 shows the midscale glitch for a fast low-glitch
DAC. The peak and net glitch areas are estimated using triangles as described
above. Settling time is measured from the time the waveform leaves the initial
1LSB error band until it enters and remains within the final 1LSB error band. The
step size between the transition regions is also 1LSB.
DAC MIDSCALE GLITCH SHOWS 1.34pV-s
NET IMPULSE AREA AND SETTLING TIME OF 4.5ns
SETTLING TIME
= 4.5ns
NET GLITCH AREA = 1.34 pV-s
PEAK GLITCH AREA = 1.36 pV-s
2 mV/DIVISION
1 LSB
1 LSB
1 LSB
5 ns/DIVISION
≈ 4.5ns
Figure 2.40
2.37
SAMPLED DATA SYSTEMS
DAC settling time is important in applications such as RGB raster scan video
display drivers, but frequency-domain specifications such as SFDR are generally
more important in communications.
If we consider the spectrum of a waveform reconstructed by a DAC from digital
data, we find that in addition to the expected spectrum (which will contain one or
more frequencies, depending on the nature of the reconstructed waveform), there
will also be noise and distortion products. Distortion may be specified in terms of
harmonic distortion, Spurious Free Dynamic Range (SFDR), intermodulation
distortion, or all of the above. Harmonic distortion is defined as the ratio of
harmonics to fundamental when a (theoretically) pure sine wave is reconstructed,
and is the most common specification. Spurious free dynamic range is the ratio of
the worst spur (usually, but not necessarily always a harmonic of the fundamental)
to the fundamental.
Code-dependent glitches will produce both out-of-band and in-band harmonics when
the DAC is reconstructing a digitally generated sinewave as in a Direct Digital
Synthesis (DDS) system. The midscale glitch occurs twice during a single cycle of a
reconstructed sinewave (at each midscale crossing), and will therefore produce a
second harmonic of the sinewave, as shown in Figure 2.41. Note that the higher
order harmonics of the sinewave, which alias back into the Nyquist bandwidth (DC
to fs/2), cannot be filtered.
EFFECT OF CODE-DEPENDENT
GLITCHES ON SPECTRAL OUTPUT
+ FULL SCALE
fO = 3MHz
MIDSCALE
fS = 10MSPS
– FULL SCALE
AMPLITUDE
CANNOT
BE FILTERED
fO
fS – 2fO
0
1
2
3
4
f S – fO
2fO
5
6
fS
2
Figure 2.41
2.38
fS
7
8
FREQUENCY (MHz)
9
10
SAMPLED DATA SYSTEMS
It is difficult to predict the harmonic distortion or SFDR from the glitch area
specification alone. Other factors, such as the overall linearity of the DAC, also
contribute to distortion. It is therefore customary to test reconstruction DACs in the
frequency domain (using a spectrum analyzer) at various clock rates and output
frequencies as shown in Figure 2.43. Typical SFDR for the 14-bit AD9772 Transmit
DAC is shown in Figure 2.44. The clock rate is 65MSPS, and the output frequency is
swept to 25MHz. As in the case of ADCs, quantization noise will appear as
increased harmonic distortion if the ratio between the clock frequency and the DAC
output frequency is an integer number. These ratios should be avoided when
making the SFDR measurements.
CONTRIBUTORS TO DDS DAC DISTORTION
■ Resolution
■ Integral Non-Linearity
■ Differential Non-Linearity
■ Code-Dependent Glitches
■ Ratio of Clock Frequency to Output Frequency (Even in an Ideal
DAC)
■ Mathematical Analysis is Difficult !
Figure 2.42
2.39
SAMPLED DATA SYSTEMS
TEST SETUP FOR MEASURING DAC SFDR
PARALLEL OR
SERIAL PORT
DDS
SYSTEM
PC
N
N
LATCH
DAC
fo
STABLE
FREQUENCY
REFERENCE
fc
Figure 2.43
AD9772 14-BIT TxDAC™ SFDR,
DATA UPDATE RATE = 65MSPS
SFDR
(dBc)
DAC OUTPUT FREQUENCY (MHz)
Figure 2.44
2.40
SPECTRUM
ANALYZER
SAMPLED DATA SYSTEMS
DAC sin (x)/x Frequency Rolloff
The output of a reconstruction DAC can be visualized as a series of rectangular
pulses whose width is equal to the reciprocal of the clock rate as shown in Figure
2.45. Note that the reconstructed signal amplitude is down 3.92dB at the Nyquist
frequency, fc/2. An inverse sin(x)/x filter can be used to compensate for this effect in
most cases. The images of the fundamental signal are also attenuated by the
sin(x)/x function.
DAC SIN X/X ROLL OFF
(AMPLITUDE NORMALIZED)
SAMPLED
SIGNAL
t
RECONSTRUCTED
SIGNAL
t
1
–3.92dB
A
1
fc
sin
A=
IMAGES
IMAGES
πf
fc
πf
fc
IMAGES
f
0
0.5fc
fc
1.5fc
2fc
2.5fc
3fc
Figure 2.45
2.41
SAMPLED DATA SYSTEMS
REFERENCES
1.
Active and Passive Electrical Wave Filter Catalog, Vol. 34, TTE,
Incorporated, 2251 Barry Avenue, Los Angeles, CA 90064.
2.
W. R. Bennett, “Spectra of Quantized Signals”, Bell System Technical
Journal, No. 27, July 1948, pp. 446-472.
3.
Steve Ruscak and Larry Singer, Using Histogram Techniques to
Measure A/D Converter Noise, Analog Dialogue, Vol. 29-2, 1995.
4.
M.J. Tant, The White Noise Book, Marconi Instruments, July 1974.
5.
G.A. Gray and G.W. Zeoli, Quantization and Saturation Noise due
to A/D Conversion, IEEE Trans. Aerospace and Electronic
Systems, Jan. 1971, pp. 222-223.
6.
Chuck Lane, A 10-bit 60MSPS Flash ADC, Proceedings of the 1989
Bipolar Circuits and Technology Meeting, IEEE Catalog No.
89CH2771-4, September 1989, pp. 44-47.
7.
F.D. Waldhauer, Analog to Digital Converter, U.S. Patent
3-187-325, 1965.
8.
J.O. Edson and H.H. Henning, Broadband Codecs for an Experimental
224Mb/s PCM Terminal, Bell System Technical Journal, 44,
November 1965, pp. 1887-1940.
9.
J.S. Mayo, Experimental 224Mb/s PCM Terminals, Bell System
Technical Journal, 44, November 1965, pp. 1813-1941.
10.
Hermann Schmid, Electronic Analog/Digital Conversions,
Van Nostrand Reinhold Company, New York, 1970.
11.
Carl Moreland, An 8-bit 150MSPS Serial ADC, 1995 ISSCC Digest
of Technical Papers, Vol. 38, p. 272.
12.
Roy Gosser and Frank Murden, A 12-bit 50MSPS Two-Stage A/D
Converter, 1995 ISSCC Digest of Technical Papers, p. 278.
13.
Carl Moreland, An Analog-to-Digital Converter Using SerialRipple Architecture, Masters' Thesis, Florida State University
College of Engineering, Department of Electrical Engineering, 1995.
14.
Practical Analog Design Techniques, Analog Devices, 1995, Chapter
4, 5, and 8.
15.
Linear Design Seminar, Analog Devices, 1995, Chapter 4, 5.
2.42
SAMPLED DATA SYSTEMS
16.
System Applications Guide, Analog Devices, 1993, Chapter 12, 13,
15,16.
17.
Amplifier Applications Guide, Analog Devices, 1992, Chapter 7.
18.
Walt Kester, Drive Circuitry is Critical to High-Speed Sampling ADCs,
Electronic Design Special Analog Issue, Nov. 7, 1994, pp. 43-50.
19.
Walt Kester, Basic Characteristics Distinguish Sampling A/D Converters,
EDN, Sept. 3, 1992, pp. 135-144.
20.
Walt Kester, Peripheral Circuits Can Make or Break Sampling ADC
Systems, EDN, Oct. 1, 1992, pp. 97-105.
21.
Walt Kester, Layout, Grounding, and Filtering Complete Sampling
ADC System, EDN, Oct. 15, 1992, pp. 127-134.
22.
Robert A. Witte, Distortion Measurements Using a Spectrum Analyzer,
RF Design, September, 1992, pp. 75-84.
23.
Walt Kester, Confused About Amplifier Distortion Specs?, Analog
Dialogue, 27-1, 1993, pp. 27-29.
24.
System Applications Guide, Analog Devices, 1993, Chapter 16.
25.
Frederick J. Harris, On the Use of Windows for Harmonic Analysis
with the Discrete Fourier Transform, IEEE Proceedings, Vol. 66, No. 1,
Jan. 1978, pp. 51-83.
26.
Joey Doernberg, Hae-Seung Lee, David A. Hodges, Full Speed Testing
of A/D Converters, IEEE Journal of Solid State Circuits, Vol. SC-19,
No. 6, Dec. 1984, pp. 820-827.
27.
Brendan Coleman, Pat Meehan, John Reidy and Pat Weeks, Coherent
Sampling Helps When Specifying DSP A/D Converters, EDN, October 15,
1987, pp. 145-152.
28.
Robert W. Ramierez, The FFT: Fundamentals and Concepts,
Prentice-Hall, 1985.
29.
R. B. Blackman and J. W. Tukey, The Measurement of Power
Spectra, Dover Publications, New York, 1958.
30.
James J. Colotti, Digital Dynamic Analysis of A/D Conversion
Systems Through Evaluation Software Based on FFT/DFT Analysis,
RF Expo East 1987 Proceedings, Cardiff Publishing Co., pp. 245-272.
31.
HP Journal, Nov. 1982, Vol. 33, No. 11.
32.
HP Product Note 5180A-2.
2.43
SAMPLED DATA SYSTEMS
33.
HP Journal, April 1988, Vol. 39, No. 2.
34.
HP Journal, June 1988, Vol. 39, No. 3.
35.
Dan Sheingold, Editor, Analog-to-Digital Conversion Handbook,
Third Edition, Prentice-Hall, 1986.
36.
Lawrence Rabiner and Bernard Gold, Theory and Application of
Digital Signal Processing, Prentice-Hall, 1975.
37.
Matthew Mahoney, DSP-Based Testing of Analog and Mixed-Signal
Circuits, IEEE Computer Society Press, Washington, D.C., 1987.
38.
IEEE Trial-Use Standard for Digitizing Waveform Recorders,
No. 1057-1988.
39.
Richard J. Higgins, Digital Signal Processing in VSLI, Prentice-Hall,
1990.
40.
M. S. Ghausi and K. R. Laker, Modern Filter Design: Active RC and
Switched Capacitors, Prentice Hall, 1981.
41.
Mathcad™ 4.0 software package available from MathSoft, Inc.,
201 Broadway, Cambridge MA, 02139.
42.
Howard E. Hilton, A 10MHz Analog-to-Digital Converter with 110dB
Linearity, H.P. Journal, October 1993, pp. 105-112.
2.44
ADCS FOR DSP APPLICATIONS
SECTION 3
ADCs FOR DSP APPLICATIONS
■
Successive Approximation ADCs
■
Sigma-Delta ADCs
■
Flash Converters
■
Subranging (Pipelined) ADCs
■
Bit-Per-Stage (Serial, or Ripple) ADCs
3.a
ADCS FOR DSP APPLICATIONS
3.b
ADCS FOR DSP APPLICATIONS
SECTION 3
ADCs FOR DSP APPLICATIONS
Walt Kester, James Bryant
The trend in ADCs and DACs is toward higher speeds and higher resolutions at
reduced power levels and supply voltages. Modern data converters generally operate
on ±5V (dual supply), +5V or +3V (single supply). In fact, the number of +3V devices
is rapidly increasing because of many new markets such as digital cameras,
camcorders, and cellular telephones. This trend has created a number of design and
applications problems which were much less important in earlier data converters,
where ±15V supplies and ±10V input ranges were the standard.
Lower supply voltages imply smaller input voltage ranges, and hence more
susceptibility to noise from all potential sources: power supplies, references, digital
signals, EMI/RFI, and probably most important, improper layout, grounding, and
decoupling techniques. Single-supply ADCs often have an input range which is not
referenced to ground. Finding compatible single-supply drive amplifiers and dealing
with level shifting of the input signal in direct-coupled applications also becomes a
challenge.
In spite of these issues, components are now available which allow extremely high
resolutions at low supply voltages and low power. This section discusses the
applications problems associated with such components and shows techniques for
successfully designing them into systems.
The most popular ADCs for DSP applications are based on five fundamental
architectures: successive approximation, sigma-delta, flash, subranging (or
pipelined), and bit-per-stage (or ripple).
LOW POWER, LOW VOLTAGE ADC DESIGN ISSUES
■ Typical Supply Voltages: ±5V, +5V, +5/+3V, +3V
■ Lower Signal Swings Increase Sensitivity to
All Types of Noise (Device, Power Supply, Logic, etc.)
■ Device Noise Increases at Low Currents
■ Common Mode Input Voltage Restrictions
■ Input Buffer Amplifier Selection Critical
■ Auto-Calibration Modes Desirable at High Resolutions
Figure 3.1
3.1
ADCS FOR DSP APPLICATIONS
ADCs FOR DSP APPLICATIONS
■ Successive Approximation
◆ Resolutions to 16-bits
◆ Minimal Throughput Delay Time (No Output Latency,
"Single-Shot" Operation Possible
◆ Used in Multiplexed Data Acquisition Systems
■ Sigma-Delta
◆ Resolutions to 24-bits
◆ Excellent Differential Linearity
◆ Internal Digital Filter (Can be Linear Phase)
◆ Long Throughput Delay Time (Output Latency)
◆ Difficult to Multiplex Inputs Due to Digital Filter Settling Time
■ High Speed Architectures:
◆ Flash Converter
◆ Subranging or Pipelined
◆ Bit-Per-Stage (Ripple)
Figure 3.2
SUCCESSIVE APPROXIMATION ADCS
The successive approximation ADC has been the mainstay of signal conditioning for
many years. Recent design improvements have extended the sampling frequency of
these ADCs into the megahertz region. The use of internal switched capacitor
techniques along with auto calibration techniques extend the resolution of these
ADCs to 16-bits on standard CMOS processes without the need for expensive thinfilm laser trimming.
The basic successive approximation ADC is shown in Figure 3.3. It performs
conversions on command. On the assertion of the CONVERT START command, the
sample-and-hold (SHA) is placed in the hold mode, and all the bits of the successive
approximation register (SAR) are reset to "0" except the MSB which is set to "1".
The SAR output drives the internal DAC. If the DAC output is greater than the
analog input, this bit in the SAR is reset, otherwise it is left set. The next most
significant bit is then set to "1". If the DAC output is greater than the analog input,
this bit in the SAR is reset, otherwise it is left set. The process is repeated with each
bit in turn. When all the bits have been set, tested, and reset or not as appropriate,
the contents of the SAR correspond to the value of the analog input, and the
conversion is complete. These bit “tests” can form the basis of a serial output
version SAR-based ADC.
3.2
ADCS FOR DSP APPLICATIONS
The end of conversion is generally indicated by an end-of-convert (EOC), data-ready
(DRDY), or a busy signal (actually, not-BUSY indicates end of conversion). The
polarities and name of this signal may be different for different SAR ADCs, but the
fundamental concept is the same. At the beginning of the conversion interval, the
signal goes high (or low) and remains in that state until the conversion is
completed, at which time it goes low (or high). The trailing edge is generally an
indication of valid output data.
SUCCESSIVE APPROXIMATION ADC
CONVERT
START
TIMING
ANALOG
INPUT
COMPARATOR
EOC,
DRDY,
OR BUSY
SHA
SUCCESSIVE
APPROXIMATION
REGISTER
(SAR)
DAC
OUTPUT
Figure 3.3
An N-bit conversion takes N steps. It would seem on superficial examination that a
16-bit converter would have twice the conversion time of an 8-bit one, but this is not
the case. In an 8-bit converter, the DAC must settle to 8-bit accuracy before the bit
decision is made, whereas in a 16-bit converter, it must settle to 16-bit accuracy,
which takes a lot longer. In practice, 8-bit successive approximation ADCs can
convert in a few hundred nanoseconds, while 16-bit ones will generally take several
microseconds.
Notice that the overall accuracy and linearity of the SAR ADC is determined
primarily by the internal DAC. Until recently, most precision SAR ADCs used lasertrimmed thin-film DACs to achieve the desired accuracy and linearity. The thin-film
resistor trimming process adds cost, and the thin-film resistor values may be
affected when subjected to the mechanical stresses of packaging.
For these reasons, switched capacitor (or charge-redistribution) DACs have become
popular in newer SAR ADCs. The advantage of the switched capacitor DAC is that
the accuracy and linearity is primarily determined by photolithography, which in
turn controls the capacitor plate area and the capacitance as well as matching. In
3.3
ADCS FOR DSP APPLICATIONS
addition, small capacitors can be placed in parallel with the main capacitors which
can be switched in and out under control of autocalibration routines to achieve high
accuracy and linearity without the need for thin-film laser trimming. Temperature
tracking between the switched capacitors can be better than 1ppm/ºC, thereby
offering a high degree of temperature stability.
A simple 3-bit capacitor DAC is shown in Figure 3.4. The switches are shown in the
track, or sample mode where the analog input voltage, AIN, is constantly charging
and discharging the parallel combination of all the capacitors. The hold mode is
initiated by opening SIN, leaving the sampled analog input voltage on the capacitor
array. Switch SC is then opened allowing the voltage at node A to move as the bit
switches are manipulated. If S1, S2, S3, and S4 are all connected to ground, a
voltage equal to –AIN appears at node A. Connecting S1 to VREF adds a voltage
equal to VREF/2 to –AIN. The comparator then makes the MSB bit decision, and
the SAR either leaves S1 connected to VREF or connects it to ground depending on
the comparator output (which is high or low depending on whether the voltage at
node A is negative or positive, respectively). A similar process is followed for the
remaining two bits. At the end of the conversion interval, S1, S2, S3, S4, and SIN
are connected to AIN, SC is connected to ground, and the converter is ready for
another cycle.
3-BIT SWITCHED CAPACITOR DAC
BIT1
(MSB)
BIT2
SC
BIT3
(LSB)
A
CTOTAL = 2C
C
C/ 2
C/ 4
_
C/ 4
+
S1
S2
S3
S4
AIN
SIN
VREF
SWITCHES SHOWN IN TRACK (SAMPLE) MODE
Figure 3.4
Note that the extra LSB capacitor (C/4 in the case of the 3-bit DAC) is required to
make the total value of the capacitor array equal to 2C so that binary division is
accomplished when the individual bit capacitors are manipulated.
3.4
ADCS FOR DSP APPLICATIONS
The operation of the capacitor DAC (cap DAC) is similar to an R/2R resistive DAC.
When a particular bit capacitor is switched to VREF, the voltage divider created by
the bit capacitor and the total array capacitance (2C) adds a voltage to node A equal
to the weight of that bit. When the bit capacitor is switched to ground, the same
voltage is subtracted from node A.
Because of their popularity, successive approximation ADCs are available in a wide
variety of resolutions, sampling rates, input and output options, package styles, and
costs. It would be impossible to attempt to list all types, but Figure 3.5 shows a
number of recent Analog Devices' SAR ADCs which are representative. Note that
many devices are complete data acquisition systems with input multiplexers which
allow a single ADC core to process multiple analog channels.
RESOLUTION / CONVERSION TIME COMPARISON
FOR REPRESENTATIVE SINGLE-SUPPLY SAR ADCs
RESOLUTION
SAMPLING
RATE
POWER
CHANNELS
AD7472
12-BITS
1.5MSPS
9mW
1
AD7891
12-BITS
500kSPS
85mW
8
AD7858/59
12-BITS
200kSPS
20mW
8
AD7887/88
12-BITS
125kSPS
3.5mW
8
AD7856/57
14-BITS
285kSPS
60mW
8
AD7660
16-BITS
100kSPS
15mW
1
AD974
16-BITS
200kSPS
120mW
4
AD7664
16-BITS
570kSPS
150mW
1
Figure 3.5
While there are some variations, the fundamental timing of most SAR ADCs is
similar and relatively straightforward (see Figure 3.6). The conversion process is
initiated by asserting a CONVERT START signal. The CONVST signal is a
negative-going pulse whose positive-going edge actually initiates the conversion.
The internal sample-and-hold (SHA) amplifier is placed in the hold mode on this
edge, and the various bits are determined using the SAR algorithm. The negativegoing edge of the CONVST pulse causes the EOC or BUSY line to go high. When
the conversion is complete, the BUSY line goes low, indicating the completion of the
conversion process. In most cases the trailing edge of the BUSY line can be used as
an indication that the output data is valid and can be used to strobe the output data
into an external register. However, because of the many variations in terminology
3.5
ADCS FOR DSP APPLICATIONS
and design, the individual data sheet should always be consulted when using a
specific ADC.
TYPICAL SAR ADC TIMING
SAMPLE X
SAMPLE X+1
SAMPLE X+2
CONVST
CONVERSION
TIME
TRACK/
ACQUIRE
CONVERSION
TIME
TRACK/
ACQUIRE
EOC,
BUSY
OUTPUT
DATA
DATA
X
DATA
X+1
Figure 3.6
It should also be noted that some SAR ADCs require an external high frequency
clock in addition to the CONVERT START command. In most cases, there is no
need to synchronize the two. The frequency of the external clock, if required,
generally falls in the range of 1MHz to 30MHz depending on the conversion time
and resolution of the ADC. Other SAR ADCs have an internal oscillator which is
used to perform the conversions and only require the CONVERT START command.
Because of their architecture, SAR ADCs allow single-shot conversion at any
repetition rate from DC to the converter's maximum conversion rate.
In a SAR ADC, the output data for a sampled input is valid at the end of the
conversion interval for that sampled input. In other ADC architectures, such as
sigma-delta or the two-stage subranging architecture shown in Figure 3.7, this is
not the case. The subranging ADC shown in the figure is a two-stage pipelined or
subranging 12-bit converter. The first conversion is done by the 6-bit ADC which
drives a 6-bit DAC. The output of the 6-bit DAC represents a 6-bit approximation to
the analog input. Note that SHA2 delays the analog signal while the 6-bit ADC
makes its decision and the 6-bit DAC settles. The DAC approximation is then
subtracted from the analog signal from SHA2, amplified, and digitized by a 7-bit
ADC. The outputs of the two conversions are combined, and the extra bit used to
correct errors made in the first conversion. The typical timing associated with this
type of converter is shown in Figure 3.8. Note that the output data presented
immediately after sample X actually corresponds to sample X–2, i.e., there is a two
clock-cycle "pipeline" delay. The pipelined ADC architecture is generally associated
with high speed ADCs, and in most cases the pipeline delay, or latency, is not a
major system problem in most applications where this type of converter is used.
3.6
ADCS FOR DSP APPLICATIONS
12-BIT TWO-STAGE PIPELINED ADC ARCHITECTURE
ANALOG
INPUT
SHA
1
SHA
2
+
_
SAMPLING
CLOCK
TIMING
6-BIT
ADC
6-BIT
DAC
6
7-BIT
ADC
BUFFER
REGISTER
7
6
ERROR CORRECTION LOGIC
12
OUTPUT REGISTERS
OUTPUT DATA
12
Figure 3.7
TYPICAL PIPELINED ADC TIMING
SAMPLE X
SAMPLE X+1
SAMPLE X+2
SAMPLING
CLOCK
OUTPUT
DATA
DATA
X–2
DATA
X–1
DATA
X
ABOVE SHOWS TWO CLOCK-CYCLES PIPELINE DELAY
Figure 3.8
3.7
ADCS FOR DSP APPLICATIONS
Pipelined ADCs may have more than two clock-cycles latency depending on the
particular architecture. For instance, the conversion could be done in three, or four,
or perhaps even more pipelined stages causing additional latency in the output
data.
Therefore, if the ADC is to be used in an event-triggered (or single-shot) mode
where there must be a one-to-one time correspondence between each sample and the
corresponding data, then the pipeline delay can be troublesome, and the SAR
architecture is advantageous. Pipeline delay or latency can also be a problem in
high speed servo-loop control systems or multiplexed applications. In addition, some
pipelined converters have a minimum allowable conversion rate and must be kept
running to prevent saturation of internal nodes.
Switched capacitor SAR ADCs generally have unbuffered input circuits similar to
the circuit shown in Figure 3.9 for the AD7858/59 ADC. During the acquisition time,
the analog input must charge the 20pF equivalent input capacitance to the correct
value. If the input is a DC signal, then the source resistance, RS, in series with the
125Ω internal switch resistance creates a time constant. In order to settle to 12-bit
accuracy, approximately 9 time constants must be allowed for settling, and this
defines the minimum allowable acquisition time. (Settling to 14-bits requires about
10 time constants, and 16-bits requires about 11).
tACQ > 9 × (RS + 125)Ω × 20pF.
For example, if RS = 50Ω, the acquisition time per the above formula must be at
least 310ns.
For AC applications, a low impedance source should be used to prevent distortion
due to the non-linear ADC input circuit. In a single supply application, a fast
settling rail-to-rail op amp such as the AD820 should be used. Fast settling allows
the op amp to settle quickly from the transient currents induced on its input by the
internal ADC switches. In Figure 3.9, the AD820 drives a lowpass filter consisting
of the 50Ω series resistor and the 10nF capacitor (cutoff frequency approximately
320kHz). This filter removes high frequency components which could result in
aliasing and decreases the noise.
Using a single supply op amp in this application requires special consideration of
signal levels. The AD820 is connected in the inverting mode and has a signal gain of
–1. The noninverting input is biased at a common mode voltage of +1.3V with the
10.7kΩ/10kΩ divider, resulting in an output voltage of +2.6V for VIN= 0V, and
+0.1V for VIN = +2.5V. This offset is provided because the AD820 output cannot go
all the way to ground, but is limited to the VCESAT of the output stage NPN
transistor, which under these loading conditions is about 50mV. The input range of
the ADC is also offset by +100mV by applying the +100mV offset from the
412Ω/10kΩ divider to the AIN– input.
3.8
ADCS FOR DSP APPLICATIONS
DRIVING SWITCHED CAPACITOR INPUTS
OF AD7858/59 12-BIT, 200kSPS ADC
+3V TO +5V
10kΩ
Ω
VIN
10kΩ
Ω
VIN : 0V TO +2.5V
AIN+ : +2.6V TO +0.1V
CUTOFF
= 320kHz
_ 0.1µF
AVDD
125Ω
Ω
50Ω
Ω
10nF
+
412Ω
Ω
+100mV
125Ω
Ω
DVDD
AD7858/59
AD820
AIN+
T
20pF
CAP
DAC
H
+
0.1µF
10kΩ
Ω
0.1µF
VCM = +1.30V
10.7kΩ
Ω
+2.5V
10kΩ
Ω
AIN–
VREF
_
H
DGND
0.1µF
0.1µF
T
T = TRACK
H = HOLD
AGND
NOTE: ONLY ONE INPUT SHOWN
Figure 3.9
SIGMA-DELTA (Σ
Σ∆) ADCS
James M. Bryant
Sigma-Delta Analog-Digital Converters (Σ∆ ADCs) have been known for nearly
thirty years, but only recently has the technology (high-density digital VLSI) existed
to manufacture them as inexpensive monolithic integrated circuits. They are now
used in many applications where a low-cost, low-bandwidth, low-power,
high-resolution ADC is required.
There have been innumerable descriptions of the architecture and theory of Σ∆
ADCs, but most commence with a maze of integrals and deteriorate from there. In
the Applications Department at Analog Devices, we frequently encounter engineers
who do not understand the theory of operation of Σ∆ ADCs and are convinced, from
study of a typical published article, that it is too complex to comprehend easily.
There is nothing particularly difficult to understand about Σ∆ ADCs, as long as you
avoid the detailed mathematics, and this section has been written in an attempt to
clarify the subject. A Σ∆ ADC contains very simple analog electronics (a comparator,
voltage reference, a switch, and one or more integrators and analog summing
circuits), and quite complex digital computational circuitry. This circuitry consists of
a digital signal processor (DSP) which acts as a filter (generally, but not invariably,
3.9
ADCS FOR DSP APPLICATIONS
a low pass filter). It is not necessary to know precisely how the filter works to
appreciate what it does. To understand how a Σ∆ ADC works, familiarity with the
concepts of over-sampling, quantization noise shaping, digital filtering, and
decimation is required.
SIGMA-DELTA ADCs
■ Low Cost, High Resolution (to 24-bits)
■ Excellent DNL
■ Low Power, but Limited Bandwidth (Voiceband, Audio)
■ Key Concepts are Simple, but Math is Complex
◆ Oversampling
◆ Quantization Noise Shaping
◆ Digital Filtering
◆ Decimation
■ Ideal for Sensor Signal Conditioning
◆ High Resolution
◆ Self, System, and Auto Calibration Modes
■ Wide Applications in Voiceband and Audio Signal
Processing
Figure 3.10
Let us consider the technique of over-sampling with an analysis in the frequency
domain. Where a DC conversion has a quantization error of up to ½ LSB, a sampled
data system has quantization noise. A perfect classical N-bit sampling ADC has an
RMS quantization noise of q/√12 uniformly distributed within the Nyquist band of
DC to fs/2 (where q is the value of an LSB and fs is the sampling rate) as shown in
Figure 3.11A. Therefore, its SNR with a full-scale sinewave input will be
(6.02N + 1.76) dB. If the ADC is less than perfect, and its noise is greater than its
theoretical minimum quantization noise, then its effective resolution will be less
than N-bits. Its actual resolution (often known as its Effective Number of Bits or
ENOB) will be defined by
ENOB =
SNR − 176
. dB
.
6.02dB
If we choose a much higher sampling rate, Kfs (see Figure 3.11B), the RMS
quantization noise remains q/√12, but the noise is now distributed over a wider
bandwidth DC to Kfs/2. If we then apply a digital low pass filter (LPF) to the
output, we remove much of the quantization noise, but do not affect the wanted
3.10
ADCS FOR DSP APPLICATIONS
signal - so the ENOB is improved. We have accomplished a high resolution A/D
conversion with a low resolution ADC. The factor K is generally referred to as the
oversampling ratio. It should be noted at this point that oversampling has an added
benefit in that it relaxes the requirements on the analog antialiasing filter.
OVERSAMPLING, DIGITAL FILTERING,
NOISE SHAPING, AND DECIMATION
A
fs
QUANTIZATION
NOISE = q / 12
q = 1 LSB
Nyquist
Operation
ADC
B
Oversampling
+ Digital Filter
Kfs
+ Decimation
ADC
C
Kfs
Σ∆
MOD
fs
fs
2
DIGITAL FILTER
DIGITAL
DEC
FILTER
Oversampling
+ Noise Shaping
+ Digital Filter
+ Decimation
fs
REMOVED NOISE
fs
2
Kfs
2
fs
DIGITAL
DEC
FILTER
Kfs
REMOVED NOISE
fs
2
Kfs
2
Kfs
Figure 3.11
Since the bandwidth is reduced by the digital output filter, the output data rate may
be lower than the original sampling rate (Kfs) and still satisfy the Nyquist criterion.
This may be achieved by passing every Mth result to the output and discarding the
remainder. The process is known as "decimation" by a factor of M. Despite the
origins of the term (decem is Latin for ten), M can have any integer value, provided
that the output data rate is more than twice the signal bandwidth. Decimation does
not cause any loss of information (see Figure 3.11B).
If we simply use over-sampling to improve resolution, we must over-sample by a
factor of 22N to obtain an N-bit increase in resolution. The Σ∆ converter does not
need such a high over-sampling ratio because it not only limits the signal passband,
but also shapes the quantization noise so that most of it falls outside this passband
as shown in Figure 3.11C.
If we take a 1-bit ADC (generally known as a comparator), drive it with the output
of an integrator, and feed the integrator with an input signal summed with the
output of a 1-bit DAC fed from the ADC output, we have a first-order Σ∆ modulator
as shown in Figure 3.12. Add a digital low pass filter (LPF) and decimator at the
digital output, and we have a Σ∆ ADC: the Σ∆ modulator shapes the quantization
3.11
ADCS FOR DSP APPLICATIONS
noise so that it lies above the passband of the digital output filter, and the ENOB is
therefore much larger than would otherwise be expected from the over-sampling
ratio.
FIRST-ORDER SIGMA-DELTA ADC
CLOCK
Kfs
INTEGRATOR
VIN
+
∫
∑
fs
A
+
_
_
DIGITAL
FILTER
AND
DECIMATOR
LATCHED
COMPARATOR
(1-BIT ADC)
B
N-BITS
fs
+VREF
1-BIT
DAC
1-BIT DATA
STREAM
1-BIT,
Kfs
–VREF
SIGMA-DELTA MODULATOR
Figure 3.12
Intuitively, a Σ∆ ADC operates as follows. Assume a DC input at VIN. The
integrator is constantly ramping up or down at node A. The output of the
comparator is fed back through a 1-bit DAC to the summing input at node B. The
negative feedback loop from the comparator output through the 1-bit DAC back to
the summing point will force the average DC voltage at node B to be equal to VIN.
This implies that the average DAC output voltage must equal to the input voltage
VIN. The average DAC output voltage is controlled by the ones-density in the 1-bit
data stream from the comparator output. As the input signal increases towards
+VREF, the number of "ones" in the serial bit stream increases, and the number of
"zeros" decreases. Similarly, as the signal goes negative towards –VREF, the
number of "ones" in the serial bit stream decreases, and the number of "zeros"
increases. From a very simplistic standpoint, this analysis shows that the average
value of the input voltage is contained in the serial bit stream out of the
comparator. The digital filter and decimator process the serial bit stream and
produce the final output data.
The concept of noise shaping is best explained in the frequency domain by
considering the simple Σ∆ modulator model in Figure 3.13.
3.12
ADCS FOR DSP APPLICATIONS
SIMPLIFIED FREQUENCY DOMAIN LINEARIZED MODEL OF
A SIGMA-DELTA MODULATOR
X
ANALOG FILTER
H(f) = 1
f
∑
+
_
Q=
QUANTIZATION
NOISE
1 (X–Y)
f
X–Y
∑
Y
Y
Y=
1 (X–Y)
+ Q
f
REARRANGING, SOLVING FOR Y:
Y=
X
+
f+1
SIGNAL TERM
Qf
f+1
NOISE TERM
Figure 3.13
The integrator in the modulator is represented as an analog lowpass filter with a
transfer function equal to H(f) = 1/f. This transfer function has an amplitude
response which is inversely proportional to the input frequency. The 1-bit quantizer
generates quantization noise, Q, which is injected into the output summing block. If
we let the input signal be X, and the output Y, the signal coming out of the input
summer must be X – Y. This is multiplied by the filter transfer function, 1/f, and the
result goes to one input to the output summer. By inspection, we can then write the
expression for the output voltage Y as:
1
Y = (X − Y ) + Q .
f
This expression can easily be rearranged and solved for Y in terms of X, f, and Q:
Y=
X
Q⋅f
+
.
f +1 f +1
Note that as the frequency f approaches zero, the output voltage Y approaches X
with no noise component. At higher frequencies, the amplitude of the signal
component approaches zero, and the noise component approaches Q. At high
frequency, the output consists primarily of quantization noise. In essence, the
analog filter has a lowpass effect on the signal, and a highpass effect on the
quantization noise. Thus the analog filter performs the noise shaping function in the
Σ∆ modulator model.
3.13
ADCS FOR DSP APPLICATIONS
For a given input frequency, higher order analog filters offer more attenuation. The
same is true of Σ∆ modulators, provided certain precautions are taken.
By using more than one integration and summing stage in the Σ∆ modulator, we can
achieve higher orders of quantization noise shaping and even better ENOB for a
given over-sampling ratio as is shown in Figure 3.14 for both a first and secondorder Σ∆ modulator. The block diagram for the second-order Σ∆ modulator is shown
in Figure 3.15. Third, and higher, order Σ∆ ADCs were once thought to be
potentially unstable at some values of input - recent analyses using finite rather
than infinite gains in the comparator have shown that this is not necessarily so, but
even if instability does start to occur, it is not important, since the DSP in the
digital filter and decimator can be made to recognize incipient instability and react
to prevent it.
Figure 3.16 shows the relationship between the order of the Σ∆ modulator and the
amount of over-sampling necessary to achieve a particular SNR. For instance, if the
oversampling ratio is 64, an ideal second-order system is capable of providing an
SNR of about 80dB. This implies approximately 13 effective number of bits (ENOB).
Although the filtering done by the digital filter and decimator can be done to any
degree of precision desirable, it would be pointless to carry more than 13 binary bits
to the outside world. Additional bits would carry no useful signal information, and
would be buried in the quantization noise unless post-filtering techniques were
employed. Additional resolution can be obtained by increasing the oversampling
ratio and/or by using a higher-order modulator.
SIGMA-DELTA MODULATORS
SHAPE QUANTIZATION NOISE
2ND ORDER
DIGITAL
FILTER
1ST ORDER
Kfs
2
fs
2
Figure 3.14
3.14
ADCS FOR DSP APPLICATIONS
SECOND-ORDER SIGMA-DELTA ADC
INTEGRATOR
VIN
+
∫
∑
+
∫
∑
_
_
CLOCK
Kfs
INTEGRATOR
+
_
1-BIT
DATA
STREAM
1-BIT
DAC
DIGITAL FILTER
AND
DECIMATOR
N-BITS
fs
Figure 3.15
SNR VERSUS OVERSAMPLING RATIO FOR FIRST,
SECOND, AND THIRD-ORDER LOOPS
120
THIRD-ORDER LOOP*
21dB / OCTAVE
100
SECOND-ORDER LOOP
15dB / OCTAVE
80
SNR
(dB)
60
FIRST-ORDER LOOP
9dB / OCTAVE
40
* > 2nd ORDER LOOPS DO NOT
OBEY LINEAR MODEL
20
0
4
8
16
32
64
128
256
OVERSAMPLING RATIO, K
Figure 3.16
3.15
ADCS FOR DSP APPLICATIONS
The AD1877 is a 16-bit 48kSPS stereo sigma-delta DAC suitable for demanding
audio applications. Key specifications are summarized in Figure 3.17. This device
has a 64X oversampling ratio, and a fourth-order modulator. The internal digital
filter is a linear phase FIR filter whose response is shown in Figure 3.18. The
passband ripple is 0.006dB, and the attenuation is greater than 90dB in the
stopband. The width of the transition region from passband to stopband is only
0.1fs, where fs is the effective sampling frequency of the AD1877 (maximum of
48kSPS). Such a filter would obviously be impossible to implement in analog form.
AD1877 16-BIT, 48kSPS STEREO SIGMA-DELTA ADC
■ Single +5 V Power Supply
■ Single-Ended Dual-Channel Analog Inputs
■ 92 dB (typ) Dynamic Range
■ 90 dB (typ) S/(THD+N)
■ 0.006 dB Decimator Passband Ripple
■ Fourth-Order, 64-Times Oversampling Σ∆ Modulator
■ Three-Stage, Linear-Phase Decimator
■ Less than 100 mW (typ)
■ Power-Down Mode
■ Input Overrange Indication
■ On-Chip Voltage Reference
■ Flexible Serial Output Interface
■ 28-Pin SOIC Package
Figure 3.17
All sigma-delta ADCs have a settling time associated with the internal digital filter,
and there is no way to remove it. In multiplexed applications the input to the ADC
is a step function if there are different input voltages on adjacent channels. In fact,
the multiplexer output can represent a fullscale step voltage to the sigma-delta ADC
when channels are switched. Adequate filter settling time must be allowed,
therefore, in such applications. This does not mean that sigma-delta ADCs
shouldn’t be used in multiplexed applications, just that the settling time of the
digital filter must be considered.
For example, the group delay through the AD1877 FIR filter is 36/fs, and
represents the time it takes for a step function input to propagate through one-half
the number of taps in the digital filter. The total time required for settling is
therefore 72fs, or approximately 1.5ms when sampling at 48kSPS with a 64X
oversampling rate.
3.16
ADCS FOR DSP APPLICATIONS
AD1877 16-BIT, 48kSPS STEREO SIGMA-DELTA ADC
FIR FILTER CHARACTERISTICS
NORMALIZED fs
■
■
■
■
fs = OUTPUT WORD RATE, 32kSPS, 44.1kSPS, OR 48kSPS TYPICAL
PASSBAND TO STOPBAND TRANSITION REGION: 0.45fs TO 0.55fs
SETTLING TIME = 72 / fs = 1.5ms FOR fs = 48kSPS
GROUP DELAY = 36 / fs = 0.75ms FOR fs = 48kSPS
Figure 3.18
In other applications, such as low frequency, high resolution 24-bit measurement
sigma-delta ADCs (such as the AD77xx-series), other types of digital filters may be
used. For instance, the SINC3 response is popular because it has zeros at multiples
of the throughput rate. For instance a 10Hz throughput rate produces zeros at
50Hz and 60Hz which aid in AC power line rejection.
So far we have considered only sigma-delta converters which contain a single-bit
ADC (comparator) and a single-bit DAC (switch). The block diagram of Figure 3.19
shows a multi-bit sigma-delta ADC which uses an n-bit flash ADC and an n-bit
DAC. Obviously, this architecture will give a higher dynamic range for a given
oversampling ratio and order of loop filter. Stabilization is easier, since second-order
and higher loops can be used. Idling patterns tend to be more random thereby
minimizing tonal effects.
The real disadvantage of this technique is that the linearity depends on the DAC
linearity, and thin film laser trimming is generally required to approach 16-bit
performance levels. This makes the multi-bit architecture extremely difficult to
implement on sigma-delta ADCs. It is, however, currently used in sigma-delta audio
DACs (AD1852, AD1853, AD1854) where special “bit scrambling” techniques are
used to ensure linearity and eliminate idle tones.
3.17
ADCS FOR DSP APPLICATIONS
MULTI-BIT SIGMA-DELTA ADC
CLOCK
Kfs
fs
INTEGRATOR
VIN
+
∑
∫
FLASH
ADC
n-BITS
_
n-BIT
DAC
DIGITAL
FILTER
AND
DECIMATOR
N-BITS
fs
n-BITS,
n-BIT DATA
STREAM
Kfs
Figure 3.19
The Σ∆ ADCs that we have described so far contain integrators, which are low pass
filters, whose passband extends from DC. Thus, their quantization noise is pushed
up in frequency. At present, most commercially available Σ∆ ADCs are of this type
(although some which are intended for use in audio or telecommunications
applications contain bandpass rather than lowpass digital filters to eliminate any
system DC offsets). But there is no particular reason why the filters of the Σ∆
modulator should be LPFs, except that traditionally ADCs have been thought of as
being baseband devices, and that integrators are somewhat easier to construct than
bandpass filters. If we replace the integrators in a Σ∆ ADC with bandpass filters
(BPFs) as shown in Figure 3.20, the quantization noise is moved up and down in
frequency to leave a virtually noise-free region in the pass-band (see Reference 1). If
the digital filter is then programmed to have its pass-band in this region, we have a
Σ∆ ADC with a bandpass, rather than a lowpass characteristic. Such devices would
appear to be useful in direct IF-to-digital conversion, digital radios, ultrasound, and
other undersampling applications. However, the modulator and the digital BPF
must be designed for the specific set of frequencies required by the system
application, thereby somewhat limiting the flexibility of this approach.
In an undersampling application of a bandpass Σ∆ ADC, the minimum sampling
frequency must be at least twice the signal bandwidth, BW. The signal is centered
around a carrier frequency, fc. A typical digital radio application using a 455kHz
center frequency and a signal bandwidth of 10kHz is described in Reference 1. An
oversampling frequency Kfs = 2MSPS and an output rate fs = 20kSPS yielded a
dynamic range of 70dB within the signal bandwidth.
3.18
ADCS FOR DSP APPLICATIONS
REPLACING INTEGRATORS WITH RESONATORS
GIVES A BANDPASS SIGMA-DELTA ADC
+
Σ
-
ANALOG
BPF
+
CLOCK
Kfs
Σ
-
ANALOG
BPF
DIGITAL
BPF AND
DECIMATOR
fs
1-BIT
DAC
fc
DIGITAL BPF
RESPONSE
fs > 2 BW
BW
SHAPED
QUANTIZATION
NOISE
f
Figure 3.20
Most sigma-delta ADCs generally have a fixed internal digital filter. The filter’s
cutoff frequency and the ADC output data rate scales with the master clock
frequency. The AD7725 is a 16-bit sigma-delta ADC with a programmable internal
digital filter. The modulator operates at a maximum oversampling rate of
19.2MSPS. The modulator is followed by a preset FIR filter which decimates the
modulator output by a factor of 8, yielding an output data rate of 2.4MSPS. The
output of the preset filter drives a programmable FIR filter. By loading the ROM
with suitable coefficient values, this filter can be programmed for the desired
frequency response.
The programmable filter is flexible with respect to number of taps and decimation
rate. The filter can have up to 108 taps, up to 5 decimation stages, and a decimation
factor between 2 and 256. Coefficient precision is 24-bits, and arithmetic precision is
30-bits.
The AD7725 contains Systolix’s PuldeDSP™ (trademark of Systolix) post processor
which permits the filter characteristics to be programmed through the parallel or
serial microprocessor interface. Or, it may boot at power-on-reset from its internal
ROM or from an external EPROM.
3.19
ADCS FOR DSP APPLICATIONS
The post processor is a fully programmable core which provides processing power of
up to 130 million multiply-accumulates (MAC) per second. To program the post
processor, the user must produce a configuration file which contains the
programming data for the filter function. This file is generated by a compiler which
is available from Analog Devices. The AD7725 compiler accepts filter coefficient
data as an input and automatically generates the required device programming
data.
The coefficient file for the FIR filter response can be generated using a digital filter
design package such as QEDesign from Momentum Data Systems. The response of
the filter can be plotted so the user knows the response before generating the filter
coefficients. The data is available to the processor at a 2.4MSPS rate. When
decimation is employed in a multistage filter, the first filter will be operated at
2.4MSPS, and the user can then decimate between stages. The number of taps
which can be contained in the processor is 108. Therefore, a single filter with 108
taps can be generated, or a multistage filter can be designed whereby the total
number of taps adds up to 108. The filter characteristic can be lowpass, highpass,
bandstop, or bandpass.
The AD7725 operates on a single +5V supply, has an on-chip 2.5V reference, and is
packaged in a 44-pin PQFP. Power dissipation is approximately 350mW when
operating at full power. A half-power mode is available with a master clock
frequency of 10MSPS maximum. Power consumption in the standby mode is 200mW
maximum. More details of the AD7725 operation can be found in Section 9.
Summary
A Σ∆ ADC works by over-sampling, where simple analog filters in the Σ∆ modulator
shape the quantization noise so that the SNR in the bandwidth of interest is much
greater than would otherwise be the case, and by using high performance digital
filters and decimation to eliminate noise outside the required passband.
Oversampling has the added benefit of relaxing the requirements on the
antialiasing filter. Because the analog circuitry is relatively undemanding, it may be
built with the same digital VLSI process that is used to fabricate the DSP circuitry
of the digital filter. Because the basic ADC is 1-bit (a comparator), the technique is
inherently linear.
Although the detailed analysis of Σ∆ ADCs involves quite complex mathematics,
their basic design can be understood without the necessity of any mathematics at
all. For further discussion on Σ∆ ADCs, refer to References 1 through 18 at the end
of this section.
3.20
ADCS FOR DSP APPLICATIONS
SIGMA-DELTA SUMMARY
■ Inherently Excellent Linearity
■ Oversampling Relaxes Analog Antialiasing Filter Requirements
■ Ideal for Mixed-Signal IC Processes, no Trimming
■ No SHA Required
■ Added Functionality: On-Chip PGAs, Analog Filters,
Autocalibration
■ On-Chip Programmable Digital Filters (AD7725: Lowpass,
Highpass, Bandpass, Bandstop)
■ Upper Sampling Rate Currently Limits Applications to
Measurement, Voiceband, and Audio, but Bandpass Sigma-Delta
Techniques May Change This
■ Analog Multiplexer Switching Speed Limited by Internal Filter
Settling Time. Consider One Sigma-Delta ADC per Channel.
Figure 3.21
FLASH CONVERTERS
Flash ADCs (sometimes called parallel ADCs) are the fastest type of ADC and use
large numbers of comparators. An N-bit flash ADC consists of 2N resistors and 2N–1
comparators arranged as in Figure 3.22. Each comparator has a reference voltage
which is 1 LSB higher than that of the one below it in the chain. For a given input
voltage, all the comparators below a certain point will have their input voltage
larger than their reference voltage and a "1" logic output, and all the comparators
above that point will have a reference voltage larger than the input voltage and a
"0" logic output. The 2N–1 comparator outputs therefore behave in a way analogous
to a mercury thermometer, and the output code at this point is sometimes called a
thermometer code. Since 2N–1 data outputs are not really practical, they are
processed by a decoder to an N-bit binary output.
The input signal is applied to all the comparators at once, so the thermometer
output is delayed by only one comparator delay from the input, and the encoder
N-bit output by only a few gate delays on top of that, so the process is very fast.
However, the architecture uses large numbers of resistors and comparators and is
limited to low resolutions, and if it is to be fast, each comparator must run at
relatively high power levels. Hence, the problems of flash ADCs include limited
resolution, high power dissipation because of the large number of high speed
comparators (especially at sampling rates greater than 50MSPS), and relatively
large (and therefore expensive) chip sizes. In addition, the resistance of the
reference resistor chain must be kept low to supply adequate bias current to the fast
comparators, so the voltage reference has to source quite large currents (>10 mA).
3.21
ADCS FOR DSP APPLICATIONS
In practice, flash converters are available up to 10-bits, but more commonly they
have 8-bits of resolution. Their maximum sampling rate can be as high as 1GHz,
with input full-power bandwidths in excess of 300 MHz.
FLASH OR PARALLEL ADC
STROBE
ANALOG
INPUT
+VREF
1.5R
R
R
PRIORITY
ENCODER
AND LATCH
R
N
DIGITAL
OUTPUT
R
R
R
0.5R
Figure 3.22
But as mentioned earlier, full-power bandwidths are not necessarily full-resolution
bandwidths. Ideally, the comparators in a flash converter are well matched both for
DC and AC characteristics. Because the strobe is applied to all the comparators
simultaneously, the flash converter is inherently a sampling converter. In practice,
there are delay variations between the comparators and other AC mismatches
which cause a degradation in ENOB at high input frequencies. This is because the
inputs are slewing at a rate comparable to the comparator conversion time.
The input to a flash ADC is applied in parallel to a large number of comparators.
Each has a voltage-variable junction capacitance, and this signal-dependent
capacitance results in most flash ADCs having reduced ENOB and higher distortion
at high input frequencies.
Adding 1 bit to the total resolution of a flash converter requires doubling the
number of comparators! This limits the practical resolution of high speed flash
converters to 8-bits because of excessive power dissipation.
3.22
ADCS FOR DSP APPLICATIONS
However, in the AD9410 10-bit, 200MSPS ADC, a technique called interpolation is
used to minimize the number of preamplifiers in the flash converter comparators
and also reduce the power (1.8W). The method is shown in Figure 3.23.
“INTERPOLATING” FLASH REDUCES THE NUMBER
OF PREAMPLIFIERS BY FACTOR OF TWO
ANALOG
INPUT
B
+
A2
V2
B
-
V2
LATCH
2
V1A
DECODE
V1A = V1 + V2
2
A1
V1
-
LATCH
1A
A
+
A
ANALOG
INPUT
V1
B
A
A
B
A
B
B
A
LATCH
1
LATCH
STROBE
AD9410: 10-Bits, 200MSPS
Figure 3.23
The preamplifiers (labeled “A1”, “A2”, etc.) are low-gain gm stages whose bandwidth
is proportional to the tail currents of the differential pairs. Consider the case for a
positive-going ramp input which is initially below the reference to AMP A1, V1. As
the input signal approaches V1, the differential output of A1 approaches zero (i.e., A
= A ), and the decision point is reached. The output of A1 drives the differential
input of LATCH 1. As the input signals continues to go positive, A continues to go
positive, and B begins to go negative. The interpolated decision point is determined
when A = B . As the input continues positive, the third decision point is reached
when B = B . This novel architecture reduces the ADC input capacitance and
thereby minimizes its change with signal level and the associated distortion. The
AD9410 also uses an input sample-and-hold circuit for improved AC linearity.
SUBRANGING (PIPELINED) ADCS
Although it is not practical to make flash ADCs with high resolution (greater than
10-bits), flash ADCs are often used as subsystems in "subranging" ADCs (sometimes
known as "half-flash ADCs"), which are capable of much higher resolutions (up to
16-bits).
3.23
ADCS FOR DSP APPLICATIONS
A block diagram of an 8-bit subranging ADC based upon two 4-bit flash converters
is shown in Figure 3.24. Although 8-bit flash converters are readily available at
high sampling rates, this example will be used to illustrate the theory. The
conversion process is done in two steps. The first four significant bits (MSBs) are
digitized by the first flash (to better than 8-bits accuracy), and the 4-bit binary
output is applied to a 4-bit DAC (again, better than 8-bit accurate). The DAC output
is subtracted from the held analog input, and the resulting residue signal is
amplified and applied to the second 4-bit flash. The outputs of the two 4-bit flash
converters are then combined into a single 8-bit binary output word. If the residue
signal range does not exactly fill the range of the second flash converter, nonlinearities and perhaps missing codes will result.
8-BIT SUBRANGING ADC
ANALOG
INPUT
GAIN
+
SHA
RESIDUE
SIGNAL
-
4-BIT
FLASH
4-BIT
FLASH
4-BIT
DAC
4
4
OUTPUT REGISTER
8
Figure 3.24
Modern subranging ADCs use a technique called digital correction to eliminate
problems associated with the architecture of Figure 3.24. A simplified block diagram
of a 12-bit digitally corrected subranging (DCS) ADC is shown in Figure 3.25. The
architecture is similar to that used in the AD6640 12-bit, 65MSPS ADC. Note that a
6-bit and an 7-bit ADC have been used to achieve an overall 12-bit output. These
are not flash ADCs, but utilize a magnitude-amplifier (MagAmp™) architecture
which will be described shortly.
If there were no errors in the first-stage conversion, the 6-bit "residue" signal
applied to the 7-bit ADC by the summing amplifier would never exceed one-half of
the range of the 7-bit ADC. The extra range in the second ADC is used in
conjunction with the error correction logic (usually just a full adder) to correct the
output data for most of the errors inherent in the traditional uncorrected
subranging converter architecture. It is important to note that the 6-bit DAC must
3.24
ADCS FOR DSP APPLICATIONS
be better than 12-bit accurate, because the digital error correction does not correct
for DAC errors. In practice, "thermometer" or "fully-decoded" DACs using one
current switch per level (63 switches in the case of a 6-bit DAC) are often used
instead of a "binary" DAC to ensure excellent differential and integral linearity and
minimum switching transients.
AD6640 12-BIT, 65MSPS PIPELINED SUBRANGING ADC
WITH DIGITAL ERROR CORRECTION
ANALOG
INPUT
SHA
2
SHA
1
GAIN
+
-
6-BIT
DAC
6-BIT
ADC
SHA
3
7-BIT
ADC
6
BUFFER
REGISTER
7
6
ERROR CORRECTION LOGIC
12
OUTPUT REGISTER
12
Figure 3.25
The second SHA delays the held output of the first SHA while the first-stage
conversion occurs, thereby maximizing throughput. The third SHA serves to
deglitch the residue output signal, thereby allowing a full conversion cycle for the 7bit ADC to make its decision (the 6 and 7-bit ADCs in the AD6640 are bit-serial
MagAmp ADCs which require more settling time than a flash converter).
This multi-stage conversion technique is sometimes referred to as "pipelining."
Additional shift registers in series with the digital outputs of the first-stage ADC
ensure that its output is ultimately time-aligned with the last 7 bits from the
second ADC when their outputs are combined in the error correction logic. A
pipelined ADC therefore has a specified number of clock cycles of latency, or pipeline
delay associated with the output data. The leading edge of the sampling clock (for
sample N) is used to clock the output register, but the data which appears as a
result of that clock edge corresponds to sample N – L, where L is the number of
clock cycles of latency. In the case of the AD6640, there are two clock cycles of
latency.
3.25
ADCS FOR DSP APPLICATIONS
The error correction scheme described above is designed to correct for errors made
in the first conversion. Internal ADC gain, offset, and linearity errors are corrected
as long as the residue signal falls within the range of the second-stage ADC. These
errors will not affect the linearity of the overall ADC transfer characteristic. Errors
made in the final conversion, however, do translate directly as errors in the overall
transfer function. Also, linearity errors or gain errors either in the DAC or the
residue amplifier will not be corrected and will show up as nonlinearities or nonmonotonic behavior in the overall ADC transfer function.
So far, we have considered only two-stage subranging ADCs, as these are easiest to
analyze. There is no reason to stop at two stages, however. Three-pass and four-pass
subranging pipelined ADCs are quite common, and can be made in many different
ways, usually with digital error correction.
A simplified block diagram of the AD9220 12-bit, 10MSPS single-supply, 250mW
CMOS ADC is shown in Figure 3.26. The AD9221 (1.25MSPS, 60mW) and the
AD9223 (3MSPS, 100mW) ADCs use the identical architecture but operate at lower
power and lower sampling rates. This is a four-stage pipelined architecture with an
additional bit in the second, third, and fourth stage for error correction. Because of
the pipelined architecture, these ADCs have a 3 clock-cycle latency (see Figure
3.27).
AD9220/9221/9223 12-BIT PIPELINED CMOS ADC
ANALOG
INPUT
SHA
1
SHA
2
+
-
5-BIT 5-BIT
ADC DAC
5
SHA
3
+
SHA
4
-
4-BIT 4-BIT
ADC DAC
3-BIT 3-BIT
ADC DAC
4
3
BUFFER REGISTERS AND
ERROR CORRECTION LOGIC
12
OUTPUT REGISTERS
12
Figure 3.26
3.26
+
-
3-BIT
ADC
3
ADCS FOR DSP APPLICATIONS
LATENCY (PIPELINE DELAY)
OF AD9220/9221/9223 ADC
ANALOG
INPUT
N+1
N
N+2
N+3
SAMPLING
CLOCK
OUTPUT
DATA
DATA N - 2
DATA N - 3
DATA N - 1
DATA N
Figure 3.27
BIT-PER-STAGE (SERIAL, OR RIPPLE) ADCS
Various architectures exist for performing A/D conversion using one stage per bit. In
fact, a multistage subranging ADC with one bit per stage and no error correction is
one form. Figure 3.28 shows the overall concept. The SHA holds the input signal
constant during the conversion cycle. There are N stages, each of which have a bit
output and a residue output. The residue output of one stage is the input to the
next. The last bit is detected with a single comparator as shown.
BIT-PER-STAGE, SERIAL, OR RIPPLE ADC
VREF
ANALOG
INPUT
SHA
STAGE
1
BIT 1
MSB
R1
STAGE
2
R2
BIT 2
STAGE
N-1
BIT N-1
+
-
BIT N
LSB
DECODE LOGIC AND OUTPUT REGISTERS
N
Figure 3.28
3.27
ADCS FOR DSP APPLICATIONS
The basic stage for performing a single binary bit conversion is shown in Figure
3.29. It consists of a gain-of-two amplifier, a comparator, and a 1-bit DAC. Assume
that this is the first stage of the ADC. The MSB is simply the polarity of the input,
and that is detected with the comparator which also controls the 1-bit DAC. The 1bit DAC output is summed with the output of the gain-of-two amplifier. The
resulting residue output is then applied to the next stage. In order to better
understand how the circuit works, the diagram shows the residue output for the
case of a linear ramp input voltage which traverses the entire ADC range, –VR to
+VR. Notice that the polarity of the residue output determines the binary bit output
of the next stage.
SINGLE-STAGE OF BINARY ADC
+VR
RESIDUE
INPUT
Σ
G=2
INPUT
t
0
+
-VR
-
+VR
+VR
-VR
SWITCH POSITION
SHOWN FOR
NEGATIVE INPUT
RESIDUE
0
BIT OUTPUT
(BINARY CODE)
t
-VR
Figure 3.29
A simplified 3-bit serial-binary ADC is shown in Figure 3.30, and the residue
outputs are shown in Figure 3.31. Again, the case is shown for a linear ramp input
voltage whose range is between –VR and +VR. Each residue output signal has
discontinuities which correspond to the point where the comparator changes state
and causes the DAC to switch. The fundamental problem with this architecture is
the discontinuity in the residue output waveforms. Adequate settling time must be
allowed for these transients to propagate through all the stages and settle at the
final comparator input. The prospects of making this architecture operate at high
speed are therefore dismal.
3.28
ADCS FOR DSP APPLICATIONS
3-BIT SERIAL ADC WITH BINARY OUTPUT
±VR
ANALOG
INPUT
R1
R2
STAGE
2
STAGE
1
SHA
BIT 1
+
-
BIT 2
BIT 3
OUTPUT REGISTER
3
Figure 3.30
INPUT AND RESIDUE WAVEFORMS OF
3-BIT BINARY RIPPLE ADC
+V R
t
0
INPUT
-VR
+V R
R1
t
0
-VR
+V R
R2
t
0
-V R
BINARY
CODE
000
001
010
011
100
101
110
111
Figure 3.31
A much better bit-per-stage architecture was developed by F.D. Waldhauer
(Reference 21) based on absolute value amplifiers (magnitude amplifiers, or simply
MagAmps™). This scheme has often been referred to as serial-Gray (since the
output coding is in Gray code), or folding converter (References 22, 23, 24). The
basic stage is shown functionally in Figure 3.32 along with its transfer function. The
input to the stage is assumed to be a linear ramp voltage whose range is between
3.29
ADCS FOR DSP APPLICATIONS
–VR and +VR. The comparator detects the polarity of the input signal and provides
the Gray bit output for the stage. It also determines whether the overall stage gain
is +2 or –2. The reference voltage VR is summed with the switch output to generate
the residue signal which is applied to the next stage. The polarity of the residue
signal determines the Gray bit for the next stage. The transfer function for the
folding stage is also shown in Figure 3.32.
MagAmp STAGE FUNCTIONAL EQUIVALENT CIRCUIT
VR
INPUT
+VR
G = +2
INPUT
Σ
RESIDUE
0
t
G = -2
SWITCH POSITION
SHOWN FOR
NEGATIVE INPUT
-VR
+VR
+
RESIDUE
0
t
BIT OUTPUT
(GRAY CODE)
-VR
Figure 3.32
A 3-bit MagAmp folding ADC is shown in Figure 3.33, and the corresponding
residue waveforms in Figure 3.34. As in the case of the binary ripple ADC, the
polarity of the residue output signal of a stage determines the value of the Gray bit
for the next stage. The polarity of the input to the first stage determines the Gray
MSB; the polarity of R1 output determines the Gray bit-2; and the polarity of R2
output determines the Gray bit-3. Notice that unlike the binary ripple ADC, there is
no abrupt transition in any of the folding stage residue output waveforms. This
makes operation at high speeds quite feasible.
The key to operating this architecture at high speeds is the folding stage. Early
designs (see References 22, 23, 24) used discrete op amps with diodes inside the
feedback loop to generate the folding transfer function. Modern IC circuit designs
implement the transfer function using current-steering open-loop gain techniques
which can be made to operate much faster. Fully differential stages (including the
SHA) also provide speed, lower distortion, and yield 8-bit accurate folding stages
with no requirement for thin film resistor laser trimming (see References 25, 26,
27).
3.30
ADCS FOR DSP APPLICATIONS
3-BIT MagAmp™ (FOLDING) ADC BLOCK DIAGRAM
±VR
ANALOG
INPUT
MAGAMP
2
MAGAMP
1
SHA
BIT 1
+
-
BIT 2
BIT 3
GRAY CODE REGISTER
3
GRAY-TO-BINARY CONVERTER
3
OUTPUT REGISTER
3
Figure 3.33
INPUT AND RESIDUE WAVEFORMS
FOR 3-BIT MagAmp ADC
+V R
INPUT
0
t
-VR
+V R
R1
t
0
-VR
+V R
R2
t
0
-VR
GRAY
CODE
000
001
011
010
110
111
101
100
Figure 3.34
3.31
ADCS FOR DSP APPLICATIONS
The MagAmp architecture can be extended to sampling rates previously dominated
by flash converters. The AD9288-100 8-bit, 100MSPS dual ADC is shown in Figure
3.35. The first five bits (Gray code) are derived from five differential MagAmp
stages. The differential residue output of the fifth MagAmp stage drives a 3-bit
flash converter, rather than a single comparator. The Gray-code output of the five
MagAmps and the binary-code output of the 3-bit flash are latched, all converted
into binary, and latched again in the output data register.
AD9288-100 DUAL 8-BIT,
100MSPS ADC FUNCTIONAL DIAGRAM
ANALOG
INPUT
SHA
DIFFERENTIAL
OUTPUTS ON
BITS 1 - 5
MAGAMP
1
MAGAMP
2
BIT
1
GRAY
BIT
2
GRAY
MAGAMP
3
BIT
3
GRAY
REGISTER
MAGAMP
4
MAGAMP
5
BIT
4
GRAY
BIT
5
GRAY
3
BINARY
5
GRAY-TO-BINARY CONVERTER
5
OUTPUT REGISTER
8
Figure 3.35
3.32
3-BIT
FLASH
ADC
3
ADCS FOR DSP APPLICATIONS
REFERENCES
1.
S. A. Jantzi, M. Snelgrove & P. F. Ferguson Jr., A 4th-Order Bandpass
Sigma-Delta Modulator, IEEE Journal of Solid State Circuits,
Vol. 38, No. 3, March 1993, pp.282-291.
2.
System Applications Guide, Analog Devices, Inc., 1993, Section 14.
3.
Mixed Signal Design Seminar, Analog Devices, Inc., 1991, Section 6.
4.
AD77XX-Series Data Sheets, Analog Devices, http://www.analog.com.
5.
Linear Design Seminar, Analog Devices, Inc., 1995, Section 8.
6.
J. Dattorro, A. Charpentier, D. Andreas, The Implementation of a OneStage Multirate 64:1 FIR Decimator for use in One-Bit Sigma-Delta A/D
Applications, AES 7th International Conference, May 1989.
7.
W.L. Lee and C.G. Sodini, A Topology for Higher-Order Interpolative
Coders, ISCAS PROC. 1987.
8.
P.F. Ferguson, Jr., A. Ganesan and R. W. Adams, One Bit Higher Order
Sigma-Delta A/D Converters, ISCAS PROC. 1990, Vol. 2, pp. 890-893.
9.
R. Koch, B. Heise, F. Eckbauer, E. Engelhardt, J. Fisher, and F. Parzefall,
A 12-bit Sigma-Delta Analog-to-Digital Converter with a 15MHz Clock
Rate, IEEE Journal of Solid-State Circuits, Vol. SC-21, No. 6,
December 1986.
10.
Wai Laing Lee, A Novel Higher Order Interpolative Modulator Topology
for High Resolution Oversampling A/D Converters, MIT Masters
Thesis, June 1987.
11.
D. R. Welland, B. P. Del Signore and E. J. Swanson, A Stereo 16-Bit
Delta-Sigma A/D Converter for Digital Audio, J. Audio Engineering
Society, Vol. 37, No. 6, June 1989, pp. 476-485.
12.
R. W. Adams, Design and Implementation of an Audio 18-Bit Analogto-Digital Converter Using Oversampling Techniques, J. Audio
Engineering Society, Vol. 34, March 1986, pp. 153-166.
13.
B. Boser and Bruce Wooley, The Design of Sigma-Delta Modulation
Analog-to-Digital Converters, IEEE Journal of Solid-State Circuits,
Vol. 23, No. 6, December 1988, pp. 1298-1308.
14.
Y. Matsuya, et. al., A 16-Bit Oversampling A/D Conversion Technology
Using Triple-Integration Noise Shaping, IEEE Journal of Solid-State
Circuits, Vol. SC-22, No. 6, December 1987, pp. 921-929.
3.33
ADCS FOR DSP APPLICATIONS
15.
Y. Matsuya, et. al., A 17-Bit Oversampling D/A Conversion Technology
Using Multistage Noise Shaping, IEEE Journal of Solid-State Circuits,
Vol. 24, No. 4, August 1989, pp. 969-975.
16.
P. Ferguson, Jr., A. Ganesan, R. Adams, et. al., An 18-Bit 20-kHz Dual
Sigma-Delta A/D Converter, ISSCC Digest of Technical Papers,
February 1991.
17.
Steven Harris, The Effects of Sampling Clock Jitter on Nyquist Sampling
Analog-to-Digital Converters and on Oversampling Delta Sigma ADCs,
Audio Engineering Society Reprint 2844 (F-4), October, 1989.
18.
Max W. Hauser, Principles of Oversampling A/D Conversion, Journal
Audio Engineering Society, Vol. 39, No. 1/2, January/February 1991,
pp. 3-26.
19.
Daniel H. Sheingold, Analog-Digital Conversion Handbook,
Third Edition, Prentice-Hall, 1986.
20.
Chuck Lane, A 10-bit 60MSPS Flash ADC, Proceedings of the 1989
Bipolar Circuits and Technology Meeting, IEEE Catalog No.
89CH2771-4, September 1989, pp. 44-47.
21.
F.D. Waldhauer, Analog to Digital Converter, U.S. Patent
3-187-325, 1965.
22.
J.O. Edson and H.H. Henning, Broadband Codecs for an Experimental
224Mb/s PCM Terminal, Bell System Technical Journal, 44,
November 1965, pp. 1887-1940.
23.
J.S. Mayo, Experimental 224Mb/s PCM Terminals, Bell System
Technical Journal, 44, November 1965, pp. 1813-1941.
24.
Hermann Schmid, Electronic Analog/Digital Conversions,
Van Nostrand Reinhold Company, New York, 1970.
25.
Carl Moreland, An 8-bit 150MSPS Serial ADC, 1995 ISSCC Digest
of Technical Papers, Vol. 38, p. 272.
26.
Roy Gosser and Frank Murden, A 12-bit 50MSPS Two-Stage A/D
Converter, 1995 ISSCC Digest of Technical Papers, p. 278.
27.
Carl Moreland, An Analog-to-Digital Converter Using SerialRipple Architecture, Masters' Thesis, Florida State University
College of Engineering, Department of Electrical Engineering, 1995.
28.
Practical Analog Design Techniques, Analog Devices, 1995, Chapter
4, 5, and 8.
29.
Linear Design Seminar, Analog Devices, 1995, Chapter 4, 5.
3.34
ADCS FOR DSP APPLICATIONS
30.
System Applications Guide, Analog Devices, 1993, Chapter 12, 13,
15,16.
31.
Amplifier Applications Guide, Analog Devices, 1992, Chapter 7.
32.
Walt Kester, Drive Circuitry is Critical to High-Speed Sampling ADCs,
Electronic Design Special Analog Issue, Nov. 7, 1994, pp. 43-50.
33.
Walt Kester, Basic Characteristics Distinguish Sampling A/D Converters,
EDN, Sept. 3, 1992, pp. 135-144.
34.
Walt Kester, Peripheral Circuits Can Make or Break Sampling ADC
Systems, EDN, Oct. 1, 1992, pp. 97-105.
35.
Walt Kester, Layout, Grounding, and Filtering Complete Sampling
ADC System, EDN, Oct. 15, 1992, pp. 127-134.
36.
High Speed Design Techniques, Analog Devices, 1996, Chapter 4, 5.
3.35
ADCS FOR DSP APPLICATIONS
3.36
DACS FOR DSP APPLICATIONS
SECTION 4
DACs FOR DSP APPLICATIONS
■
DAC Structures
■
Low Distortion DAC Architectures
■
DAC Logic
■
Sigma-Delta DACs
■
Direct Digital Synthesis (DDS)
4.a
DACS FOR DSP APPLICATIONS
4.b
DACS FOR DSP APPLICATIONS
SECTION 4
DACs FOR DSP APPLICATIONS
Walt Kester, James Bryant
DAC STRUCTURES
The most commonly used DAC structures (other than a simple 1-bit DAC based on a
single switch used with a reference voltage) are binary weighted DACs or ladder
networks, but these, though relatively simple in structure, require quite complex
analysis. We will start by examining one of the simplest structures of all, the Kelvin
divider shown in Figure 4.1. An N-bit version of this DAC simply consists of 2N
equal resistors in series. The output is taken from the appropriate tap by closing
one of the 2N switches by decoding 1 of 2N switches from the N-bit data. Recent
DACs using this architecture are referred to as “string DACs.”
SIMPLEST VOLTAGE OUTPUT DAC: THE KELVIN DIVIDER
(“STRING DAC”)
VREF
8 FROM 3
DECODER
3-BIT
DIGITAL
INPUT
ANALOG
OUTPUT
Figure 4.1
This architecture is simple, has a voltage output (but a code-varying ZOUT), and is
inherently monotonic (even if a resistor is zero, OUTPUTN cannot exceed
OUTPUTN+1). It is linear if all the resistors are equal, but may be made
deliberately non-linear if a non-linear DAC is required. Since only two switches
operate during a transition, it is a low-glitch architecture. Its major drawback is the
4.1
DACS FOR DSP APPLICATIONS
large number of resistors required for high resolution, and as a result it is not
commonly used - but, as we shall see later, it is used as a component in more
complex DAC structures.
There is an analogous current output DAC which consists, again, of 2N resistors (or
current sources), but in this case they are all connected in parallel between the
reference voltage input and the virtual ground output (see Figure 4.2).
THE SIMPLEST CURRENT OUTPUT DAC
VREF
7 FROM 3
DECODER
CURRENT
OUTPUT INTO
VIRTUAL
GROUND
(USUALLY AN
OP-AMP I-V
CONVERTER)
3-BIT
DIGITAL
INPUT
Figure 4.2
In this DAC, once a resistor is switched into circuit by increasing digital code, any
further increases do not switch it out again. The structure is thus inherently
monotonic, irrespective of inaccuracies in the resistors, and like the previous case,
may be made intentionally non-linear where a specific non-linearity is required.
Again, as in the previous case, the architecture is rarely, if ever, used to fabricate a
complete DAC because of the large numbers of resistors and switches required.
However, it is often used as a component in a more complex DAC structure.
Unlike the Kelvin divider, this type of DAC does not have a unique name, although
both types are referred to as fully decoded DACs or thermometer DACs or string
DACs.
Fully-decoded DACs are often used as components of more complex DACs. The most
common are "segmented DACs" where part of the output of a fully decoded DAC is
further subdivided. The structure is used because the fully decoded DAC is
inherently monotonic, so if the subdivision is also monotonic, the whole resulting
DAC is also monotonic.
4.2
DACS FOR DSP APPLICATIONS
A voltage segmented DAC (see Figure 4.3) works by further sub-dividing the voltage
across one resistor of a Kelvin divider. The sub-division may be done with a further
Kelvin divider (in which case the whole structure is known as a "Kelvin-Varley
divider", or with some other DAC structure.
SEGMENTED VOLTAGE DACs
KELVIN-VARLEY DIVIDER
(“STRING DAC”)
VREF
VREF
KELVIN DIVIDER AND
R-2R LADDER NETWORK
OUTPUT
OUTPUT
NOTE:
MSB OF LADDER
ON RIGHT
IF THE LADDER NETWORK
IS MONOTONIC, THE
WHOLE DAC IS
MONOTONIC
Figure 4.3
In all DACs, the output is the product of the reference voltage and the digital code,
so in that sense, all DACs are multiplying DACs, but many DACs operate well only
over a limited range of Vref. True MDACs, however, are designed to operate over a
wide range of Vref. A strict definition of a multiplying DAC demands that its
reference voltage range includes 0V, and many, especially current mode ladder
networks with CMOS switches, permit positive, negative, and AC Vref. DACs which
do not work down to 0V Vref are still useful, however, and types where Vref can
vary by 10:1 or so are often called MDACs, although a more accurate description
might be "semi-multiplying" DACs.
LOW DISTORTION DAC ARCHITECTURES
Because of the emphasis in communications systems on DDS DACs with high
SFDR, much effort has been placed on determining optimum DAC architectures.
Practically all low distortion high speed DACs make use of some form of nonsaturating current-mode switching. A straight binary DAC with one current switch
per bit produces code-dependent glitches as discussed above and is certainly not the
most optimum architecture (Figure 4.4). A DAC with one current source per code
level can be shown not to have code-dependent glitches, but it is not practical to
implement for high resolutions. However, this performance can be approached by
4.3
DACS FOR DSP APPLICATIONS
decoding the first few MSBs into a "thermometer" code and have one current switch
per level. For example, a 5-bit thermometer DAC would have an architecture
similar to that shown in Figure 4.5.
5-BIT BINARY DAC ARCHITECTURES
I
I
I
I
I
I
MSB
I/2
I/4
I/8
I/16
MSB
R
OUTPUT
R
2R
R
2R
R
2R
R
OUTPUT
R (CAN BE EXTERNAL)
Figure 4.4
5-BIT “THERMOMETER” OR “FULLY-DECODED” DAC
MINIMIZES CODE-DEPENDENT GLITCHES
MSB
5-BIT
LATCH
5-TO-31
DECODE
LOGIC
31
LINES
31-BIT
LATCH
LSB
CLOCK
Figure 4.5
4.4
31
CURRENT
EQUAL
31
OUTPUT
CURRENT
LINES
SWITCHES
DACS FOR DSP APPLICATIONS
The input binary word is latched and then decoded into 31 outputs which drive a
second latch. The output of the second latch drives 31 equally weighted current
switches whose outputs are summed together. This scheme effectively removes
nearly all the code-dependence of the output glitch. The residual glitch that does
occur at the output is equal regardless of the output code change (it is codeindependent) and can be filtered because it occurs at the DAC update frequency and
its harmonics. The distortion mechanisms associated with the full-decoded
architecture are primarily asymmetrical output slewing, finite switch turn-on and
turn-off times, and integral nonlinearity.
The obvious disadvantage of this type of thermometer DAC is the large number of
latches and switches required to make a 14, 12, 10, or even 8-bit DAC. However, if
this technique is used on the 5 MSBs of an 8, 10, 12, or 14-bit DAC, a significant
reduction in the code-dependent glitch is possible. This process is called
segmentation and is quite common in low distortion DACs.
Figure 4.6 shows a scheme whereby the first 5 bits of a 10-bit DAC are decoded as
described above and drive 31 equally weighted switches. The last 5 bits are derived
from binarily weighted current sources. Equally weighted current sources driving
an R/2R resistor ladder could be used to derive the LSBs, however, this approach
requires thin film resistors which are not generally available on a low-cost CMOS
process. Also, the use of R/2R networks lowers the DAC output impedance, thereby
requiring more drive current to develop the same voltage across a fixed load
resistance.
10-BIT SEGMENTED DAC
5
31
31
MSB
DECODE
10
10-BIT
LATCH
36-BIT
LATCH
5
FULLY
DECODED
MSB
DAC
CURRENT
OUTPUT
5
BINARY
LSB
DAC
CLOCK
Figure 4.6
4.5
DACS FOR DSP APPLICATIONS
The AD9772 14-bit, 150MSPS TxDAC™ uses three sections of segmentation as
shown in Figure 4.7. Other members of the AD977x-family and the AD985x-family
also use this same core.
The first 5 bits (MSBs) are fully decoded and drive 31 equally weighted current
switches, each supplying 512 LSBs of current. The next 4 bits are decoded into 15
lines which drive 15 current switches, each supplying 32 LSBs of current. The 5
LSBs are latched and drive a traditional binary weighted DAC which supplies 1
LSB per output level. A total of 51 current switches and latches are required to
implement this architecture.
AD9772 TxDAC™ 14-BIT CMOS DAC CORE
5
31
31
BITS 1-5
DECODE
14
14-BIT
LATCH
5-TO-31
4
15
51-BIT
LATCH
I = 512 LSB
15
BITS 6-9
DECODE
15
CURRENT
SWITCHES
I = 32 LSB
4-TO-15
5
31
CURRENT
SWITCHES
5
5
5 BINARY
CURRENT
SWITCHES
I = 1 LSB
CURRENT
OUTPUT
FS =
2mA20mA
CLOCK
Figure 4.7
The basic current switching cell is made up of a differential PMOS transistor pair
as shown in Figure 4.8. The differential pairs are driven with low-level logic to
minimize switching transients and time skew. The DAC outputs are symmetrical
differential currents which help to minimize even-order distortion products
(especially which driving a differential output such as a transformer or an op amp
differential I/V converter).
The overall architecture of the AD977x TxDAC™ family and the AD985x-DDS
family is an excellent tradeoff between power/performance and allows the entire
DAC function to be implemented on a standard CMOS process with no thin film
resistors. Single-supply operation on +3.3V or +5V make the devices extremely
attractive for portable and low power applications.
4.6
DACS FOR DSP APPLICATIONS
PMOS TRANSISTOR CURRENT SWITCHES
+VS
RL
RL
Figure 4.8
DAC LOGIC
The earliest monolithic DACs contained little, if any, logic circuitry, and parallel
data had to be maintained on the digital input to maintain the digital output. Today
almost all DACs have input latches, and data need only be written once, not
maintained.
There are innumerable variations of DAC input structure which will not be
discussed here, but the majority today are "double-buffered". A double-buffered DAC
has two sets of latches. Data is initially latched in the first rank and subsequently
transferred to the second as shown in Figure 4.9. There are three reasons why this
arrangement is useful.
The first is that it allows data to enter the DAC in many different ways. A DAC
without a latch, or with a single latch, must be loaded with all bits at once, in
parallel, since otherwise its output during loading may be totally different from
what it was or what it is to become. A double-buffered DAC, on the other hand, may
be loaded with parallel data, serial data, or with 4-bit or 8-bit words, or whatever,
and the output will be unaffected until the new data is completely loaded and the
DAC receives its update instruction.
4.7
DACS FOR DSP APPLICATIONS
DOUBLE-BUFFERED DAC PERMITS COMPLEX
INPUT STRUCTURES AND SIMULTANEOUS UPDATE
DIGITAL
INPUT
INPUT STRUCTURE:
MAY BE SERIAL,
PARALLEL, BYTE-WIDE,
ETC.
OUTPUT LATCH
TRANSFERS DATA
TO DAC TIMING IS
INDEPENDENT OF
INPUT
OUTPUT
DAC
fc = SAMPLING FREQUENCY
OUTPUT STROBE MAY GO TO MANY DACs
Figure 4.9
The second feature of this type of input structure is that the output clock can
operate at a fixed frequency (the DAC update rate), while the input latch can be
loaded asynchronously. This is useful in real-time signal reconstruction
applications.
The third convenience of the double-buffered structure is that many DACs may be
updated simultaneously: data is loaded into the first rank of each DAC in turn, and
when all is ready, the output buffers of all DACs are updated at once. There are
many DAC applications where the output of several DACs must change
simultaneously, and the double-buffered structure allows this to be done very easily.
Most early monolithic high resolution DACs had parallel or byte-wide data ports
and tended to be connected to parallel data buses and address decoders and
addressed by microprocessors as if they were very small write-only memories (some
DACs are not write-only, but can have their contents read as well - this is
convenient for some applications but is not very common). A DAC connected to a
data bus is vulnerable to capacitive coupling of logic noise from the bus to the
analog output, and many DACs today have serial data structures. These are less
vulnerable to such noise (since fewer noisy pins are involved), use fewer pins and
therefore take less space, and are frequently more convenient for use with modern
micro-processors, many of which have serial data ports. Some, but not all, of such
serial DACs have data outputs as well as data inputs so that several DACs may be
connected in series and data clocked to them all from a single serial port. The
arrangement is referred to as "daisy-chaining".
Another development in DACs is the ability to make several on a single chip, which
is useful to reduce PCB sizes and assembly costs. Today (2000) it is possible to buy
sixteen 8-bit, eight 12-bit, four 14-bit, or two 16-/18/20/22/24-bit DACs in a single
package. In the future, even higher densities are probable.
4.8
DACS FOR DSP APPLICATIONS
INTERPOLATING DACS
In ADC-based systems, oversampling can ease the requirements on the antialiasing
filter, and a sigma-delta ADC has this inherent advantage. In a DAC-based system
(such as DDS), the concept of interpolation can be used in a similar manner. This
concept is common in digital audio CD players, where the basic update rate of the
data from the CD is about 44kSPS. "Zeros" are inserted into the parallel data,
thereby increasing the effective update rate to 4-times, 8-times, or 16-times the
fundamental throughput rate. The 4x, 8x, or 16x data stream is passed through a
digital interpolation filter which generates the extra data points. The high
oversampling rate moves the image frequencies higher, thereby allowing a less
complex filter with a wider transition band. The sigma-delta 1-bit DAC architecture
represents the ultimate extension of this concept and has become popular in modern
CD players.
The same concept can be applied to a high speed DAC. Assume a traditional DAC is
driven at an input word rate of 30MSPS (see Figure 4.10A). Assume the DAC
output frequency is 10MHz. The image frequency component at 30–10 = 20MHz
must be attenuated by the analog antialiasing filter, and the transition band of the
filter is 10 to 20MHz. Assume that the image frequency must be attenuated by
60dB. The filter must therefore go from a passband of 10MHz to 60dB stopband
attenuation over the transition band lying between 10 and 20MHz (one octave). A
Butterworth filter design gives 6dB attenuation per octave for each pole. Therefore,
a minimum of 10 poles is required to provide the desired attenuation. Filters
become even more complex as the transition band becomes narrower.
Assume that we increase the DAC update rate to 60MSPS and insert a "zero"
between each original data sample. The parallel data stream is now 60MSPS, but
we must now determine the value of the zero-value data points. This is done by
passing the 60MSPS data stream with the added zeros through a digital
interpolation filter which computes the additional data points. The response of the
digital filter relative to the 2-times oversampling frequency is shown in Figure
4.10B. The analog antialiasing filter transition zone is now 10 to 50MHz (the first
image occurs at 2fc–fo=60–10=50MHz). This transition zone is a little greater than
2 octaves, implying that a 5- or 6-pole Butterworth filter is sufficient.
The AD9772 is a 2-times oversampling interpolating 14-bit DAC, and a simplified
block diagram is shown in Figure 4.11. The device is designed to handle 14-bit input
word rates up to 150MSPS. The output word rate is 300MSPS maximum. For an
output frequency of 60MHz, and an update rate of 150MHz, and an oversampling
ratio of 2, the image frequency occurs at 300MHz – 60MHz = 240MHz. The
transition band for the analog filter is therefore 60MHz to 240MHz. Without
oversampling, the image frequency occurs at 150MHz – 60MHz = 90MHz, and the
filter transition band is 60MHz to 90MHz.
4.9
DACS FOR DSP APPLICATIONS
ANALOG FILTER REQUIREMENTS FOR fo = 10MHz:
fc = 30MSPS, AND fc = 60MSPS
A
ANALOG LPF
fCLOCK = 30MSPS
dB
fo
IMAGE
IMAGE
10
20
30
IMAGE
40
50
IMAGE
60
70
80
FREQUENCY (MHz)
B
fCLOCK = 60MSPS
dB
fo
ANALOG
LPF
IMAGE
IMAGE
10
20
30
40
50
60
70
80
Figure 4.10
AD9772 14-BIT, 150MSPS INTERPOLATING TxDAC™
14
fc
LATCH
14
PLL
DIGITAL
14
INTERPOLATION
FILTER
LATCH
K•fc
TYPICAL APPLICATION: fc = 150MSPS
fo = 60MHz
K=2
Figure 4.11
4.10
14
DAC
LPF
fo
DACS FOR DSP APPLICATIONS
SIGMA-DELTA DACS
Another way of obtaining high resolution is to use oversampling techniques and a
1-bit DAC. The technique, known as sigma-delta (Σ-∆), is computation intensive, so
has only recently become practical for the manufacture of high resolution DACs, but
since it uses a 1-bit DAC, it is intrinsically linear and monotonic.
A Σ-∆ DAC, unlike the Σ-∆ ADC, is mostly digital (see Figure 4.12A). It consists of
an "interpolation filter" (a digital circuit which accepts data at a low rate, inserts
zeros at a high rate, and then applies a digital filter algorithm and outputs data at a
high rate), a Σ-∆ modulator (which effectively acts as a low pass filter to the signal
but as a high pass filter to the quantization noise, and converts the resulting data to
a high speed bit stream), and a 1-bit DAC whose output switches between equal
positive and negative reference voltages. The output is filtered in an external analog
LPF. Because of the high oversampling frequency, the complexity of the LPF is
much less than the case of traditional Nyquist operation.
SIGMA-DELTA (Σ∆
Σ∆)
Σ∆ DACs:
1-BIT AND MULTIBIT
SINGLE BIT
N-BITS @ fs
N-BITS @ K fs
DIGITAL
INTERPOLATION
FILTER
1- BIT @ KfS
DIGITAL
Σ∆
MODULATOR
ANALOG SIGNAL:
2 LEVELS
1-BIT
DAC
ANALOG
OUTPUT
FILTER
ANALOG
OUTPUT
MULTIBIT (AD185x-SERIES)
N-BITS @ fs
N-BITS @ K fs
DIGITAL
INTERPOLATION
FILTER
n- BITS @ KfS
DIGITAL
MULTIBIT
Σ∆
ANALOG SIGNAL:
2n LEVELS
n-BIT
DAC
MODULATOR
ANALOG
OUTPUT
FILTER
ANALOG
OUTPUT
N = 16 / 18 / 20 / 24 BITS, fs = 192kSPS
Figure 4.12
It is possible to use more than one bit in the DAC, and this leads to the multibit
architecture shown in Figure 4.12B. The concept is similar to that of the
interpolating DAC previously discussed, with the addition of the digital sigma-delta
modulator. In the past, multibit DACs have been difficult to design because of the
accuracy requirement on the n-bit internal DAC (this DAC, although only n-bits,
must have the linearity of the final number of bits, N). The AD185x-series of audio
DACs, however use a proprietary data scrambling technique (called direct data
4.11
DACS FOR DSP APPLICATIONS
scrambling, or D2S) which overcomes this problem and produces excellent
performance with respect to THD + N. For instance, the AD1853 dual 24-bit,
192kSPS DAC has greater than 115dB THD + N at a 48kSPS sampling rate.
DIRECT DIGITAL SYNTHESIS (DDS)
A frequency synthesizer generates multiple frequencies from one or more frequency
references. These devices have been used for decades, especially in communications
systems. Many are based upon switching and mixing frequency outputs from a bank
of crystal oscillators. Others have been based upon well understood techniques
utilizing phase-locked loops (PLLs). This mature technology is illustrated in Figure
4.13. A fixed-frequency reference drives one input of the phase comparator. The
other phase comparator input is driven from a divide-by-N counter which is in turn
driven by a voltage-controlled-oscillator (VCO). Negative feedback forces the output
of the internal loop filter to a value which makes the VCO output frequency N-times
the reference frequency. The time constant of the loop is controlled by the loop filter.
There are many tradeoffs in designing a PLL, such a phase noise, tuning speed,
frequency resolution, etc., and there are many good references on the subject (see
References 1, 2, and 3).
FREQUENCY SYNTHESIS USING
OSCILLATORS AND PHASE-LOCKED LOOPS
OSCILLATOR BANK
PHASE-LOCKED LOOP
MIXER
XO
1
PHASE
COMPARATOR
fc
FIXED
FREQUENCY
REFERENCE
XO
2
VCO
LOOP
FILTER
f out
f out
÷ N
MIXER
XO
3
f out = N • f c
SW
XO
n
Figure 4.13
With the widespread use of digital techniques in instrumentation and
communications systems, a digitally-controlled method of generating multiple
frequencies from a reference frequency source has evolved called Direct Digital
Synthesis (DDS). The basic architecture is shown in Figure 4.14. In this simplified
4.12
DACS FOR DSP APPLICATIONS
model, a stable clock drives a programmable-read-only-memory (PROM) which
stores one or more integral number of cycles of a sinewave (or other arbitrary
waveform, for that matter). As the address counter steps through each memory
location, the corresponding digital amplitude of the signal at each location drives a
DAC which in turn generates the analog output signal. The spectral purity of the
final analog output signal is determined primarily by the DAC. The phase noise is
basically that of the reference clock.
FUNDAMENTAL DIRECT DIGITAL SYNTHESIS SYSTEM
CLOCK
fc
ADDRESS
COUNTER
PROM
SIN
LOOKUP
TABLE
N-BITS
REGISTER
N-BITS
LOOKUP TABLE CONTAINS SIN
DATA FOR INTEGRAL NUMBER
OF CYCLES
DAC
fout
LPF
Figure 4.14
The DDS system differs from the PLL in several ways. Because a DDS system is a
sampled data system, all the issues involved in sampling must be considered:
quantization noise, aliasing, filtering, etc. For instance, the higher order harmonics
of the DAC output frequencies fold back into the Nyquist bandwidth, making them
unfilterable, whereas, the higher order harmonics of the output of PLL-based
synthesizers can be filtered. There are other considerations which will be discussed
shortly.
A fundamental problem with this simple DDS system is that the final output
frequency can be changed only by changing the reference clock frequency or by
reprogramming the PROM, making it rather inflexible. A practical DDS system
implements this basic function in a much more flexible and efficient manner using
digital hardware called a Numerically Controlled Oscillator (NCO). A block diagram
of such a system is shown in Figure 4.15.
4.13
DACS FOR DSP APPLICATIONS
A FLEXIBLE DDS SYSTEM
PHASE ACCUMULATOR
n = 24-32 BITS
SERIAL
OR BYTE
LOAD
REGISTER
n
n
PARALLEL
DELTA
PHASE
REGISTER
M
n
n
PHASE n
REGISTER
SIN ROM
LOOKUP
TABLE
CLOCK
PHASE
TRUNCATION
12-16 BITS
FREQUENCY CONTROL
N-BITS
AMPLITUDE
TRUNCATION
fc
DAC
fo =
M • fc
2n
LPF
Figure 4.15
The heart of the system is the phase accumulator whose contents is updated once
each clock cycle. Each time the phase accumulator is updated, the digital number,
M, stored in the delta phase register is added to the number in the phase
accumulator register. Assume that the number in the delta phase register is 00...01
and that the initial contents of the phase accumulator is 00...00. The phase
accumulator is updated by 00...01 on each clock cycle. If the accumulator is 32-bits
wide, 232 clock cycles (over 4 billion) are required before the phase accumulator
returns to 00...00, and the cycle repeats.
The truncated output of the phase accumulator serves as the address to a sine (or
cosine) lookup table. Each address in the lookup table corresponds to a phase point
on the sinewave from 0° to 360°. The lookup table contains the corresponding digital
amplitude information for one complete cycle of a sinewave. (Actually, only data for
90° is required because the quadrature data is contained in the two MSBs). The
lookup table therefore maps the phase information from the phase accumulator into
a digital amplitude word, which in turn drives the DAC.
Consider the case for n=32, and M=1. The phase accumulator steps through each of
232 possible outputs before it overflows. The corresponding output sinewave
frequency is equal to the clock frequency divided by 232. If M=2, then the phase
accumulator register "rolls over" twice as fast, and the output frequency is doubled.
This can be generalized as follows.
4.14
DACS FOR DSP APPLICATIONS
For an n-bit phase accumulator (n generally ranges from 24 to 32 in most DDS
systems), there are 2n possible phase points. The digital word in the delta phase
register, M, represents the amount the phase accumulator is incremented each clock
cycle. If fc is the clock frequency, then the frequency of the output sinewave is equal
to:
fo =
M ⋅ fc
.
2n
This equation is known as the DDS "tuning equation." Note that the frequency
resolution of the system is equal to fc/2n. For n=32, the resolution is greater than
one part in four billion! In a practical DDS system, all the bits out of the phase
accumulator are not passed on to the lookup table, but are truncated, leaving only
the first 12 to 16 MSBs. This reduces the size of the lookup table and does not affect
the frequency resolution. The phase truncation only adds a small but acceptable
amount of phase noise to the final output; the bulk of the output distortion comes
from the DAC itself.
The basic DDS system described above is extremely flexible and has high resolution.
The frequency can be changed instantaneously with no phase discontinuity by
simply changing the contents of the M-register. However, practical DDS systems
first require the execution of a serial, or byte-loading sequence to get the new
frequency word into an internal buffer register which precedes the parallel-output
M-register. This is done to minimize package pin count. After the new word is
loaded into the buffer register, the parallel-output delta phase register is clocked,
thereby changing all the bits simultaneously. The number of clock cycles required to
load the delta-phase buffer register determines the maximum rate at which the
output frequency can be changed.
The AD9850 125MSPS DDS system (Figure 4.16) uses a 32-bit phase accumulator
which is truncated to 14-bits (MSBs) before being passed to the lookup table. The
final digital output is 10-bits to the internal DAC. The AD9850 allows the output
phase to be modulated using an additional register and an adder placed between the
output of the phase accumulator register and the input to the lookup table. The
AD9850 uses a 5-bit word to control the phase which allows shifting the phase in
increments of 180°, 90°, 45°, 22.5°, 11.25°, and any combination thereof. The device
also contains an internal high speed comparator which can be configured to accept
the (externally) filtered output of the DAC to generate a low-jitter output pulse
suitable for driving the sampling clock input of an ADC. The full scale output
current can be adjusted from 10 to 20mA using a single external resistor, and the
output voltage compliance is +1V.
The frequency tuning (delta-phase register input word) and phase modulation words
are loaded into the AD9850 via a parallel or serial loading format. The parallel load
format consists of five consecutive loads of an 8-bit control word (byte). The first 8bit byte controls phase modulation (5-bits), power-down enable (1-bit), and loading
format (2-bits). Bytes 2-5 comprise the 32-bit frequency tuning word. The maximum
control register update frequency is 23MHz. Serial loading of the AD9850 is
accomplished via a 40-bit serial data stream on a single pin. Maximum update rate
of the control register in the serial-load mode is 3MHz.
4.15
DACS FOR DSP APPLICATIONS
The AD9850 consumes only 380mW of power on a single +5V supply at a maximum
125MSPS clock rate. The device is available in a 28-pin surface mount SSOP
(Shrink Small Outline Package).
AD9850 CMOS 125MSPS DDS/DAC SYNTHESIZER
5
5
BYTE LOAD
PHASE
CONTROL
8 BITS X 5
DATA AND
CONTROL
INPUT
REGISTER
SERIAL LOAD
32
PHASE
ACCUMULATOR
32
14
SIN
LOOKUP
TABLE
1 BIT X 40
10
10-BIT
DAC
WORD LOAD CLOCK
ANALOG
OUT
RSET
REFERENCE CLOCK IN
HIGH SPEED
COMPARATOR
+
-
Figure 4.16
Analog Devices offers a number of DDS systems for a variety of applications. The
AD983x family are low-cost 10-bit systems with clock frequencies up to 50MSPS.
The AD985x-family offers 10-bit and 12-bit systems with clock frequencies up to
300MSPS and additional functions such as quadrature and phase modulation,
chirp-mode capability, and programmable on- chip reference clock multipliers.
4.16
DACS FOR DSP APPLICATIONS
REFERENCES
1.
R.E. Best, Phase-Locked Loops, McGraw-Hill, New York, 1984.
2.
F.M. Gardner, Phaselock Techniques, 2nd Edition, John Wiley,
New York, 1979.
3.
Phase-Locked Loop Design Fundamentals, Applications Note AN-535,
Motorola, Inc.
4.
The ARRL Handbook for Radio Amateurs, American Radio
Relay League, Newington, CT, 1992.
5.
Richard J. Kerr and Lindsay A. Weaver, Pseudorandom Dither for
Frequency Synthesis Noise, United States Patent Number 4,901,265,
February 13, 1990.
6.
Henry T. Nicholas, III and Henry Samueli, An Analysis of the Output
Spectrum of Direct Digital Frequency Synthesizers in the Presence of
Phase-Accumulator Truncation, IEEE 41st Annual Frequency Control
Symposium Digest of Papers, 1987, pp. 495-502, IEEE Publication No.
CH2427-3/87/0000-495.
7.
Henry T. Nicholas, III and Henry Samueli, The Optimization of Direct
Digital Frequency Synthesizer Performance in the Presence of Finite Word
Length Effects, IEEE 42nd Annual Frequency Control Symposium
Digest of Papers, 1988, pp. 357-363, IEEE Publication No. CH25882/88/0000-357.
4.17
DACS FOR DSP APPLICATIONS
4.18
FAST FOURIER TRANSFORMS
SECTION 5
FAST FOURIER TRANSFORMS
■
The Discrete Fourier Transform
■
The Fast Fourier Transform
■
FFT Hardware Implementation and Benchmarks
■
DSP Requirements for Real Time FFT Applications
■
Spectral Leakage and Windowing
5.a
FAST FOURIER TRANSFORMS
5.b
FAST FOURIER TRANSFORMS
SECTION 5
FAST FOURIER TRANSFORMS
Walt Kester
THE DISCRETE FOURIER TRANSFORM
In 1807 the French mathematician and physicist Jean Baptiste Joseph Fourier
presented a paper to the Institut de France on the use of sinusoids to represent
temperature distributions. The paper made the controversial claim that any
continuous periodic signal could be represented by the sum of properly chosen
sinusoidal waves. Among the publication review committee were two famous
mathematicians: Joseph Louis Lagrange, and Pierre Simon de Laplace. Lagrange
objected strongly to publication on the basis that Fourier’s approach would not work
with signals having discontinuous slopes, such as square waves. Fourier’s work was
rejected, primarily because of Lagrange’s objection, and was not published until the
death of Lagrange, some 15 years later. In the meantime, Fourier’s time was
occupied with political activities, expeditions to Egypt with Napoleon, and trying to
avoid the guillotine after the French Revolution! (This bit of history extracted from
Reference 1, p.141).
It turns out that both Fourier and Lagrange were at least partially correct.
Lagrange was correct that a summation of sinusoids cannot exactly form a signal
with a corner. However, you can get very close if enough sinusoids are used. (This is
described by the Gibbs effect, and is well understood by scientists, engineers, and
mathematicians today).
Fourier analysis forms the basis for much of digital signal processing. Simply
stated, the Fourier transform (there are actually several members of this family)
allows a time domain signal to be converted into its equivalent representation in the
frequency domain. Conversely, if the frequency response of a signal is known, the
inverse Fourier transform allows the corresponding time domain signal to be
determined.
In addition to frequency analysis, these transforms are useful in filter design, since
the frequency response of a filter can be obtained by taking the Fourier transform of
its impulse response. Conversely, if the frequency response is specified, then the
required impulse response can be obtained by taking the inverse Fourier transform
of the frequency response. Digital filters can be constructed based on their impulse
response, because the coefficients of an FIR filter and its impulse response are
identical.
The Fourier transform family (Fourier Transform, Fourier Series, Discrete Time
Fourier Series, and Discrete Fourier Transform) is shown in Figure 5.2. These
accepted definitions have evolved (not necessarily logically) over the years and
depend upon whether the signal is continuous–aperiodic, continuous–periodic,
sampled–aperiodic, or sampled–periodic. In this context, the term sampled is the
same as discrete (i.e., a discrete number of time samples).
5.1
FAST FOURIER TRANSFORMS
APPLICATIONS OF THE DISCRETE
FOURIER TRANSFORM (DFT)
Sampled
Time Domain
Discrete Fourier Transform (DFT)
Sampled
Frequency Domain
Inverse DFT (IDFT)
■ Digital Spectral Analysis
◆ Spectrum Analyzers
◆ Speech Processing
◆ Imaging
◆ Pattern Recognition
■ Filter Design
◆ Calculating Impulse Response from Frequency Response
◆ Calculating Frequency Response from Impulse Response
■ The Fast Fourier Transform (FFT) is Simply an Algorithm
for Efficiently Calculating the DFT
Figure 5.1
FOURIER TRANSFORM FAMILY
AS A FUNCTION OF TIME DOMAIN SIGNAL TYPE
FOURIER TRANSFORM:
Signal is Continuous
and Aperiodic
t
FOURIER SERIES:
Signal is Continuous
and Periodic
t
DISCRETE TIME FOURIER SERIES:
Signal is Sampled
and Aperiodic
t
DISCRETE FOURIER TRANSFORM:
(Discrete Fourier Series)
Signal is Sampled
and Periodic
N=8
t
Sample 0
Figure 5.2
5.2
Sample N – 1
FAST FOURIER TRANSFORMS
The only member of this family which is relevant to digital signal processing is the
Discrete Fourier Transform (DFT) which operates on a sampled time domain signal
which is periodic. The signal must be periodic in order to be decomposed into the
summation of sinusoids. However, only a finite number of samples (N) are available
for inputting into the DFT. This dilemma is overcome by placing an infinite number
of groups of the same N samples “end-to-end,” thereby forcing mathematical (but
not real-world) periodicity as shown in Figure 5.2.
The fundamental analysis equation for obtaining the N-point DFT is as follows:
X( k ) =
1
N
N −1
∑
n =0
x(n )e − j2 πnk / N =
1
N
N −1
x(n )[cos(2πnk / N ) − j sin(2 πnk / N )]
n =0
∑
At this point, some terminology clarifications are in order regarding the above
equation (also see Figure 5.3). X(k) (capital letter X) represents the DFT frequency
output at the kth spectral point, where k ranges from 0 to N–1. The quantity N
represents the number of sample points in the DFT data frame.
Note that “N” should not be confused with ADC or DAC resolution, which is also
given by the quantity N in other places in this book.
The quantity x(n) (lower case letter x) represents the nth time sample, where n also
ranges from 0 to N – 1. In the general equation, x(n) can be real or complex.
Notice that the cosine and sine terms in the equation can be expressed in either
polar or rectangular coordinates using Euler’s equation:
e jθ = cos θ + j sin θ
The DFT output spectrum, X(k), is the correlation between the input time samples
and N cosine and N sine waves. The concept is best illustrated in Figure 5.4. In this
figure, the real part of the first four output frequency points is calculated, therefore,
only the cosine waves are shown. A similar procedure is used with sine waves in
order to calculate the imaginary part of the output spectrum.
The first point, X(0), is simply the sum of the input time samples, because cos(0) =
1. The scaling factor, 1/N, is not shown, but must be present in the final result. Note
that X(0) is the average value of the time samples, or simply the DC offset. The
second point, ReX(1), is obtained by multiplying each time sample by each
corresponding point on a cosine wave which makes one complete cycle in the
interval N and summing the results. The third point, ReX(2), is obtained by
multiplying each time sample by each corresponding point of a cosine wave which
has two complete cycles in the interval N and then summing the results. Similarly,
the fourth point, ReX(3), is obtained by multiplying each time sample by the
corresponding point of a cosine wave which has three complete cycles in the interval
N and summing the results. This process continues until all N outputs have been
computed. A similar procedure is followed using sine waves in order to calculate the
imaginary part of the frequency spectrum. The cosine and sine waves are referred to
as basis functions.
5.3
FAST FOURIER TRANSFORMS
THE DISCRETE FOURIER TRANSFORM (DFT)
■ A Periodic Signal Can be Decomposed into the Sum of Properly
Chosen Cosine and Sine Waves (Jean Baptiste Joseph Fourier,
1807)
■ The DFT Operates on a Finite Number (N) of Digitized Time
Samples, x(n). When These Samples are Repeated and Placed
“End-to-End”, they Appear Periodic to the Transform.
■ The Complex DFT Output Spectrum X(k) is the Result of
Correlating the Input Samples with sine and cosine Basis
Functions:
–j2π
πnk
1 N–1
1 N–1
2π
π nk
2π
π nk
N
= N Σ x(n) cos N – j sin N
X(k) = N Σ x(n) e
n=0
n=0
0≤ k≤ N–1
Figure 5.3
CORRELATION OF TIME SAMPLES WITH
BASIS FUNCTIONS USING THE DFT FOR N = 8
TIME DOMAIN
BASIS FUNCTIONS
(k = 0)
FREQUENCY DOMAIN
ReX(0)
cos 0
n
0
(k = 1)
πn
x(7) cos 2π
8
x(0)
N/2
N–1
N/2
k=0
N/2
N–1
(k = 3)
π3n
cos 2π
8
k
n
N–1
(k = 2)
π2n
cos 2π
8
0
0 k=1
N/2
N–1
ReX(2)
k
n
N/2
N–1
0
k=2
N/2
N–1
ReX(3)
N/2
N–1 n
k
0
0
Figure 5.4
5.4
N–1
ReX(1)
0
n
0
N/2
k
k=3 N/2
N–1
FAST FOURIER TRANSFORMS
Assume that the input signal is a cosine wave having a period of N, i.e., it makes
one complete cycle during the data window. Also assume its amplitude and phase is
identical to the first cosine wave of the basis functions, cos(2πn/8). The output
ReX(1) contains a single point, and all the other ReX(k) outputs are zero. Assume
that the input cosine wave is now shifted to the right by 90º. The correlation
between it and the corresponding basis function is zero. However, there is an
additional correlation required with the basis function sin(2πn/8) to yield ImX(1).
This shows why both real and imaginary parts of the frequency spectrum need to be
calculated in order to determine both the amplitude and phase of the frequency
spectrum.
Notice that the correlation of a sine/cosine wave of any frequency other than that of
the basis function produces a zero value for both ReX(1) and ImX(1).
A similar procedure is followed when using the inverse DFT (IDFT) to reconstruct
the time domain samples, x(n), from the frequency domain samples X(k). The
synthesis equation is given by:
x(n ) =
N −1
∑
k =0
X(k )e j2πnk / N =
N −1
X( k )[cos(2πnk / N ) + j sin(2 πnk / N )]
k =0
∑
There are two basic types of DFTs: real, and complex. The equations shown in
Figure 5.5 are for the complex DFT, where the input and output are both complex
numbers. Since time domain input samples are real and have no imaginary part,
the imaginary part of the input is always set to zero. The output of the DFT, X(k),
contains a real and imaginary component which can be converted into ampltude and
phase.
The real DFT, although somewhat simpler, is basically a simplification of the
complex DFT. Most FFT routines are written using the complex DFT format,
therefore understanding the complex DFT and how it relates to the real DFT is
important. For instance, if you know the real DFT frequency outputs and want to
use a complex inverse DFT to calculate the time samples, you need to know how to
place the real DFT oututs points into the complex DFT format before taking the
complex inverse DFT.
Figure 5.6 shows the input and output of a real and a complex FFT. Notice that the
output of the real DFT yields real and imaginary X(k) values, where k ranges from
only 0 to N/2. Note that the imaginary points ImX(0) and ImX(N/2) are always zero
because sin(0) and sin(nπ) are both always zero.
The frequency domain output X(N/2) corresponds to the frequency output at onehalf the sampling frequency, fs. The width of each frequency bin is equal to fs/N.
5.5
FAST FOURIER TRANSFORMS
THE COMPLEX DFT
Frequency Domain
←←
–j2π
πnk
Σ x(n) e N
n=0
N–1
1
X(k) =
N
–j2π
π
WN = e N
←←
Time Domain
2π
π nk
2π
π nk
1 N–1
x(n) cos N – j sin N
Σ
N
n=0
=
1 N–1
x(n) WNnk
N Σ
n=0
INVERSE DFT
j2π
πnk
N
X(k) e
Σ
k=0
Time Domain
=
N–1
x(n) =
←←
DFT
=
=
N–1
Σ
←←
Frequency Domain
2π
π nk
2π
π nk
+ j sin N
N
X(k) cos
k=0
N–1
Σ
k=0
, 0≤ k≤ N–1
X(k) WN–nk
, 0≤ n≤ N–1
Figure 5.5
DFT INPUT/OUTPUT SPECTRUM
Time Domain, x(n)
REAL DFT
DC Offset
Real
Real
N/2
0
Frequency Domain, X(k)
N–1
N Points
0 ≤ n ≤ N–1
fs
N
zero
zero
Imaginary
0 ≤ k ≤ N/2
0
N/2
COMPLEX DFT
Real
0
2N Points
0
N/2
N–1
0
zero
N–1
N–1
N/2
N/2
0 ≤ k ≤ N–1
Figure 5.6
2N Points
zero
Imaginary
0
0 ≤ n ≤ N–1
5.6
fs
N
Real
Imaginary
N/2
N Points
+ Two Zero
Points
N/2 ⇒ fs/2
0
N–1
FAST FOURIER TRANSFORMS
The complex DFT has real and imaginary values both at its input and output. In
practice, the imaginary parts of the time domain samples are set to zero. If you are
given the output spectrum for a complex DFT, it is useful to know how to relate
them to the real DFT output and vice versa. The crosshatched areas in the diagram
correspond to points which are common to both the real and complex DFT.
Figure 5.7 shows the relationship between the real and complex DFT in more detail.
The real DFT output points are from 0 to N/2, with ImX(0) and ImX(N/2) always
zero. The points between N/2 and N – 1 contain the negative frequencies in the
complex DFT. Note that ReX(N/2 + 1) has the same value as Re(N/2 – 1). Similarly,
ReX(N/2 + 2) has the same value as ReX(N/2 – 2), etc. Also, note that ImX(N/2 + 1)
is the negative of ImX(N/2 – 1), and ImX(N/2 + 2) is the negative of ImX(N/2 – 2),
etc. In other words, ReX(k) has even symmetry about N/2 and ImX(k) has odd
symmetry about N/2. In this way, the negative frequency components for the
complex FFT can be generated if you are only given the real DFT components.
The equations for the complex and the real DFT are summarized in Figure 5.8. Note
that the equations for the complex DFT work nearly the same whether taking the
DFT, X(k) or the IDFT, x(n). The real DFT does not use complex numbers, and the
equations for X(k) and x(n) are significantly different. Also, before using the x(n)
equation, ReX(0) and ReX(N/2) must be divided by two. These details are explained
in Chapter 31 of Reference 1, and the reader should study this chapter before
attemping to use these equations.
CONSTRUCTING THE COMPLEX DFT NEGATIVE
FREQUENCY COMPONENTS FROM THE REAL DFT
Frequency Domain
Time Domain
Real Part
0
N/2
Even
Symmetry
About N/2
(fs/2)
Real Part
N–1
0
N/2
N–1
“Negative” Frequency
Imaginary Part
Imaginary Part
(All Zeros)
0
N/2
N–1
0
N/2
Odd
Symmetry
About N/2
N–1
(fs/2)
Axis of
Symmetry
Figure 5.7
5.7
FAST FOURIER TRANSFORMS
The DFT output spectrum can be represented in either polar form (magnitude and
phase) or rectangular form (real and imaginary) as shown in Figure 5.9. The
conversion between the two forms is straightforward.
COMPLEX AND REAL DFT EQUATIONS
REAL TRANSFORM
COMPLEX TRANSFORM
–j2π
πnk
N–1
1
N
X(k) =
N Σ x(n) e
n=0
ReX(k) =
ImX(k) =
x(n) =
N–1
Σ X(k) e
k=0
j2π
πnk
N
2
N
N–1
Σ
n=0
–2
N
N–1
Σ
n=0
N/2
x(n) =
x(n) cos(2π
πnk/N)
Σ
k=0
x(n) sin(2π
πnk/N)
ReX(k) cos(2π
πnk/N)
– ImX(k) cos(2π
πnk/N)
Time Domain: x(n) is complex, discrete,
and periodic. n runs from 0 to N– 1
Time Domain: x(n) is real, discrete, and periodic.
n runs from 0 to N – 1
Frequency Domain: X(k) is complex,
discrete, and periodic. k runs from 0 to N–1
k = 0 to N/2 are positive frequencies.
k = N/2 to N–1 are negative frequencies
Frequency domain:
ReX(k) is real, discrete, and periodic.
ImX(k) is real, discrete, and periodic.
k runs from 0 to N/2
Before using x(n) equation, ReX(0) and
ReX(N/2) must be divided by two.
Figure 5.8
CONVERTING REAL AND IMAGINARY DFT
OUTPUTS INTO MAGNITUDE AND PHASE
X(k)
=
ReX(k) + j ImX(k)
ReX(k) 2 + ImX(k) 2
MAG [X(k)] =
ϕ [X(k)]
tan–1
=
ImX(k)
ReX(k)
Im X(k)
X(k)
MAG[X(k)]
ϕ
Re X(k)
Figure 5.9
5.8
FAST FOURIER TRANSFORMS
THE FAST FOURIER TRANSFORM
In order to understand the development of the FFT, consider first the 8-point DFT
expansion shown in Figure 5.10. In order to simplify the diagram, note that the
quantity WN is defined as:
WN = e − j2π / N .
This leads to the definition of the twiddle factors as:
WN nk = e − j2πnk / N .
The twiddle factors are simply the sine and cosine basis functions written in polar
form. Note that the 8-point DFT shown in the diagram requires 64 complex
multiplications. In general, an N-point DFT requires N2 complex multiplications.
The number of multiplications required is significant because the multiplication
function requires a relatively large amount of DSP processing time. In fact, the total
time required to compute the DFT is directly proportional to the number of
multiplications plus the required amount of overhead.
THE 8-POINT DFT (N = 8)
–j2π
πnk
N–1
N–1
1
1
X(k) =
N
= N Σ x(n) WNnk
x(n)
e
Σ
Nn=0
n=0
–j2π
π
WN = e N
X(0) =
x(0)W80
+ x(1)W80 + x(2)W80 + x(3)W80 + x(4)W80 + x(5)W80 + x(6)W80 + x(7)W80
X(1) =
x(0)W80
+ x(1)W81 + x(2)W82 + x(3)W83 + x(4)W84 + x(5)W85 + x(6)W86 + x(7)W87
X(2) =
x(0)W80
+ x(1)W82 + x(2)W84 + x(3)W86 + x(4)W88 + x(5)W810 + x(6)W812 + x(7)W814
X(3) =
x(0)W80
+ x(1)W83 + x(2)W86 + x(3)W89 + x(4)W812 + x(5)W815 + x(6)W818 + x(7)W821
X(4) =
x(0)W80
+ x(1)W84 + x(2)W88 + x(3)W812 + x(4)W816 + x(5)W820 + x(6)W824 + x(7)W828
X(5) =
x(0)W80
+ x(1)W85 + x(2)W810 + x(3)W815 + x(4)W820 + x(5)W825 + x(6)W830 + x(7)W835
X(6) =
x(0)W80
+ x(1)W86 + x(2)W812 + x(3)W818 + x(4)W824 + x(5)W830 + x(6)W836 + x(7)W842
X(7) =
x(0)W80
+ x(1)W87 + x(2)W814 + x(3)W821 + x(4)W828 + x(5)W835 + x(6)W842 + x(7)W849
N2 Complex Multiplications
1
Scaling Factor Omitted
N
Figure 5.10
5.9
FAST FOURIER TRANSFORMS
The FFT is simply an algorithm to speed up the DFT calculation by reducing the
number of multiplications and additions required. It was popularized by J. W.
Cooley and J. W. Tukey in the 1960s and was actually a rediscovery of an idea of
Runge (1903) and Danielson and Lanczos (1942), first occurring prior to the
availability of computers and calculators – when numerical calculation could take
many man hours. In addition, the German mathematician Karl Friedrich Gauss
(1777 – 1855) had used the method more than a century earlier.
In order to understand the basic concepts of the FFT and its derivation, note that
the DFT expansion shown in Figure 5.10 can be greatly simplified by taking
advantage of the symmetry and periodicity of the twiddle factors as shown in Figure
5.11. If the equations are rearranged and factored, the result is the Fast Fourier
Transform (FFT) which requires only (N/2) log2(N) complex multiplications. The
computational efficiency of the FFT versus the DFT becomes highly significant
when the FFT point size increases to several thousand as shown in Figure 5.12.
However, notice that the FFT computes all the output frequency components (either
all or none!). If only a few spectral points need to be calculated, the DFT may
actually be more efficient. Calculation of a single spectral output using the DFT
requires only N complex multiplications.
APPLYING THE PROPERTIES OF
SYMMETRY AND PERIODICITY TO WNr FOR N = 8
Symmetry:
N=8
WN r+ N/2 = – WNr
, Periodicity:
W84 =
W80+4 =
– W80 = – 1
W85 =
W81+4 =
– W81
W86 =
W82+4 =
– W82
W87 =
W83+4 =
– W83
W88 =
W80+8 =
+ W80 = + 1
W89 =
W81+8 =
+ W81
W810 =
W82+8 =
+ W82
W811 =
W83+8 =
+ W83
Figure 5.11
5.10
WN r+ N
=
WNr
FAST FOURIER TRANSFORMS
THE FAST FOURIER TRANSFORM (FFT) VS.
THE DISCRETE FOURIER TRANSFORM (DFT)
■ The FFT is Simply an Algorithm for Efficiently Calculating the DFT
■ Computational Efficiency of an N-Point FFT:
◆ DFT:
N2
Complex Multiplications
◆ FFT:
(N/2) log2(N) Complex Multiplications
N
DFT Multiplications FFT Multiplications FFT Efficiency
256
65,536
1,024
64 : 1
512
262,144
2,304
114 : 1
1,024
1,048,576
5,120
205 : 1
2,048
4,194,304
11,264
372 : 1
4,096
16,777,216
24,576
683 : 1
Figure 5.12
The radix-2 FFT algorithm breaks the entire DFT calculation down into a number of
2-point DFTs. Each 2-point DFT consists of a multiply-and-accumulate operation
called a butterfly, as shown in Figure 5.13. Two representations of the butterfly are
shown in the diagram: the top diagram is the actual functional representation of the
butterfly showing the digital multipliers and adders. In the simplified bottom
diagram, the multiplications are indicated by placing the multiplier over an arrow,
and addition is indicated whenever two arrows converge at a dot.
The 8-point decimation-in-time (DIT) FFT algorithm computes the final output in
three stages as shown in Figure 5.14. The eight input time samples are first divided
(or decimated) into four groups of 2-point DFTs. The four 2-point DFTs are then
combined into two 4-point DFTs. The two 4-point DFTs are then combined to
produce the final output X(k). The detailed process is shown in Figure 5.15, where
all the multiplications and additions are shown. Note that the basic two-point DFT
butterfly operation forms the basis for all computation. The computation is done in
three stages. After the first stage computation is complete, there is no need to store
any previous results. The first stage outputs can be stored in the same registers
which originally held the time samples x(n). Similarly, when the second stage
computation is completed, the results of the first stage computation can be deleted.
In this way, in-place computation proceeds to the final stage. Note that in order for
the algorithm to work properly, the order of the input time samples, x(n), must be
properly re-ordered using a bit reversal algorithm.
5.11
FAST FOURIER TRANSFORMS
THE BASIC BUTTERFLY COMPUTATION IN
THE DECIMATION-IN-TIME FFT ALGORITHM
+
a
Σ
A = a + bWNr
Σ
B = a – bWNr
+
WN r
+
b
–
SIMPLIFIED REPRESENTATION
a
b
A = a + bWNr
WNr
–1
B = a – bWNr
Figure 5.13
COMPUTATION OF AN 8-POINT DFT IN THREE STAGES
USING DECIMATION-IN-TIME
x(0)
x(4)
x(2)
x(6)
x(1)
x(5)
x(3)
x(7)
2-POINT
DFT
2-POINT
DFT
2-POINT
DFT
2-POINT
DFT
X(0)
X(2)
COMBINE
4-POINT
DFTs
COMBINE
2-POINT
DFTs
X(3)
X(4)
X(5)
X(6)
X(7)
Figure 5.14
5.12
X(1)
COMBINE
2-POINT
DFTs
FAST FOURIER TRANSFORMS
EIGHT-POINT DECIMATION-IN-TIME FFT ALGORITHM
STAGE 1
STAGE 2
STAGE 3
X(0)
x(0)
x(4)
W80
W80
x(2)
x(6)
X(1)
–1
W80
W82
X(3)
–1
–1
W80
x(1)
x(5)
–1
W80
W81
–1
–1
W80
x(3)
x(7)
X(2)
–1
W82
–1
–1
W80
W82
–1
W83
–1
–1
N
log2N
2
Bit-Reversed
Inputs
X(4)
X(5)
X(6)
X(7)
Complex
Multiplications
Figure 5.15
The bit reversal algorithm used to perform this re-ordering is shown in Figure 5.16.
The decimal index, n, is converted to its binary equivalent. The binary bits are then
placed in reverse order, and converted back to a decimal number. Bit reversing is
often performed in DSP hardware in the data address generator (DAG), thereby
simplifying the software, reducing overhead, and speeding up the computations.
The computation of the FFT using decimation-in-frequency (DIF) is shown in
Figures 5.17 and 5.18. This method requires that the bit reversal algorithm be
applied to the output X(k). Note that the butterfly for the DIF algorithm differs
slightly from the decimation-in-time butterfly as shown in Figure 5.19.
The use of decimation-in-time versus decimation-in-frequency algorithms is largely
a matter of preference, as either yields the same result. System constraints may
make one of the two a more optimal solution.
It should be noted that the algorithms required to compute the inverse FFT are
nearly identical to those required to compute the FFT, assuming complex FFTs are
used. In fact, a useful method for verifying a complex FFT algorithm consists of first
taking the FFT of the x(n) time samples and then taking the inverse FFT of the
X(k). At the end of this process, the original time samples, Re x(n), should be
obtained and the imaginary part, Im x(n), should be zero (within the limits of the
mathematical round off errors).
5.13
FAST FOURIER TRANSFORMS
BIT REVERSAL EXAMPLE FOR N = 8
■ Decimal Number :
0
1
2
3
4
5
6
7
■ Binary Equivalent :
000
001
010
011
100
101
110
111
■ Bit-Reversed Binary :
000
100
010
110
001
101
011
111
■ Decimal Equivalent :
0
4
2
6
1
5
3
7
Figure 5.16
COMPUTATION OF AN 8-POINT DFT IN THREE STAGES
USING DECIMATION-IN-FREQUENCY
x(0)
x(1)
REDUCE
TO 2-POINT
DFTs
x(2)
x(3)
x(4)
x(5)
x(6)
REDUCE
TO 4-POINT
DFTs
REDUCE
TO 2-POINT
DFTs
x(7)
Figure 5.17
5.14
2-POINT
DFT
2-POINT
DFT
2-POINT
DFT
2-POINT
DFT
X(0)
X(4)
X(2)
X(6)
X(1)
X(5)
X(3)
X(7)
FAST FOURIER TRANSFORMS
EIGHT-POINT DECIMATION-IN-FREQUENCY
FFT ALGORITHM
STAGE 1
STAGE 2
STAGE 3
X(0)
x(0)
W80
x(1)
–1
W80
x(2)
X(2)
–1
W82
x(3)
–1
W80
–1
W80
x(4)
W81
W80
–1
–1
W82
x(6)
W80
W83
W82
–1
–1
X(5)
X(3)
–1
–1
x(7)
X(6)
X(1)
–1
x(5)
X(4)
W80
–1
X(7)
Bit-Reversed
Outputs
Figure 5.18
THE BASIC BUTTERFLY COMPUTATION IN
THE DECIMATION-IN-FREQUENCY FFT ALGORITHM
+
a
Σ
A=a+b
+
WN r
+
b
–
B = (a – b)WNr
Σ
SIMPLIFIED REPRESENTATION
a
A=a+b
WNr
b
–1
B = (a – b)WNr
Figure 5.19
5.15
FAST FOURIER TRANSFORMS
The FFTs discussed up to this point are radix-2 FFTs, i.e., the computations are
based on 2-point butterflies. This implies that the number of points in the FFT must
be a power of 2. If the number of points in an FFT is a power of 4, however, the FFT
can be broken down into a number of 4-point DFTs as shown in Figure 5.20. This is
called a radix-4 FFT. The fundamental decimation-in-time butterfly for the radix-4
FFT is shown in Figure 5.21.
The radix-4 FFT requires fewer complex multiplications but more additions than
the radix-2 FFT for the same number of points. Compared to the radix-2 FFT, the
radix-4 FFT trades more complex data addressing and twiddle factors with less
computation. The resulting savings in computation time varies between different
DSPs but a radix-4 FFT can be as much as twice as fast as a radix-2 FFT for DSPs
with optimal architectures.
COMPUTATION OF A 16-POINT DFT IN THREE STAGES
USING RADIX-4 DECIMATION-IN-TIME ALGORITHM
x(0)
x(4)
x(8)
X(0)
4-POINT
DFT
x(12)
x(1)
x(5)
x(9)
4-POINT
DFT
X(1)
x(2)
x(10)
4-POINT
DFT
x(14)
x(3)
x(7)
x(11)
4-POINT
DFT
COMBINE
4-POINT
DFTs
X(4)
X(5)
X(6)
X(7)
X(8)
X(9)
X(10)
X(11)
X(12)
X13)
X(14)
X(15)
x(15)
Figure 5.20
5.16
X(3)
COMBINE
8-POINT
DFTs
x(13)
x(6)
X(2)
COMBINE
4-POINT
DFTs
FAST FOURIER TRANSFORMS
RADIX-4 FFT DECIMATION-IN-TIME BUTTERFLY
WN0
WNq
–j
–1
j
–1
1
WN2q
–1
j
–1
–j
WN3q
Figure 5.21
FFT HARDWARE IMPLEMENTATION AND
BENCHMARKS
In general terms, the memory requirements for an N-point FFT are N locations for
the real data, N locations for the imaginary data, and N locations for the sinusoid
data (sometimes referred to as twiddle factors). Additional memory locations will be
required if windowing is used. Assuming the memory requirements are met, the
DSP must perform the necessary calculations in the required time. Many DSP
vendors will either give a performance benchmark for a specified FFT size or
calculation time for a butterfly. When comparing FFT specifications, it is important
to make sure that the same type of FFT is used in all cases. For example, the 1024point FFT benchmark on one DSP derived from a radix-2 FFT should not be
compared with the radix-4 benchmark from another DSP.
Another consideration regarding FFTs is whether to use a fixed-point or a floatingpoint processor. The results of a butterfly calculation can be larger than the inputs
to the butterfly. This data growth can pose a potential problem in a DSP with a
fixed number of bits. To prevent data overflow, the data needs to be scaled
beforehand leaving enough extra bits for growth. Alternatively, the data can be
scaled after each stage of the FFT computation. The technique of scaling data after
each pass of the FFT is known as block floating point. It is called this because a full
array of data is scaled as a block regardless of whether or not each element in the
block needs to be scaled. The complete block is scaled so that the relative
relationship of each data word remains the same. For example, if each data word is
shifted right by one bit (divided by 2), the absolute values have been changed, but
relative to each other, the data stays the same.
5.17
FAST FOURIER TRANSFORMS
In a 16-bit fixed-point DSP, a 32-bit word is obtained after multiplication. The
Analog Device’s 21xx-series DSPs have extended dynamic range by providing a 40bit internal register in the multiply-accumulator (MAC).
The use of a floating-point DSP eliminates the need for data scaling and therefore
results in a simpler FFT routine, however the tradeoff is the increased processing
time required for the complex floating-point arithmetic. In addition, a 32-bit
floating-point DSP will obviously have less round off noise than a 16-bit fixed-point
DSP. Figure 5.22 summarizes the FFT benchmarks for popular Analog Devices’
DSPs. Notice in particular that the ADSP-TS001 TigerSHARC™ DSP offers both
fixed-point and floating-point modes, thereby providing an exceptional degree of
programming flexibility.
RADIX-2 COMPLEX FFT
HARDWARE BENCHMARK COMPARISONS
■ ADSP-2189M, 16-bit, Fixed-Point @ 75MHz
◆ 453µs (1024-Point)
■ ADSP-21160 SHARC™, 32-bit, Floating-Point @ 100MHz
◆ 180µs (1024-Point), 2 channels, SIMD Mode
◆ 115µs (1024-Point), 1 channel, SIMD Mode
■ ADSP-TS001 TigerSHARC™ @ 150MHz,
◆ 16-bit, Fixed-Point Mode
● 7.3µs (256-Point FFT)
◆ 32-bit, Floating-Point Mode
● 69µs (1024-Point)
Figure 5.22
DSP REQUIREMENTS FOR REAL TIME FFT
APPLICATIONS
There are two basic ways to acquire data from a real-world signal, either one
sample at a time (continuous processing), or one frame at a time (batch processing).
Sample-based systems, like a digital filter, acquire data one sample at a time. For
each sample clock, a sample comes into the system, and a processed sample is sent
to the output. Frame-based systems, like an FFT-based digital spectrum analyzer,
acquire a frame (or block of samples). Processing occurs on the entire frame of data
and results in a frame of transformed output data.
In order to maintain real time operation, the entire FFT must therefore be
calculated during the frame period. This assumes that the DSP is collecting the
data for the next frame while it is calculating the FFT for the current frame of data.
Acquiring the data is one area where special architectural features of DSPs come
into play. Seamless data acquisition is facilitated by the DSP’s flexible data
5.18
FAST FOURIER TRANSFORMS
addressing capabilities in conjunction with its direct memory accessing (DMA)
channels.
Assume the DSP is the ADSP-TS001 TigerSHARC which can calculate a 1024-point
32-bit complex floating-point FFT in 69µs. The maximum sampling frequency is
therefore 1024/69µs = 14.8MSPS. This implies a signal bandwidth of less than
7.4MHz. It is also assumed that there is no additional FFT overhead or data
transfer limitation.
REAL-TIME FFT PROCESSING EXAMPLE
■ Assume 69µs Execution Time for Radix-2, 1024-point FFT
(TigerSHARC, 32-bit Mode)
■ fs (maximum) <
■
1024 Samples
69µs
= 14.8MSPS
Therefore Input Signal Bandwidth < 7.4MHz
■ This Assumes No Additional FFT Overhead and
No Input/Output Data Transfer Limitations
Figure 5.23
The above example will give an estimate of the maximum bandwidth signal which
can be handled by a given DSP using its FFT benchmarks. Another way to approach
the issue is to start with the signal bandwidth and develop the DSP requirements.
If the signal bandwidth is known, the required sampling frequency can be estimated
by multiplying by a factor between 2 and 2.5 (the increased sampling rate may be
required to ease the requirements on the antialiasing filter which precedes the
ADC). The next step is to determine the required number of points in the FFT to
achieve the desired frequency resolution. The frequency resolution is obtained by
dividing the sampling rate fs by N, the number of points in the FFT. These and
other FFT considerations are shown in Figure 5.24.
The number of FFT points also determines the noise floor of the FFT with respect to
the broadband noise level, and this may also be a consideration. Figure 5.25 shows
the relationships between the system fullscale signal level, the broadband noise
level (measured over the bandwidth DC to fs/2), and the FFT noise floor. Notice that
the FFT processing gain is determined by the number of points in the FFT. The FFT
acts like an analog spectrum analyzer with a sweep bandwidth of fs/N. Increasing
the number of points increases the FFT resolution and narrows its bandwidth,
thereby reducing the noise floor. This analysis neglects noise caused by the FFT
round off error. In practice, the ADC which is used to digitize the signal produces
quantization noise which is the dominant noise source.
At this point it is time to examine actual DSPs and their FFT processing times to
make sure real time operation can be achieved. This means that the FFT must be
5.19
FAST FOURIER TRANSFORMS
calculated during the acquisition time for one frame of data which is N/fs. Other
considerations such as fixed-point vs. floating-point, radix-2 vs. radix-4, and
processor power dissipation and cost may be other considerations.
REAL TIME FFT CONSIDERATIONS
■ Signal Bandwidth
■ Sampling Frequency, fs
■ Number of Points in FFT, N
■ Frequency Resolution = fs / N
■ Maximum Time to Calculate N-Point FFT = N / fs
■ Fixed-Point vs. Floating Point DSP
■ Radix-2 vs. Radix-4 Execution Time
■ FFT Processing Gain = 10 log10(N / 2)
■ Windowing Requirements
Figure 5.24
FFT PROCESSING GAIN
NEGLECTING ROUND OFF ERROR
SIGNAL
LEVEL
(dB)
Fullscale RMS Signal Level
SNR
RMS Noise Level (DC to fs / 2)
FFT Processing Gain =
10 log10
N
2
= 27dB for N = 1024
= 30dB for N = 2048
= 33dB for N = 4096
FFT
Noise Floor
f
fs
N
Figure 5.25
5.20
fs
2
FAST FOURIER TRANSFORMS
SPECTRAL LEAKAGE AND WINDOWING
Spectral leakage in FFT processing can best be understood by considering the case
of performing an N-point FFT on a pure sinusoidal input. Two conditions will be
considered. In Figure 5.26, the ratio between the sampling frequency and the input
sinewave frequency is such that precisely an integral number of cycles are contained
within the data window (frame, or record). Recall that the DFT assumes that an
infinite number of these windows are placed end-to-end to form a periodic waveform
as shown in the diagram as the periodic extensions. Under these conditions, the
waveform appears continuous (no discontinuities), and the DFT or FFT output will
be a single tone located at the input signal frequency.
FFT OF SINEWAVE HAVING INTEGRAL
NUMBER OF CYCLES IN DATA WINDOW
fin
0
N-1 0
Periodic
Extension
N-1 0
N-1
t
Periodic
Extension
Data
Window
fin
N
fin
= C
N
fs
N = Record Length
NC = Number of Cycles
in Data Window
fs/N
f
Figure 5.26
Figure 5.27 shows the condition where there are not an integral number of sinewave
cycles within the data window. The discontinuities which occur at the endpoints of
the data window result in leakage in the frequency domain because of the
harmonics which are generated. In addition to the sidelobes, the main lobe of the
sinewave is smeared over several frequency bins. This process is equivalent to
multiplying the input sinewave by a rectangular window pulse which has the
familiar sin(x)/x frequency response and associated smearing and sidelobes.
5.21
FAST FOURIER TRANSFORMS
FFT OF SINEWAVE HAVING NON-INTEGRAL
NUMBER OF CYCLES IN DATA WINDOW
fin
N-1
0
0
N-1 0
0dB
t
Periodic
Extension
Data
Window
Periodic
Extension
N-1
fin
NC
fin
≠
N
fs
N = Record Length
NC = Number of Cycles
in Data Window
–12dB
6dB/OCT
f
fs/N
Figure 5.27
Notice that the first sidelobe is only 12dB below the fundamental, and that the
sidelobes roll off at only 6dB/octave. This situation would be unsuitable for most
spectral analysis applications. Since in practical FFT spectral analysis applications
the exact input frequencies are unknown, something must be done to minimize
these sidelobes. This is done by choosing a window function other than the
rectangular window. The input time samples are multiplied by an appropriate
window function which brings the signal to zero at the edges of the window as
shown in Figure 5.28. The selection of a window function is primarily a tradeoff
between main lobe spreading and sidelobe roll off. Reference 7 is highly
recommended for an in-depth treatment of windows.
The mathematical functions which describe four popular window functions
(Hamming, Blackman, Hanning, and Minimum 4-term Blackman-Harris) are
shown in Figure 5.29. The computations are straightforward, and the window
function data points are usually precalculated and stored in the DSP memory to
minimize their impact on FFT processing time. The frequency response of the
rectangular, Hamming, and Blackman windows are shown in Figure 5.30. Figure
5.31 shows the tradeoff between main lobe spreading and sidelobe amplitude and
roll off for the popular window functions.
5.22
FAST FOURIER TRANSFORMS
WINDOWING TO REDUCE SPECTRAL LEAKAGE
Input Data
x(n)
t
Window
Function
w(n)
t
Windowed
Input Data
w(n)•x(n)
t
n=0
n=N–1
Data
Window
Figure 5.28
SOME POPULAR WINDOW FUNCTIONS
Hamming:
w(n) = 0.54 – 0.46 cos
2π
πn
N
Blackman:
w(n) = 0.42
– 0.5 cos
πn
2π
N
Hanning:
w(n) =
– 0.5 cos
πn
2π
N
Minimum
4-Term
Blackman
Harris
w(n) = 0.35875
0 ≤ n ≤ N–1
0.5
πn
+ 0.08 cos 4π
N
– 0.48829 cos
πn
2π
N
+ 0.14128 cos
πn
4π
N
– 0.01168 cos
6π
πn
N
Figure 5.29
5.23
FAST FOURIER TRANSFORMS
FREQUENCY RESPONSE OF RECTANGULAR,
HAMMING, AND BLACKMAN WINDOWS FOR N = 256
dB
RECTANGULAR
0
BIN WIDTH =
fs
N
–40
–80
–120
–10
–5
0
5
FREQUENCY BINS
dB
HAMMING
dB
0
0
–40
–40
–80
–80
–120
–10
–5
0
5
FREQUENCY BINS
–120
10
10
BLACKMAN
–10
–5
0
5
FREQUENCY BINS
10
Figure 5.30
POPULAR WINDOWS AND FIGURES OF MERIT
WINDOW
FUNCTION
3dB BW
(Bins)
6dB BW
(Bins)
HIGHEST
SIDELOBE
(dB)
SIDELOBE
ROLLOFF
(dB/Octave)
Rectangle
0.89
1.21
–12
6
Hamming
1.3
1.81
– 43
6
Blackman
1.68
2.35
–58
18
Hanning
1.44
2.00
–32
18
Minimum
4-Term
BlackmanHarris
1.90
2.72
–92
6
Figure 5.31
5.24
FAST FOURIER TRANSFORMS
REFERENCES
1.
Steven W. Smith, The Scientist and Engineer’s Guide to Digital Signal
Processing, Second Edition, 1999, California Technical Publishing,
P.O. Box 502407, San Diego, CA 92150. Also available for free download at:
http://www.dspguide.com or http://www.analog.com
2.
C. Britton Rorabaugh, DSP Primer, McGraw-Hill, 1999.
3.
Richard J. Higgins, Digital Signal Processing in VLSI, Prentice-Hall,
1990.
4.
A. V. Oppenheim and R. W. Schafer, Digital Signal Processing, PrenticeHall, 1975.
5.
L. R. Rabiner and B. Gold, Theory and Application of Digital Signal
Processing, Prentice-Hall, 1975.
6.
John G. Proakis and Dimitris G. Manolakis, Introduction to Digital
Signal Processing, MacMillian, 1988.
7.
Fredrick J. Harris, On the Use of Windows for Harmonic Analysis with
the Discrete Fourier Transform, Proc. IEEE, Vol. 66, No. 1, 1978 pp. 51-83.
8.
R. W. Ramirez, The FFT: Fundamentals and Concepts,
Prentice-Hall, 1985.
9.
J. W. Cooley and J. W. Tukey, An Algorithm for the Machine Computation
of Complex Fourier Series, Mathematics Computation, Vol. 19,
pp. 297-301, April 1965.
10.
Digital Signal Processing Applications Using the ADSP-2100
Family, Vol. 1 and Vol. 2, Analog Devices, Free Download at:
http://www.analog.com
11.
ADSP-21000 Family Application Handbook, Vol. 1, Analog Devices,
Free Download at: http://www.analog.com
5.25
FAST FOURIER TRANSFORMS
5.26
DIGITAL FILTERS
SECTION 6
DIGITAL FILTERS
■
Finite Impulse Response (FIR) Filters
■
Infinite Impulse Response (IIR) Filters
■
Multirate Filters
■
Adaptive Filters
6.a
DIGITAL FILTERS
6.b
DIGITAL FILTERS
SECTION 6
DIGITAL FILTERS
Walt Kester
INTRODUCTION
Digital filtering is one of the most powerful tools of DSP. Apart from the obvious
advantages of virtually eliminating errors in the filter associated with passive
component fluctuations over time and temperature, op amp drift (active filters), etc.,
digital filters are capable of performance specifications that would, at best, be
extremely difficult, if not impossible, to achieve with an analog implementation. In
addition, the characteristics of a digital filter can be easily changed under software
control. Therefore, they are widely used in adaptive filtering applications in
communications such as echo cancellation in modems, noise cancellation, and
speech recognition.
The actual procedure for designing digital filters has the same fundamental
elements as that for analog filters. First, the desired filter responses are
characterized, and the filter parameters are then calculated. Characteristics such as
amplitude and phase response are derived in the same way. The key difference
between analog and digital filters is that instead of calculating resistor, capacitor,
and inductor values for an analog filter, coefficient values are calculated for a digital
filter. So for the digital filter, numbers replace the physical resistor and capacitor
components of the analog filter. These numbers reside in a memory as filter
coefficients and are used with the sampled data values from the ADC to perform the
filter calculations.
The real-time digital filter, because it is a discrete time function, works with
digitized data as opposed to a continuous waveform, and a new data point is
acquired each sampling period. Because of this discrete nature, data samples are
referenced as numbers such as sample 1, sample 2, sample 3, etc. Figure 6.1 shows
a low frequency signal containing higher frequency noise which must be filtered out.
This waveform must be digitized with an ADC to produce samples x(n). These data
values are fed to the digital filter, which in this case is a lowpass filter. The output
data samples, y(n), are used to reconstruct an analog waveform using a low glitch
DAC.
Digital filters, however, are not the answer to all signal processing filtering
requirements. In order to maintain real-time operation, the DSP processor must be
able to execute all the steps in the filter routine within one sampling clock period,
1/fs. A fast general purpose fixed-point DSP such as the ADSP-2189M at 75MIPS
can execute a complete filter tap multiply-accumulate instruction in 13.3ns. The
ADSP-2189M requires N+5 instructions for an N-tap filter. For a 100-tap filter, the
total execution time is approximately 1.4µs. This corresponds to a maximum
possible sampling frequency of 714kHz, thereby limiting the upper signal
bandwidth to a few hundred kHz.
6.1
DIGITAL FILTERS
DIGITAL FILTERING
H(f)
t
t
f
ANALOG
ANTIALIASING
FILTER
x(n)
ADC
DIGITAL
LOWPASS
FILTER
fs
y(n)
DAC
ANALOG
ANTI-IMAGING
FILTER
fs
Figure 6.1
However, it is possible to replace a general purpose DSP chip and design special
hardware digital filters which will operate at video-speed sampling rates. In other
cases, the speed limitations can be overcome by first storing the high speed ADC
data in a buffer memory. The buffer memory is then read at a rate which is
compatible with the speed of the DSP-based digital filter. In this manner, pseudorealtime operation can be maintained as in a radar system, where signal processing
is typically done on bursts of data collected after each transmitted pulse.
Another option is to use a third-party dedicated DSP filter engine like the Systolix
PulseDSP™ filter core. The AD7725 16-bit sigma-delta ADC has an on-chip
PulseDSP filter which can do 125 million multiply-accumulates per second.
Even in highly oversampled sampled data systems, an analog antialiasing filter is
still required ahead of the ADC and a reconstruction (anti-imaging) filter after the
DAC. Finally, as signal frequencies increase sufficiently, they surpass the
capabilities of available ADCs, and digital filtering then becomes impossible. Active
analog filtering is not possible at extremely high frequencies because of op amp
bandwidth and distortion limitations, and filtering requirements must then be met
using purely passive components. The primary focus of the following discussions
will be on filters which can run in real-time under DSP program control.
As an example, consider the comparison between an analog and a digital filter
shown in Figure 6.3. The cutoff frequency of the both filters is 1kHz. The analog
filter is realized as a 6-pole Chebyshev Type 1 filter (ripple in passband, no ripple in
stopband). In practice, this filter would probably be realized using three 2-pole
stages, each of which requires an op amp, and several resistors and capacitors. The
6-pole design is certainly not trivial, and maintaining the 0.5dB ripple specification
requires accurate component selection and matching.
6.2
DIGITAL FILTERS
On the other hand, the digital FIR filter shown has only 0.002dB passband ripple,
linear phase, and a much sharper roll off. In fact, it could not be realized using
analog techniques! In a practical application, there are many other factors to
consider when evaluating analog versus digital filters. Most modern signal
processing systems use a combination of analog and digital techniques in order to
accomplish the desired function and take advantage of the best of both the analog
and the digital world.
DIGITAL VERSUS ANALOG FILTERING
DIGITAL FILTERS
ANALOG FILTERS
High Accuracy
Less Accuracy - Component
Tolerances
Linear Phase (FIR Filters)
Non-Linear Phase
No Drift Due to Component
Variations
Drift Due to Component
Variations
Flexible, Adaptive Filtering Possible
Adaptive Filters Difficult
Easy to Simulate and Design
Difficult to Simulate and Design
Computation Must be Completed in
Sampling Period - Limits Real Time
Operation
Analog Filters Required at
High Frequencies and for
Anti-Aliasing Filters
Requires High Performance ADC,
DAC & DSP
No ADC, DAC, or DSP Required
Figure 6.2
There are many applications where digital filters must operate in real-time. This
places specific requirements on the DSP depending upon the sampling frequency
and the filter complexity. The key point is that the DSP must finish all
computations during the sampling period so it will be ready to process the next data
sample. Assume that the analog signal bandwidth to be processed is fa. This
requires the ADC sampling frequency fs to be at least 2fa. The sampling period is
1/fs. All DSP filter computations (including overhead) must be completed during
this interval. The computation time depends on the number of taps in the filter and
the speed and efficiency of the DSP. Each tap on the filter requires one
multiplication and one addition (multiply-accumulate). DSPs are generally
optimized to perform fast multiply-accumulates, and many DSPs have additional
features such as circular buffering and zero-overhead looping minimize the
“overhead” instructions that otherwise would be needed.
6.3
DIGITAL FILTERS
ANALOG VERSUS DIGITAL FILTER
FREQUENCY RESPONSE COMPARISON
dB
ANALOG FILTER
DIGITAL FILTER
Chebyshev Type 1
6 Pole, 0.5dB Ripple
FIR, 129-Tap, 0.002dB Ripple,
Linear Phase, fs = 10kSPS
dB
0
0
–20
–20
–40
–40
–60
–60
–80
–80
–100
–100
0
1
2
3
4
5
0
FREQUENCY (kHz)
1
2
3
FREQUENCY (kHz)
Figure 6.3
PROCESSING REQUIREMENTS
FOR REAL TIME DIGITAL FILTERING
■ Signal Bandwidth = fa
■ Sampling Frequency fs > 2fa
■ Sampling Period = 1 / fs
■ Filter Computational Time + Overhead < Sampling Period
◆ Depends on Number of Taps
◆ Speed of DSP Multiplication-Accumulates (MACs)
◆ Efficiency of DSP
●
Circular Buffering
●
Zero Overhead Looping
●
etc.
Figure 6.4
6.4
4
5
DIGITAL FILTERS
FINITE IMPULSE RESPONSE (FIR) FILTERS
There are two fundamental types of digital filters: finite impulse response (FIR) and
infinite impulse response (IIR). As the terminology suggests, these classifications
refer to the filter’s impulse response. By varying the weight of the coefficients and
the number of filter taps, virtually any frequency response characteristic can be
realized with an FIR filter. As has been shown, FIR filters can achieve performance
levels which are not possible with analog filter techniques (such as perfect linear
phase response). However, high performance FIR filters generally require a large
number of multiply-accumulates and therefore require fast and efficient DSPs. On
the other hand, IIR filters tend to mimic the performance of traditional analog
filters and make use of feedback. Therefore their impulse response extends over an
infinite period of time. Because of feedback, IIR filters can be implemented with
fewer coefficients than for an FIR filter. Lattice filters are simply another way to
implement either FIR or IIR filters and are often used in speech processing
applications. Finally, digital filters lend themselves to adaptive filtering
applications simply because of the speed and ease with which the filter
characteristics can be changed by varying the filter coefficients.
TYPES OF DIGITAL FILTERS
■ Moving Average
■ Finite Impulse Response (FIR)
◆ Linear Phase
◆ Easy to Design
◆ Computationally Intensive
■ Infinite Impulse Response (IIR)
◆ Based on Classical Analog Filters
◆ Computationally Efficient
■ Lattice Filters (Can be FIR or IIR)
■ Adaptive Filters
Figure 6.5
The most elementary form of an FIR filter is a moving average filter as shown in
Figure 6.6. Moving average filters are popular for smoothing data, such as in the
analysis of stock prices, etc. The input samples, x(n) are passed through a series of
buffer registers (labeled z–1, corresponding to the z-transform representation of a
delay element). In the example shown, there are four taps corresponding to a fourpoint moving average. Each sample is multiplied by 0.25, and these results are
added to yield the final moving average output y(n). The figure also shows the
general equation of a moving average filter with N taps. Note again that N refers to
the number of filter taps, and not the ADC or DAC resolution as in previous
sections.
6.5
DIGITAL FILTERS
4-POINT MOVING AVERAGE FILTER
x(n)
x(n–1)
x(n–2)
Z–1
h(0)
=
1
4
Z–1
h(1)
=
x(n–3)
Z–1
h(2)
1
4
=
h(3)
1
4
=
1
4
Σ
y(n)
y(n) = h(0) x(n) + h(1) x(n – 1) + h(2) x(n – 2) + h(3) x(n – 3)
= 1 x(n) + 1
4
= 1 x(n)
4
4
+
1 x(n – 2) +
4
x(n – 1) +
x(n – 1)
For N-Point
Moving Average Filter:
+
x(n – 2)
1 x(n – 3)
4
+
x(n – 3)
N–1
y(n) =
1
x(n – k)
Nk=0
Σ
Figure 6.6
Since the coefficients are equal, an easier way to perform a moving average filter is
shown in Figure 6.7. Note that the first step is store the first four samples, x(0),
x(1), x(2), x(3) in a register. These quantities are added and then multiplied by 0.25
to yield the first output, y(3). Note that the initial outputs y(0), y(1), and y(2) are not
valid because all registers are not full until sample x(3) is received.
When sample x(4) is received, it is added to the result, and sample x(0) is subtracted
from the result. The new result must then be multiplied by 0.25. Therefore, the
calculations required to produce a new output consist of one addition, one
subtraction, and one multiplication, regardless of the length of the moving average
filter.
The step function response of a 4-point moving average filter is shown in Figure 6.8.
Notice that the moving average filter has no overshoot. This makes it useful in
signal processing applications where random white noise must be filtered but pulse
response preserved. Of all the possible linear filters that could be used, the moving
average produces the lowest noise for a given edge sharpness. This is illustrated in
Figure 6.9, where the noise level becomes lower as the number of taps are
increased. Notice that the 0% to 100% risetime of the pulse response is equal to the
total number of taps in the filter multiplied by the sampling period.
6.6
DIGITAL FILTERS
CALCULATING OUTPUT OF
4-POINT MOVING AVERAGE FILTER
y(3) = 0.25
x(3) + x(2) + x(1) + x(0)
y(4) = 0.25
x(4) + x(3) + x(2) + x(1)
y(5) = 0.25
x(5) + x(4) + x(3) + x(2)
x(6) + x(5) + x(4) + x(3)
y(6) = 0.25
y(7) = 0.25
x(7) + x(6) + x(5) + x(4)
Each Output Requires:
1 multiplication, 1 addition, 1 subtraction
Figure 6.7
4-TAP MOVING AVERAGE FILTER STEP RESPONSE
= Input
x(n)
= Output
y(n)
n
0
1
2
3
4
5
6
7
8
9
10
11
12
N–1
General:
1
y(n) =
x(n – k)
Nk=0
For N = 4:
1
y(n) =
x(n – k)
4 k=0
Σ
3
Σ
Figure 6.8
6.7
DIGITAL FILTERS
MOVING AVERAGE FILTER RESPONSE
TO NOISE SUPERIMPOSED ON STEP INPUT
Note:
Latency not shown
Note:
Latency not shown
0
0
Sample Number
Sample Number
500
0
500
Sample Number
500
Figure 6.9
The frequency response of the simple moving average filter is sin(x)/x and is shown
on a linear amplitude scale in Figure 6.10. Adding more taps to the filter sharpens
the rolloff, but does not significantly reduce the amplitude of the sidelobes which
are approximately 14dB down for the 11 and 31-tap filter. These filters are
definitely not suitable where high stopband attenuation is required.
MOVING AVERAGE FILTER FREQUENCY RESPONSE
1.0
0.8
3 point
Amplitude
0.6
11 point
0.4
31 point
0.2
0
0
0.1
0.2
0.3
Frequency (fraction of fs)
Figure 6.10
6.8
0.4
0.5
DIGITAL FILTERS
It is possible to dramatically improve the performance of the simple FIR moving
average filter by properly selecting the individual weights or coefficients rather than
giving them equal weight. The sharpness of the rolloff can be improved by adding
more stages (taps), and the stopband attenuation characteristics can be improved by
properly selecting the filter coefficients. Note that unlike the moving average filter,
one multiply-accumulate cycle is now required per tap for the generalized FIR filter.
The essence of FIR filter design is the appropriate selection of the filter coefficients
and the number of taps to realize the desired transfer function H(f). Various
algorithms are available to translate the frequency response H(f) into a set of FIR
coefficients. Most of this software is commercially available and can be run on PCs.
The key theorem of FIR filter design is that the coefficients h(n) of the FIR filter are
simply the quantized values of the impulse response of the frequency transfer
function H(f). Conversely, the impulse response is the discrete Fourier transform of
H(f).
N-TAP FINITE IMPULSE RESPONSE (FIR) FILTER
x(n)
x(n–1)
Z–1
h(0)
x(n–N+2)
Z–1
h(1)
x(n–N+1)
Z–1
h(N–2)
h(N–1)
Σ
y(n)
N–1
y(n) = h(n) ✳ x(n)
=
Σ
k=0
h(k) x(n – k)
✳ = Symbol for Convolution
Requires N multiply-accumulates for each output
Figure 6.11
The generalized form of an N-tap FIR filter is shown in Figure 6.11. As has been
discussed, an FIR filter must perform the following convolution equation:
y(n) = h(k)*x(n) =
N −1
∑ h(k )x(n − k ) .
k =0
where h(k) is the filter coefficient array and x(n-k) is the input data array to the
filter. The number N, in the equation, represents the number of taps of the filter
and relates to the filter performance as has been discussed above. An N-tap FIR
filter requires N multiply-accumulate cycles.
6.9
DIGITAL FILTERS
FIR filter diagrams are often simplified as shown in Figure 6.12. The summations
are represented by arrows pointing into the dots, and the multiplications are
indicated by placing the h(k) coefficients next to the arrows on the lines. The z–1
delay element is often shown by placing the label above or next to the appropriate
line.
SIMPLIFIED FILTER NOTATIONS
x(n)
Z–1
h(0)
x(n–1)
x(n–N+2)
Z–1
h(N–2)
h(1)
Σ
x(n)
h(0)
Z–1
x(n–1)
h(1)
Z–1
x(n–N+1)
h(N–1)
Σ
Z–1
Σ
x(n–N+2) Z–1
h(N–2)
y(n)
x(n–N+1)
h(N–1)
y(n)
Figure 6.12
FIR FILTER IMPLEMENTATION IN DSP HARDWARE USING
CIRCULAR BUFFERING
In the series of FIR filter equations, the N coefficient locations are always accessed
sequentially from h(0) to h(N–1). The associated data points circulate through the
memory; new samples are added replacing the oldest each time a filter output is
computed. A fixed boundary RAM can be used to achieve this circulating buffer
effect as shown in Figure 6.13 for a 4 tap FIR filter. The oldest data sample is
replaced by the newest after each convolution. A "time history" of the four most
recent data samples is always stored in RAM.
To facilitate memory addressing, old data values are read from memory starting
with the value one location after the value that was just written. For example, x(4)
is written into memory location 0, and data values are then read from locations 1, 2,
3, and 0. This example can be expanded to accommodate any number of taps. By
addressing data memory locations in this manner, the address generator need only
supply sequential addresses, regardless of whether the operation is a memory read
or write. This data memory buffer is called circular because when the last location
is reached, the memory pointer is reset to the beginning of the buffer.
6.10
DIGITAL FILTERS
CALCULATING OUTPUTS OF 4-TAP FIR
FILTER USING A CIRCULAR BUFFER
Memory
Location
Read
Write
Read
0
x(0)
x(4)
x(4)
1
x(1)
x(1)
2
x(2)
x(2)
x(2)
3
x(3)
x(3)
x(3)
y(3) =
y(4) =
Write
Read
x(4)
x(5)
x(5)
h(0) x(3) + h(1) x(2) + h(2) x(1) + h(3) x(0)
h(0) x(4) + h(1) x(3) + h(2) x(2) + h(3) x(1)
y(5) = h(0) x(5) + h(1) x(4) + h(2) x(3) + h(3) x(2)
Figure 6.13
The coefficients are fetched simultaneously with the data. Due to the addressing
scheme chosen, the oldest data sample is fetched first. Therefore, the last coefficient
must be fetched first. The coefficients can be stored backwards in memory: h(N-1) is
the first location, and h(0) is the last, with the address generator providing
incremental addresses. Alternatively, coefficients can be stored in a normal manner
with the accessing of coefficients starting at the end of the buffer, and the address
generator being decremented. In the example shown in Figure 6.13, the coefficients
are stored in a reverse manner.
A simple summary flowchart for these operations is shown in Figure 6.14. For
Analog Devices’ DSPs, all operations within the filter loop are completed in one
instruction cycle, thereby greatly increasing efficiency. This is referred to as zerooverhead looping. The actual FIR filter assembly code for the ADSP-21XX family of
fixed point DSPs is shown in Figure 6.15. The arrows in the diagram point to the
actual executable instructions, and the rest of the code are simply comments added
for clarification.
6.11
DIGITAL FILTERS
The first instruction (labeled fir:) sets up the computation by clearing the MR
register and loading the MX0 and MY0 registers with the first data and coefficient
values from data and program memory. The multiply-accumulate with dual data
fetch in the convolution loop is then executed N–1 times in N cycles to compute the
sum of the first N–1 products. The final multiply-accumulate instruction is
performed with the rounding mode enabled to round the result to the upper 24 bits
of the MR register. The MR1 register is then conditionally saturated to its most
positive or negative value based on the status of the overflow flag contained in the
MV register. In this manner, results are accumulated to the full 40-bit precision of
the MR register, with saturation of the output only if the final result overflowed
beyond the least significant 32 bits of the MR register.
The limit on the number of filter taps attainable for a real-time implementation of
the FIR filter subroutine is determined primarily by the processor cycle time, the
sampling rate, and the number of other computations required. The FIR subroutine
presented here requires a total of N+5 cycles for a filter of length N. For the ADSP2189M 75MIPS DSP, one instruction cycle is 13.3ns, so a 100-tap filter would
require 13.3ns×100 + 5×13.3ns = 1330ns + 66.5ns = 1396.5ns = 1.4µs.
PSEUDOCODE FOR FIR FILTER PROGRAM
USING A DSP WITH CIRCULAR BUFFERING
1. Obtain sample from ADC (typically interrupt driven)
2. Move sample into input signal’s circular buffer
3. Update the pointer for the input signal’s circular buffer
4. Zero the accumulator
5. Implement filter (control the loop through each of the coefficients)
6. Fetch the coefficient from the coefficient’s circular buffer
7. Update the pointer for the coefficient’s circular buffer
8. Fetch the sample from the input signal’s circular buffer
9. Update the pointer for the input signal’s circular buffer
10. Multiply the coefficient by the sample
11. Add the product to the accumulator
12. Move the filtered sample to the DAC
ADSP-21xx Example code:
CNTR = N-1;
DO convolution UNTIL CE;
convolution:
MR = MR + MX0 * MY0(SS), MX0 = DM(I0,M1), MY0 = PM(I4,M5);
Figure 6.14
6.12
DIGITAL FILTERS
ADSP-21XX FIR FILTER ASSEMBLY CODE
(SINGLE PRECISION)
.MODULE
{
fir_sub;
FIR Filter Subroutine
Calling Parameters
I0 --> Oldest input data value in delay line
I4 --> Beginning of filter coefficient table
L0 = Filter length (N)
L4 = Filter length (N)
M1,M5 = 1
CNTR = Filter length - 1 (N-1)
Return Values
MR1 = Sum of products (rounded and saturated)
I0 --> Oldest input data value in delay line
I4 --> Beginning of filter coefficient table
Altered Registers
MX0,MY0,MR
Computation Time
(N - 1) + 6 cycles = N + 5 cycles
All coefficients are assumed to be in 1.15 format. }
.ENTRY
fir:
fir;
MR=0, MX0=DM(I0,M1), MY0=PM(I4,M5);
CNTR = N-1;
DO convolution UNTIL CE;
MR=MR+MX0*MY0(SS), MX0=DM(I0,M1), MY0=PM(I4,M5);
MR=MR+MX0*MY0(RND);
IF MV SAT MR;
RTS;
convolution:
.ENDMOD;
Figure 6.15
DESIGNING FIR FILTERS
FIR filters are relatively easy to design using modern CAD tools. Figure 6.16
summarizes the characteristics of FIR filters as well as the most popular design
techniques. The fundamental concept of FIR filter design is that the filter frequency
response is determined by the impulse response, and the quantized impulse response
and the filter coefficients are identical. This can be understood by examining Figure
6.17. The input to the FIR filter is an impulse, and as the impulse propagates
through the delay elements, the filter output is identical to the filter coefficients.
The FIR filter design process therefore consists of determining the impulse response
from the desired frequency response, and then quantizing the impulse response to
generate the filter coefficients.
It is useful to digress for a moment and examine the relationship between the time
domain and the frequency domain to better understand the principles behind digital
filters such as the FIR filter. In a sampled data system, a convolution operation can
be carried out by performing a series of multiply-accumulates. The convolution
operation in the time or frequency domain is equivalent to point-by-point
multiplication in the opposite domain. For example, convolution in the time domain
is equivalent to multiplication in the frequency domain. This is shown graphically in
Figure 6.18. It can be seen that filtering in the frequency domain can be
accomplished by multiplying all frequency components in the passband by a 1 and
all frequencies in the stopband by 0. Conversely, convolution in the frequency
domain is equivalent to point by point multiplication in the time domain.
6.13
DIGITAL FILTERS
CHARACTERISTICS OF FIR FILTERS
■ Impulse Response has a Finite Duration (N Cycles)
■ Linear Phase, Constant Group Delay (N Must be Odd)
■ No Analog Equivalent
■ Unconditionally Stable
■ Can be Adaptive
■ Computational Advantages when Decimating Output
■ Easy to Understand and Design
◆ Windowed-Sinc Method
◆ Fourier Series Expansion with Windowing
◆ Frequency Sampling Using Inverse FFT - Arbitrary Frequency
Response
◆ Parks-McClellan Program with Remez Exchange Algorithm
Figure 6.16
FIR FILTER IMPULSE RESPONSE
DETERMINES THE FILTER COEFFICIENTS
x(n)
h(0)
Z–1
h(1)
x(n–1)
Z–1
x(n–7)
x(n–8)
Z–1
h(7)
N=9
h(8)
Σ
y(n)
y(n)
x(n)
1
h(4)
h(2)
h(6)
h(8)
h(0)
n
n
h(1)
h(7)
h(3)
Figure 6.17
6.14
h(5)
DIGITAL FILTERS
DUALITY OF TIME AND FREQUENCY
TIME DOMAIN
FREQUENCY DOMAIN
x(n)
X(k)
INPUT
t
f
FFT
FILTER
RESPONSE
1
h(m)
t
OUTPUT
x(n) ✳ h(m)
H(k)
f
0
X(k) • H(k)
t
f
Figure 6.18
The transfer function in the frequency domain (either a 1 or a 0) can be translated
to the time domain by the discrete Fourier transform (in practice, the fast Fourier
transform is used). This transformation produces an impulse response in the time
domain. Since the multiplication in the frequency domain (signal spectrum times
the transfer function) is equivalent to convolution in the time domain (signal
convolved with impulse response), the signal can be filtered by convolving it with
the impulse response. The FIR filter is exactly this process. Since it is a sampled
data system, the signal and the impulse response are quantized in time and
amplitude yielding discrete samples. The discrete samples comprising the desired
impulse response are the FIR filter coefficients.
The mathematics involved in filter design (analog or digital) generally make use of
transforms. In continuous-time systems, the Laplace transform can be considered to
be a generalization of the Fourier Transform. In a similar manner, it is possible to
generalize the Fourier transform for discrete-time sampled data systems, resulting
in what is commonly referred to as the z-transform. Details describing the use of the
z-transform in digital filter design are given in References 1, 2, 3, 4, 5, and 6, but
the theory is not necessary for the rest of this discussion.
6.15
DIGITAL FILTERS
FIR Filter Design Using the Windowed-Sinc Method
An ideal lowpass filter frequency response is shown in Figure 6.19A. The
corresponding impulse response in the time domain is shown in Figure 6.19B, and
follows the sin(x)/x (sinc) function. If an FIR filter is used to implement this
frequency response, an infinite number of taps are required. The windowed-sinc
method is used to implement the filter as follows. First, the impulse response is
truncated to a reasonable number of N taps as in Figure 6.19C. As has been
discussed in Section 5, the frequency response corresponding to Figure 6.19C has
relatively poor sidelobe performance because of the end-point discontinuities in the
truncated impulse response. The next step in the design process is to apply an
appropriate window function as shown in Figure 6.19D to the truncated impulse.
This forces the endpoints to zero. The particular window function chosen determines
the rolloff and sidelobe performance of the filter. Window functions have been
discussed in detail in Section 5, and there are several good choices depending upon
the desired frequency response. The frequency response of the truncated and
windowed-sinc impulse response of Figure 6.19E is shown in Figure 6.19F.
FIR FILTER DESIGN USING
THE WINDOWED-SINC METHOD
Ideal Lowpass Filter
Frequency Response
1
(A)
Ideal Lowpass Filter
Impulse Response
(B)
sin x
x
(sinc function)
0
(C)
t
f
fc
Window
Function
N
Final Filter
Frequency Response
1
(E)
t
N
Figure 6.19
t
N
Windowed
Impulse Response
(D)
6.16
Truncated
Impulse Response
t
0
(F)
fc
f
DIGITAL FILTERS
FIR Filter Design Using the Fourier Series Method with Windowing
The Fourier series with windowing method (Figure 6.20) starts by defining the
transfer function H(f) mathematically and expanding it in a Fourier series. The
Fourier series coefficients define the impulse response and therefore the coefficients
of the FIR filter. However, the impulse response must be truncated and windowed
as in the previous method. After truncation and windowing, an FFT is used to
generate the corresponding frequency response. The frequency response can be
modified by choosing different window functions, although precise control of the
stopband characteristics is difficult in any method which uses windowing.
FIR FILTER DESIGN USING
FOURIER SERIES METHOD WITH WINDOWING
■ Specify H(f)
■ Expand H(f) in a Fourier series: The Fourier series coefficients are
the coefficients of the FIR filter, h(m), and its Impulse Response
■ Truncate the Impulse Response to N points (taps)
■ Apply a suitable Window function to h(m) to smooth the effects of
truncation
■ Lacks precise control of cutoff frequency; Highly dependent
on Window function
Figure 6.20
FIR Filter Design Using the Frequency Sampling Method
This method is extremely useful in generating an FIR filter with an arbitrary
frequency response. H(f) is specified as a series of amplitude and phase points in the
frequency domain. The points are then converted into real and imaginary
components. Next, the impulse response is obtained by taking the complex inverse
FFT of the frequency response. The impulse response is then truncated to N points,
and a window function is applied to minimize the effects of truncation. The filter
design should then be tested by taking its FFT and evaluating the frequency
response. Several iterations may be required to achieve the desired response.
6.17
DIGITAL FILTERS
FREQUENCY SAMPLING METHOD FOR FIR FILTERS
WITH ARBITRARY FREQUENCY RESPONSE
■ Specify H(k) as a Finite Number of Spectral Points Spread
Uniformly Between 0 and 0.5fs (512 Usually Sufficient)
■ Specify Phase Points (Can Make Equal to Zero)
■ Convert Rectangular Form (Real + Imaginary)
■ Take the Complex Inverse FFT of H(k) Array to Obtain the Impulse
Response
■ Truncate the Impulse Response to N Points
■ Apply a suitable Window function to h(m) to smooth the effects of
truncation
■ Test Filter Design and Modify if Necessary
■ CAD Design Techniques More Suitable for Lowpass, Highpass,
Bandpass, or Bandstop Filters
Figure 6.21
FIR Filter Design Using the Parks-McClellan Program
Historically, the design method based on the use of windows to truncate the impulse
response and to obtain the desired frequency response was the first method used for
designing FIR filters. The frequency-sampling method was developed in the 1970s
and is still popular where the frequency response is an arbitrary function.
Modern CAD programs are available today which greatly simplify the design of
lowpass, highpass, bandpass, or bandstop FIR filters. A popular one was developed
by Parks and McClellan and uses the Remez exchange algorithm. The filter design
begins by specifying the parameters shown in Figure 6.22: passband ripple,
stopband ripple (same as attenuation), and the transition region. For this design
example, the QED1000 program from Momentum Data Systems was used (a demo
version is free and downloadable from http://www.mds.com).
For this example, we will design an audio lowpass filter that operates at a sampling
rate of 44.1kHz. The filter is specified as shown in Figure 6.22: 18kHz passband
frequency, 21kHz stopband frequency, 0.01dB passband ripple, 96dB stopband
ripple (attenuation). We must also specify the wordlength of the coefficients, which
in this case is 16 bits, assuming a 16-bit fixed-point DSP is to be used.
6.18
DIGITAL FILTERS
FIR CAD TECHNIQUES: PARKS McCLELLAN PROGRAM
WITH REMEZ EXCHANGE ALGORITHM
δ1 = Passband Ripple
1
GAIN
fc
Passband Ending
Frequency
fp
Stopband Beginning
Frequency
δ2 = Stopband Ripple
(Attenuation)
0
FREQUENCY
Ripple Ratio =
δ2
δ1
Figure 6.22
PARKS McCLELLAN EQUIRIPPLE FIR
FILTER DESIGN: PROGRAM INPUTS
■ Filter Type:
◆ Lowpass
◆ Highpass
◆ Bandpass
◆ Bandstop
◆ Differentiator
◆ Multiband
■ Sampling Frequency: 44,100Hz
■ Passband Frequency: 18,000Hz
■ Stopband Frequency: 21,000Hz
■ Passband Ripple:
■ Stopband Ripple (Attenuation):
0.01dB
96dB
■ Wordlength: 16-bits
Figure 6.23
6.19
DIGITAL FILTERS
The program allows us to choose between a window-based design or the equiripple
Parks-McClellan program. We will choose the latter. The program now estimates
the number of taps required to implement the filter based on the above
specifications. In this case, it is 69 taps. At this point, we can accept this and
proceed with the design or decrease the number of taps and see what degradation is
specifications occur.
We will accept this number and let the program complete the calculations. The
program outputs the frequency response (Figure 6.25), step function response
(Figure 6.26), s and z-plane analysis data, and the impulse response (Figure 6.27).
The QED1000 program then outputs the quantized filter coefficients to a program
which generates the actual DSP assembly code for a number of popular DSPs,
including Analog Devices’. The program is quite flexible and allows the user to
perform a number of scenarios to optimize the filter design.
FIR FILTER PROGRAM OUTPUTS
■ Estimated Number of Taps Required: 69
◆ Accept? Change?
Accept
■ Frequency Response (Linear and Log Scales)
■ Step Response
■ S - and Z - Plane Analysis
■ Impulse Response: Filter Coefficients (Quantized)
■ DSP FIR Filter Assembly Code
Figure 6.24
The 69-tap FIR filter requires 69 + 5 = 74 instruction cycles using the ADSP-2189M
75MIPS processor, which yields a total computation time per sample of 74×13.3ns =
984ns. The sampling interval is 1/44.1kHz, or 22.7µs. This allows 22.7µs – 0.984µs =
21.7µs for overhead and other operations.
Other options are to use a slower processor for this application (3.3MIPS), a more
complex filter which takes more computation time (up to N = 1700), or increase the
sampling frequency to about 1MSPS.
6.20
DIGITAL FILTERS
FIR DESIGN EXAMPLE: FREQUENCY RESPONSE
Passband Ripple < 0.01dB
18kHz
GAIN
96dB
fs = 44.1kSPS
FREQUENCY (Hz),
Courtesy Momentum Data Systems
21kHz
fs
2
Figure 6.25
FIR FILTER DESIGN EXAMPLE: STEP RESPONSE
AMPLITUDE
TIME (MILLISECONDS)
Courtesy Momentum Data Systems
Figure 6.26
6.21
DIGITAL FILTERS
FIR DESIGN EXAMPLE: IMPULSE RESPONSE
(FILTER COEFFICIENTS)
AMPLITUDE
N = 69
TIME (MILLISECONDS)
Courtesy Momentum Data Systems
Figure 6.27
DESIGN EXAMPLE USING ADSP-2189M:
PROCESSOR TIME FOR 69-TAP FIR FILTER
■ Sampling Frequency fs = 44.1kSPS
■ Sampling Interval = 1 / fs = 22.7µs
■ Number of Filter Taps, N = 69
■ Number of Required Instructions = N + 5 = 74
■ Processing Time / Instruction = 13.3ns (75MIPS)
(ADSP-2189M)
■ Total Processing Time = 74×13.3ns = 984ns
■ Total Processing Time < Sampling Interval with
22.7µs – 0.984µs = 21.7µs for Other Operations
◆ Increase Sampling Frequency to 1MHz
◆ Use Slower DSP (3.3MIPS)
◆ Add More Filter Taps (Up to N = 1700)
Figure 6.28
6.22
DIGITAL FILTERS
Designing Highpass, Bandpass, and Bandstop Filters Based on Lowpass
Filter Design
Converting a lowpass filter design impulse response into a highpass filter impulse
response can be accomplished in one of two ways. In the spectral inversion method,
the sign of each filter coefficient in the lowpass filter impulse response is changed.
Next, 1 is added to the center coefficient. In the spectral reversal method, the sign of
every other coefficient is changed. This reverses the frequency domain plot. In other
words, if the cutoff of the lowpass filter is 0.2fs, the resulting highpass filter will
have a cutoff frequency of 0.5fs – 0.2fs = 0.3fs. This must be considered when doing
the original lowpass filter design.
DESIGNING HIGHPASS FILTERS
USING LOWPASS FILTER IMPULSE RESPONSE
■ Spectral Inversion Technique:
◆ Design Lowpass Filter (Linear Phase, N odd)
◆ Change the Sign of Each Coefficient in the Impulse
Response, h(m)
◆ Add 1 to the Coefficient at the Center of Symmetry
■ Spectral Reversal Technique:
◆ Design Lowpass Filter
◆ Change the Sign of Every Other Coefficient in the
Impulse Response, h(m)
◆ This Reverses the Frequency Domain left-for-right:
0 becomes 0.5, and 0.5 becomes 0;
i.e., if the cutoff frequency of the lowpass filter is 0.2,
the cutoff of the resulting highpass filter is 0.3
Figure 6.29
Bandpass and bandstop filters can be designed by combining individual lowpass and
highpass filters in the proper manner. Bandpass filters are designed by placing the
lowpass and highpass filters in cascade. The equivalent impulse response of the
cascaded filters is then obtained by convolving the two individual impulse
responses.
A bandstop filter is designed by connecting the lowpass and highpass filters in
parallel and adding their outputs. The equivalent impulse response is then obtained
by adding the two individual impulse responses.
6.23
DIGITAL FILTERS
BANDPASS AND BANDSTOP FILTERS
DESIGNED FROM LOWPASS AND HIGHPASS FILTERS
x(n)
h1(k)
h2(k)
y(n)
x(n)
h1(k) ✳ h2(k)
y(n)
✳ = Convolution
Lowpass
Highpass
Bandpass
Highpass
x(n)
h1(k)
Σ
x(n)
y(n)
h1(k) + h2(k)
y(n)
h2(k)
Lowpass
Bandstop
Figure 6.30
INFINITE IMPULSE RESPONSE (IIR) FILTERS
As was mentioned previously, FIR filters have no real analog counterparts, the
closest analogy being the weighted moving average. In addition, FIR filters have
only zeros and no poles. On the other hand, IIR filters have traditional analog
counterparts (Butterworth, Chebyshev, Elliptic, and Bessel) and can be analyzed
and synthesized using more familiar traditional filter design techniques.
Infinite impulse response filters get their name because their impulse response
extends for an infinite period of time. This is because they are recursive, i.e., they
utilize feedback. Although they can be implemented with fewer computations than
FIR filters, IIR filters do not match the performance achievable with FIR filters, and
do not have linear phase. Also, there is no computational advantage achieved when
the output of an IIR filter is decimated because each output value must always be
calculated.
6.24
DIGITAL FILTERS
INFINITE IMPULSE RESPONSE (IIR) FILTERS
■ Uses Feedback (Recursion)
■ Impulse Response has an Infinite Duration
■ Potentially Unstable
■ Non-Linear Phase
■ More Efficient than FIR Filters
■ No Computational Advantage when Decimating
Output
■ Usually Designed to Duplicate Analog Filter Response
■ Usually Implemented as Cascaded Second-Order
Sections (Biquads)
Figure 6.31
IIR filters are generally implemented in two-pole sections called biquads because
they are described with a biquadratic equation in the z-domain. Higher order filters
are designed using cascaded biquad sections, i.e., a 6-pole filter requires 3 biquad
sections.
The basic IIR biquad is shown in Figure 6.32. The zeros are formed by the
feedforward coefficients b0, b1, and b2; the poles are formed by the feedback
coefficients a1, and a2.
HARDWARE IMPLEMENTATION OF
SECOND-ORDER IIR FILTER (BIQUAD) DIRECT FORM 1
b0
x(n)
Σ
y(n)
z–1
z–1
b1
–a1
z–1
z–1
b2
–a2
y(n) = b0x(n) + b1x(n–1) + b2x(n–2) – a1y(n–1) – a2y(n–2)
M
M
Σ
N
y(n) = Σ bkx(n–k) – Σ aky(n–k)
k=0
k=1
k=0
H(z) =
bkz–k
(Zeros)
N
1
+
Σ
akz–k
k=1
(Poles)
Figure 6.32
6.25
DIGITAL FILTERS
The general digital filter equation is shown in Figure 6.32 which gives rise to the
general transfer function H(z) which contains polynomials in both the numerator
and the denominator. The roots of the denominator determine the pole locations of
the filter, and the roots of the numerator determine the zero locations. Although it
is possible to construct a high order IIR filter directly from this equation (called the
direct form implementation), accumulation errors due to quantization errors (finite
wordlength arithmetic) may give rise to instability and large errors. For this reason,
it is common to cascade several biquad sections with appropriate coefficients rather
than use the direct form implementation. The biquads can be scaled separately and
then cascaded in order to minimize the coefficient quantization and the recursive
accumulation errors. Cascaded biquads execute more slowly than their direct form
counterparts, but are more stable and minimize the effects of errors due to finite
arithmetic errors.
The Direct Form 1 biquad section shown in Figure 6.32 requires four registers. This
configuration can be changed into an equivalent circuit shown in Figure 6.33 which
is called the Direct Form 2 which requires only two registers. It can be shown that
the equations describing the Direct Form 2 IIR biquad filter are the same as those
for Direct Form 1. As in the case of FIR filters, the notation for an IIR filter is often
simplified as shown in Figure 6.34.
IIR BIQUAD FILTER DIRECT FORM 2
x(n)
b0
Σ
Σ
z–1
–a1
b1
z–1
–a2
b2
Reduces to the same equation as Direct Form 1:
y(n) = b0x(n) + b1x(n–1) + b2x(n–2) – a1y(n–1) – a2y(n–2)
Requires Only 2 Delay Elements (Registers)
Figure 6.33
6.26
y(n)
DIGITAL FILTERS
IIR BIQUAD FILTER SIMPLIFIED NOTATIONS
b0
x(n)
y(n)
z–1
–a1
b1
z–1
–a2
b2
b0
x(n)
–a1
z–1
y(n)
b1
z–1
–a2
b2
Figure 6.34
IIR FILTER DESIGN TECHNIQUES
A popular method for IIR filter design is to first design the analog-equivalent filter
and then mathematically transform the transfer function H(s) into the z-domain,
H(z). Multiple pole designs are implemented using cascaded biquad sections. The
most popular analog filters are the Butterworth, Chebyshev, Elliptical, and Bessel
(see Figure 6.35). There are many CAD programs available to generate the Laplace
transform, H(s), for these filters.
The all-pole Butterworth (also called maximally flat) has no ripple in the passband
or stopband and has monotonic response in both regions. The all-pole Type 1
Chebyshev filter has a faster rolloff than the Butterworth (for the same number of
poles) and has ripple in the passband. The Type 2 Chebyschev filter is rarely used,
but has ripple in the stopband rather than the passband.
The Elliptical (Cauer) filter has poles and zeros and ripple in both the passband and
stopband. This filter has even faster rolloff than the Chebyshev for the same
number of poles. The Elliptical filter is often used where degraded phase response
can be tolerated.
Finally, the Bessel (Thompson) filter is an all-pole filter which is optimized for pulse
response and linear phase but has the poorest rolloff of any of the types discussed
for the same number of poles.
6.27
DIGITAL FILTERS
REVIEW OF POPULAR ANALOG FILTERS
■ Butterworth
◆ All Pole, No Ripples in Passband or Stopband
◆ Maximally Flat Response (Fastest Roll-off with No Ripple)
■ Chebyshev (Type 1)
◆ All Pole, Ripple in Passband, No Ripple in Stopband
◆ Shorter Transition Region than Butterworth for Given Number
of Poles
◆ Type 2 has Ripple in Stopband, No Ripple in Passband
■ Elliptical (Cauer)
◆ Has Poles and Zeros, Ripple in Both Passband and Stopband
◆ Shorter Transition Region than Chebyshev for Given Number
of Poles
◆ Degraded Phase Response
■ Bessel (Thompson)
◆ All Pole, No Ripples in Passband or Stopband
◆ Optimized for Linear Phase and Pulse Response
◆ Longest Transition Region of All for Given Number of Poles
Figure 6.35
All of the above types of analog filters are covered in the literature, and their
Laplace transforms, H(s), are readily available – either from tables or CAD
programs. There are three methods used to convert the Laplace transform into the
z-transform: impulse invariant transformation, bilinear transformation, and the
matched z-transform. The resulting z-transforms can be converted into the
coefficients of the IIR biquad. These techniques are highly mathematically intensive
and will not be discussed further.
A CAD approach for IIR filter design is similar to the Parks-McClellan program
used for FIR filters. This technique uses the Fletcher-Powell algorithm.
In calculating the throughput time of a particular DSP IIR filter, one should
examine the benchmark performance specification for a biquad filter section. For
the ADSP-21xx-family, seven instruction cycles are required to execute a biquad
filter output sample. For the ADSP-2189M, 75MIPS DSP, this corresponds to 7 ×
13.3ns = 93ns, corresponding to a maximum possible sampling frequency of
10MSPS (neglecting overhead).
6.28
DIGITAL FILTERS
IIR FILTER DESIGN TECHNIQUES
■ Impulse Invarient Transformation Method
◆ Start with H(s) for Analog Filter
◆ Take Inverse Laplace Transform to get Impulse Response
◆ Obtain z-Transform H(z) from Sampled Impulse Response
◆ z-Transform Yields Filter Coefficients
◆ Aliasing Effects Must be Considered
■ Bilinear Transformation Method
◆ Another Method for Transforming H(s) into H(z)
◆ Performance Determined by the Analog System’s
Differential Equation
◆ Aliasing Effects do not Occur
■ Matched z-Transform Method
◆ Maps H(s) into H(z) for filters with both poles and zeros
■ CAD Methods
◆ Fletcher-Powell Algorithm
◆ Implements Cascaded Biquad Sections
Figure 6.36
THROUGHPUT CONSIDERATIONS FOR IIR FILTERS
■ Determine How Many Biquad Sections (N) are Required to
Realize the Desired Frequency Response
■ Multiply this by the number of instruction cycles per Biquad
for the DSP and add overhead cycles (5N + 2 cycles for the
ADSP-21xx series, for example).
■ The Result (plus overhead) is the Minimum Allowable
Sampling Period (1 / fs) for Real-Time Operation
Figure 6.37
6.29
DIGITAL FILTERS
Summary: FIR Versus IIR Filters
Choosing between FIR and IIR filter designs can be somewhat of a challenge, but a
few basic guidelines can be given. Typically, IIR filters are more efficient than FIR
filters because they require less memory and fewer multiply-accumulates are
needed. IIR filters can be designed based upon previous experience with analog
filter designs. IIR filters may exhibit instability problems, but this is much less
likely to occur if higher order filters are designed by cascading second-order
systems.
On the other hand, FIR filters require more taps and multiply-accumulates for a
given cutoff frequency response, but have linear phase characteristics. Since FIR
filters operate on a finite history of data, if some data is corrupted (ADC sparkle
codes, for example) the FIR filter will ring for only N–1 samples. Because of the
feedback, however, an IIR filter will ring for a considerably longer period of time.
If sharp cutoff filters are needed and processing time is at a premium, IIR elliptic
filters are a good choice. If the number of multiply/accumulates is not prohibitive,
and linear phase is a requirement, then the FIR should be chosen.
COMPARISON BETWEEN FIR AND IIR FILTERS
IIR FILTERS
FIR FILTERS
More Efficient
Less Efficient
Analog Equivalent
No Analog Equivalent
May Be Unstable
Always Stable
Non-Linear Phase Response
Linear Phase Response
More Ringing on Glitches
Less Ringing on Glitches
CAD Design Packages Available
CAD Design Packages Available
No Efficiency Gained by Decimation
Decimation Increases Efficiency
Figure 6.38
6.30
DIGITAL FILTERS
MULTIRATE FILTERS
There are many applications where it is desirable to change the effective sampling
rate in a sampled data system. In many cases, this can be accomplished simply by
changing the sampling frequency to the ADC or DAC. However, it is often desirable
to accomplish the sample rate conversion after the signal has been digitized. The
most common techniques used are decimation (reducing the sampling rate by a
factor of M), and interpolation (increasing the sampling rate by a factor of L). The
decimation and interpolation factors (M and L) are normally integer numbers. In a
generalized sample-rate converter, it may be desirable to change the sampling
frequency by a non-integer number. In the case of converting the CD sampling
frequency of 44.1kHz to the digital audio tape (DAT) sampling rate of 48kHz,
interpolating by L =160 followed by decimation by M = 147 accomplishes the desired
result.
The concept of decimation is illustrated in Figure 6.39. The top diagram shows the
original signal, fa, which is sampled at a frequency fs. The corresponding frequency
spectrum shows that the sampling frequency is much higher than required to
preserve information contained in fa, i.e., fa is oversampled. Notice that there is no
information contained between the frequencies fa and fs – fa. The bottom diagram
shows the same signal where the sampling frequency has been reduced (decimated)
by a factor of M. Notice that even though the sampling rate has been reduced, there
is no aliasing and loss of information. Decimation by a larger factor than shown in
Figure 6.39 will cause aliasing.
DECIMATION OF A
SAMPLED SIGNAL BY A FACTOR OF M
A: ORIGINAL OVERSAMPLED SIGNAL
1
fs
t
fs
2
fa
fs–fa fs
f
B: SIGNAL DECIMATED BY M
M
fs
t
fs
2M
fs
M
f
Figure 6.39
6.31
DIGITAL FILTERS
Figure 6.40A shows how to decimate the output of an FIR filter. The filtered data
y(n) is stored in a data register which is clocked at the decimated frequency fs/M.
This does not change the number of computations required of the digital filter, i.e.,
it still must calculate each output sample y(n).
DECIMATION COMBINED WITH FIR FILTERING
A: No Change in Computational Efficiency
x(n)
z–1
h(0)
z–1
h(1)
z–1
h(2)
Σ
Clock =
h(N–1)
Σ
y(n)
Σ
fs
M
Data
Register
y(m) ↓ M
B: Computational Efficiency Increased by Factor of M
x(n)
z–1
z–1
z–1
Clock =
N Data Registers
h(0)
h(1)
h(2)
Σ
h(N–1)
Σ
fs
M
y(m) ↓ M
Σ
Figure 6.40
Figure 6.40B shows a method for increasing the computational efficiency of the FIR
filter by a factor of M. The data from the delay registers are simply stored in N data
registers which are clocked at the decimated frequency fs/M. The FIR
multiply/accumulates now only have to be done every Mth clock cycle. This increase
in efficiency could be utilized by adding more taps to the FIR filter, doing other
computations in the extra time, or using a slower DSP.
Figure 6.41 shows the concept of interpolation. The original signal in 6.41A is
sampled at a frequency fs. In 6.41B, the sampling frequency has been increased by a
factor of L, and zeros have been added to fill in the extra samples. The signal with
added zeros is passed through an interpolation filter which provides the extra data
values.
6.32
DIGITAL FILTERS
INTERPOLATION BY A FACTOR OF L
A:
ORIGINAL
SAMPLED SIGNAL
SAMPLING RATE = fs
1
fs
B:
t
SIGNAL WITH
ADDED ZEROS
t
1
Lfs
C:
SIGNAL AFTER
INTERPOLATING
FILTER
SAMPLING RATE = Lfs
1
Lfs
t
Figure 6.41
The frequency domain effects of interpolation are shown in Figure 6.42. The original
signal is sampled at a frequency fs and is shown in 6.42A. The interpolated signal in
6.42B is sampled at a frequency Lfs. An example of interpolation is a CD player
DAC, where the CD data is generated at a frequency of 44.1kHz. If this data is
passed directly to a DAC, the frequency spectrum shown in Figure 6.42A results,
and the requirements on the anti-imaging filter which precedes the DAC are
extremely stringent to overcome this. An oversampling interpolating DAC is
normally used, and the spectrum shown in Figure 6.42B results. Notice that the
requirements on the analog anti-imaging filter are now easier to realize. This is
important in maintaining relatively linear phase and also reducing the cost of the
filter.
The digital implementation of interpolation is shown in Figure 6.43. The original
signal x(n) is first passed through a rate expander which increases the sampling
frequency by a factor of L and inserts the extra zeros. The data then passes through
an interpolation filter which smoothes the data and interpolates between the
original data points. The efficiency of this filter can be improved by using a filter
algorithm which takes advantage of the fact that the zero-value input samples do
not require multiply-accumulates. Using a DSP which allows circular buffering and
zero-overhead looping also improves efficiency.
6.33
DIGITAL FILTERS
EFFECTS OF INTERPOLATION
ON FREQUENCY SPECTRUM
ORIGINAL SIGNAL
1
fs
fs
2
t
fs
3fs
2fs
4fs
f
Lfs
f
SIGNAL INTERPOLATED BY L
1
Lfs
Lfs
2
t
Figure 6.42
TYPICAL INTERPOLATION IMPLEMENTATION
Interpolation Filter
Clock = Lfs
x(n)
z–1
↑L
h(0)
z–1
h(1)
z–1
h(2)
Σ
h(N–1)
Σ
Rate Expander
Increases Sample Rate
and Inserts Zeros
Efficient DSP algorithms take advantage of:
Multiplications by zero
Circular Buffers
Zero-Overhead Looping
Figure 6.43
6.34
Σ
y(m) ↑ L
DIGITAL FILTERS
Interpolators and decimators can be combined to perform fractional sample rate
conversion as shown in Figure 6.44. The input signal x(n) is first interpolated by a
factor of L and then decimated by a factor of M. The resulting output sample rate is
Lfs/M. To maintain the maximum possible bandwidth in the intermediate signal,
the interpolation must come before the decimation; otherwise, some of the desired
frequency content in the original signal would be filtered out by the decimator.
An example is converting from the CD sampling rate of 44.1kHz to the digital audio
tape (DAT) sampling rate of 48.0kHz. The interpolation factor is 160, and the
decimation factor of 147. In practice, the interpolating filter h΄(k) and the
decimating filter h"(k) are combined into a single filter, h(k).
The entire sample-rate conversion function is integrated into the AD1890, AD1891,
AD1892, and AD1893-family which operates at frequencies between 8kHz and
56kHz (48kHz for the AD1892). The new AD1896 operates up to 196kHz.
SAMPLE RATE CONVERTERS
INTERPOLATOR
x(n)
↑L
DECIMATOR
h'(k)
h"(k)
Output Sample Rate =
fs
x(n)
↓M
↑L
h(k)
↓M
y(m)
L f
s
M
y(m)
Example: Convert CD Sampling Rate = 44.1kHz to
DAT Sampling Rate = 48.0kHz
Use L = 160, M = 147
fout =
L f
160
× 44.1kHz = 48.0kHz
=
M s
147
Figure 6.44
ADAPTIVE FILTERS
Unlike analog filters, the characteristics of digital filters can easily be changed
simply by modifying the filter coefficients. This makes digital filters attractive in
communications applications such as adaptive equalization, echo cancellation, noise
reduction, speech analysis and synthesis, etc. The basic concept of an adaptive filter
is shown in Figure 6.45. The objective is to filter the input signal, x(n), with an
adaptive filter in such a manner that it matches the desired signal, d(n). The
6.35
DIGITAL FILTERS
desired signal, d(n), is subtracted from the filtered signal, y(n), to generate an error
signal. The error signal drives an adaptive algorithm which generates the filter
coefficients in a manner which minimizes the error signal. The least-mean-square
(LMS) or recursive-least-squares (RLS) algorithms are two of the most popular.
ADAPTIVE FILTER
d(n)
Desired Signal
x(n)
Input Signal
ADAPTIVE
N-TAP
FIR FlLTER
N
–
y(n)
h(k),
Filter
Coefficients
+
∑
Error
Signal
ADAPTIVE
ALGORITHM
(LMS, RLS)
Figure 6.45
Adaptive filters are widely used in communications to perform such functions as
equalization, echo cancellation, noise cancellation, and speech compression. Figure
6.46 shows an application of an adaptive filter used to compensate for the effects of
amplitude and phase distortion in the transmission channel. The filter coefficients
are determined during a training sequence where a known data pattern is
transmitted. The adaptive algorithm adjusts the filter coefficients to force the
receive data to match the training sequence data. In a modem application, the
training sequence occurs after the initial connection is made. After the training
sequence is completed, the switches are put in the other position, and the actual
data is transmitted. During this time, the error signal is generated by subtracting
the input from the output of the adaptive filter.
Speech compression and synthesis also makes extensive use of adaptive filtering to
reduce data rates. The linear predictive coding (LPC) model shown in Figure 6.47
models the vocal tract as a variable frequency impulse generator (for voiced portions
of speech) and a random noise generator (for unvoiced portions of speech such as
consonant sounds). These these two generators drive a digital filter which in turn
generates the actual voice signal.
6.36
DIGITAL FILTERS
DIGITAL TRANSMISSION
USING ADAPTIVE EQUALIZATION
ENCODER,
MODULATOR,
FILTER
TRANSMIT
DATA
DAC
ANALOG
FILTER
TRANSMISSION
CHANNEL
TRAINING
SEQUENCE
RECEIVE
DATA
ADAPTIVE
FILTER
DECODER,
DEMODULATOR,
FILTER
Adds Noise,
Amplitude and
Phase Distortion
ADC
ANALOG
FILTER
N
h(k)
ADAPTIVE
ALGORITHM
Error
+
∑
TRAINING
SEQUENCE
–
Figure 6.46
LINEAR PREDICTIVE CODING (LPC)
MODEL OF SPEECH PRODUCTION
NOSE
MOUTH
(Tongue & Lips)
PHARYNX
VELUM
LARYNX
(Vocal Cords)
LUNGS
PITCH
PERIOD
IMPULSE
TRAIN
GENERATOR
RANDOM
NOISE
GENERATOR
Voiced
TIME-VARYING
DIGITAL
FILTER
Unvoiced
GAIN
Figure 6.47
6.37
DIGITAL FILTERS
The application of LPC in a communication system such as GSM is shown in Figure
6.48. The speech input is first digitized by a 16-bit ADC at a sampling frequency of
8kSPS. This produces output data at 128kBPS which is much too high to be
transmitted directly. The transmitting DSP uses the LPC algorithm to break the
speech signal into digital filter coefficients and pitch. This is done in 20ms windows
which has been found to be optimum for most speech applications. The actual
transmitted data is only 2.4kBPS which represents a 53.3 compression factor. The
receiving DSP uses the LPC model to reconstruct the speech from the coefficients
and the excitation data. The final output data rate of 128kBPS then drives a 16-bit
DAC for final reconstruction of the speech data.
LPC SPEECH COMPANDING SYSTEM
fs = 8kSPS
Speech
Input
16-Bit
ADC
TRANSMITTING DSP
WINDOWING
(20ms)
PRE-EMPHASIS
COEFFICIENT
AND PITCH
128kBPS
2.4kBPS
fs = 8kSPS
Speech
Output
RECEIVING DSP
P
16-Bit
DAC
128kBPS
s(n)
=
Σ
k=1
Speech
Output
ak s(n – k) + G • x(n)
All-Pole
Filter
Gain Excitation
Data Rate Reduction Factor = 53.3
Figure 6.48
The digital filters used in LPC speech applications can either be FIR or IIR,
although all-pole IIR filters are the most widely used. Both FIR and IIR filters can
be implemented in a lattice structure as shown in Figure 6.49 for a recursive allpole filter. This structure can be derived from the IIR structure, but the advantage
of the lattice filter is that the coefficients are more directly related to the outputs of
algorithms which use the vocal tract model shown in Figure 6.47 than the
coefficients of the equivalent IIR filter.
The parameters of the all-pole lattice filter model are determined from the speech
samples by means of linear prediction as shown in Figure 6.50. Due to the nonstationary nature of speech signals, this model is applied to short segments
(typically 20ms) of the speech signal. A new set of parameters is usually determined
for each time segment unless there are sharp discontinuities, in which case the data
may be smoothed between segments.
6.38
DIGITAL FILTERS
ALL POLE LATTICE FILTER
EXCITATION
INPUT
SPEECH
OUTPUT
y(n)
XM(n)
–kM
kM
–k2
k2
z–1
–k1
k1
z–1
uM(n)
u2(n)
z–1
u1(n)
Figure 6.49
ESTIMATION OF LATTICE FILTER
COEFFICIENTS IN TRANSMITTING DSP
SPEECH
SAMPLES
+
Σ
ERROR
SIGNAL
–
Z–1
ADAPTIVE
FIR
PREDICTOR
COEFFICIENTS
ADAPTIVE
ALGORITHM
Figure 6.50
6.39
DIGITAL FILTERS
REFERENCES
1.
Steven W. Smith, The Scientist and Engineer’s Guide to Digital Signal
Processing, Second Edition, 1999, California Technical Publishing,
P.O. Box 502407, San Diego, CA 92150. Also available for free download at:
http://www.dspguide.com or http://www.analog.com
2.
C. Britton Rorabaugh, DSP Primer, McGraw-Hill, 1999.
3.
Richard J. Higgins, Digital Signal Processing in VLSI, Prentice-Hall,
1990.
4.
A. V. Oppenheim and R. W. Schafer, Digital Signal Processing, PrenticeHall, 1975.
5.
L. R. Rabiner and B. Gold, Theory and Application of Digital Signal
Processing, Prentice-Hall, 1975.
6.
John G. Proakis and Dimitris G. Manolakis, Introduction to Digital
Signal Processing, MacMillian, 1988.
7.
J.H. McClellan, T.W. Parks, and L.R. Rabiner, A Computer Program
for Designing Optimum FIR Linear Phase Digital Filters, IEEE
Trasactions on Audio and Electroacoustics, Vol. AU-21, No. 6,
December, 1973.
8.
Fredrick J. Harris, On the Use of Windows for Harmonic Analysis with
the Discrete Fourier Transform, Proc. IEEE, Vol. 66, No. 1, 1978 pp. 51-83.
9.
Momentum Data Systems, Inc., 17330 Brookhurst St., Suite 140,
Fountain Valley, CA 92708, http://www.mds.com
10.
Digital Signal Processing Applications Using the ADSP-2100
Family, Vol. 1 and Vol. 2, Analog Devices, Free Download at:
http://www.analog.com
11.
ADSP-21000 Family Application Handbook, Vol. 1, Analog Devices,
Free Download at: http://www.analog.com
12.
B. Widrow and S.D. Stearns, Adaptive Signal Processing, Prentice-Hall,
1985.
13.
S. Haykin, Adaptive Filter Theory, 3rd Edition, Prentice-Hall, 1996.
14.
Michael L. Honig and David G. Messerschmitt, Adaptive Filters Structures, Algorithms, and Applications, Kluwer Academic Publishers,
Hingham, MA 1984.
15.
J.D. Markel and A.H. Gray, Jr., Linear Prediction of Speech, SpringerVerlag, New York, NY, 1976.
6.40
DIGITAL FILTERS
16.
L.R. Rabiner and R.W. Schafer, Digital Processing of Speech
Signals, Prentice-Hall, 1978.
6.41
DIGITAL FILTERS
6.42
DSP HARDWARE
SECTION 7
DSP HARDWARE
■
Microcontrollers, Microprocessors, and Digital Signal
Processors (DSPs)
■
DSP Requirements
■
ADSP-21xx 16-Bit Fixed-Point DSP Core
■
Fixed-Point Versus Floating Point
■
ADI SHARC® Floating Point DSPs
■
ADSP-2116x Single-Instruction, Multiple Data (SIMD)
Core Architecture
■
TigerSHARC™: The ADSP-TS001
Static Superscalar DSP
■
DSP Benchmarks
■
DSP Evaluation and Development Tools
7.a
DSP HARDWARE
7.b
DSP HARDWARE
SECTION 7
DSP HARDWARE
Dan King, Greg Geerling, Ken Waurin, Noam
Levine, Jesse Morris, Walt Kester
MICROCONTROLLERS, MICROPROCESSORS, AND
DIGITAL SIGNAL PROCESSORS (DSPS)
Computers are extremely capable in two broad areas: (1) data manipulation, such
as word processing and database management, and (2) mathematical calculation,
used in science, engineering, and Digital Signal Processing. However, most
computers are not optimized to perform both functions. In computing applications
such as word processing, data must be stored, sorted, compared, moved, etc., and
the time to execute a particular instruction is not critical, as long as the program’s
overall response time to various commands and operations is adequate enough to
satisfy the end user. Occasionally, mathematical operations may also be performed,
as in a spreadsheet or database program, but speed of execution is generally not the
governing factor. In most general purpose computing applications there is no
concentrated attempt by software companies to make the code efficient. Application
programs are loaded with “features” which require more memory and faster
processors with every new release or upgrade.
GENERAL COMPUTING APPLICATIONS
DATA MANIPULATION
■
■
■
■
MATH CALCULATION
■
■
■
■
Word Processing
Database Management
Spread Sheets
Operating Systems
Digital Signal Processing
Motion Control
Engineering Simulations
Real-Time Signal Processing
■ Data Movement (A→
→ B)
■ Value Testing (If A = B, then…)
■ Addition (C = A + B)
■ Multiplication (C = A × B)
■ Time to Execute not Critical,
not Predictable
■ Time to Execute Critical
and Predictable
Figure 7.1
On the other hand, digital signal processing applications require that mathematical
operations be performed quickly, and the time to execute a given instruction must
be known precisely, and it must be predictable. Both code and hardware must be
extremely efficient to accomplish this. As has been shown in the last two sections of
this book, the most fundamental mathematical operation or kernel in all of DSP is
7.1
DSP HARDWARE
the sum-of-products (or dot-product). Fast execution of the dot product is critical to
fast Fourier transforms (FFTs), real time digital filters, matrix multiplications,
graphics pixel manipulation, etc.
Based on this introductory discussion of DSP requirements, it is important to
understand the differences between microcontrollers, microprocessors, and DSPs.
While microcontrollers used in industrial process control applications can perform
functions such as multiplication, addition, and division, they are much more suited
to applications where I/O capability and control is more important than speed.
Microcontrollers such as the 8051-family typically contain a CPU, RAM, ROM,
serial/parallel interfaces, timers, and interrupt circuitry. The MicroConverter™
series from Analog Devices contains not only the 8051 core but also high
performance ADC and DAC functions along with flash memory.
MICROCONTROLLERS, MICROPROCESSORS,
AND DIGITAL SIGNAL PROCESSORS (DSPs)
■ Microcontrollers:
◆ CPU, RAM, ROM, Serial/Parallel Interface, Timer, Interrupt Circuitry
◆ Well Suited for Toasters as well Industrial Process Control
◆ Speed is not Generally a Requirement!
◆ Compact Instruction Sets
◆ Example: 8051, 68HC11, PIC
■ Microprocessors:
◆ Single Chip CPU - Requires Additional External Circuitry
◆ RISC: Reduced Instruction Set Computer
◆ CISC: Complex Instruction Set Computer
◆ Example: Pentium-Series, PowerPC, MIPS
■ Digital Signal Processors (DSPs):
◆ RAM, ROM, Serial/Parallel Interface, Interrupt Circuitry
◆ CPU Optimized for Fast Repetitive Math for Real Time Processing
◆ Example: ADSP-21XX, ADSP-21K
Figure 7.2
Microprocessors, such as the Pentium-series from Intel, are basically single-chip
CPUs which require additional circuitry to make up the total computing function.
Microprocessor instruction sets can be either complex-instruction-set computer
(CISC) or reduced-instruction-set computer (RISC). The complex-instruction-set
computer (CISC) includes instructions for basic processor operations, plus single
instructions that are highly sophisticated; for example, to evaluate a high-order
polynomial. But CISC has a price: many of the instructions execute via microcode
in the CPU and require numerous clock cycles plus silicon real estate for code
storage memory.
7.2
DSP HARDWARE
In contrast, the reduced-instruction-set computer (RISC) recognizes that, in many
applications, basic instructions such as LOAD and STORE - with simple addressing
schemes - are used much more frequently than the advanced instructions, and
should not incur an execution penalty. These simpler instructions are hardwired in
the CPU logic to execute in a single clock cycle, reducing execution time and CPU
complexity.
Although the RISC approach offers many advantages in general purpose computing,
it is not well suited to DSP. For example, most RISCs do not support singleinstruction multiplication, a very common and repetitive operation in DSP. The
DSP is optimized to accomplish these tasks fast enough to maintain real-time
operation in the context of the application. This requires single-cycle arithmetic
operations and accumulations.
DSP REQUIREMENTS
The most fundamental mathematical operation in DSP is shown in Figure 7.3: the
sum of products (dot product). It is common to digital filters, FFTs, and many other
DSP applications. A Digital Signal Processor (DSP) is optimized to perform
repetitive mathematical operations such as the dot product. There are five basic
requirements of a DSP to optimize this performance: fast arithmetic, extended
precision, dual operand fetch, circular buffering, and zero-overhead looping.
THE MOST FUNDAMENTAL MATHEMATICAL
OPERATION IN DSP: THE SUM OF PRODUCTS
y(n) = h(0) • x(n) + h(1) • x(n–1) + ... + h(N–1) • x(n–N)
■ For Example: Digital Filtering
◆ Multiply Data Sample Times Filter Coefficient (Twiddle
Factor for FFTs)
◆ Add Contents to Accumulator
◆ Repeat N times
■ DSP Requirements:
◆ Fast Multiply-Accumulates
◆ Extended Precision (Accumulator Register)
◆ Dual Operand Fetch
◆ Circular Buffering
◆ Zero-Overhead Looping
■ In One Instruction Cycle Using ADSP-21XX Core :
◆ Fetch Data Sample from Data Memory
◆ Fetch Coefficient from Program Memory
◆ Perform Multiply-Accumulate
◆ Update Pointers
Figure 7.3
7.3
DSP HARDWARE
FAST ARITHMETIC
Fast arithmetic is the simplest of these requirements to understand. Since real-time
DSP applications are driven by performance, the multiply-accumulate or MAC time
is a basic requirement; faster MACs mean potentially higher bandwidth. It is
critical to remember that MAC time alone does not define DSP performance. This
often forgotten fact leads to an inadequate measure of processor performance by
simply examining its MIPS (million instructions per second) rating. Since most DSP
and DSP-like architectures feature MACs that can execute an instruction every
cycle, most processors are given a MIPS rating equal to its MAC throughput. This
does not necessarily account for the other factors that can degrade a processors
overall performance in real-world applications. The other four criteria can wipe out
MAC gains if they are not satisfied.
In addition to the requirement for fast arithmetic, a DSP should be able to support
other general purpose math functions and should therefore have an appropriate
arithmetic logic unit (ALU) and a programmable shifter function for bit
manipulation.
EXTENDED PRECISION
Apart from the obvious need for fast multiplication and addition (MAC), there is
also a requirement for extended precision in the accumulator register. For example,
when two 16-bit words are multiplied, the result is a 32-bit word. The Analog
Devices ADSP-21xx 16-bit fixed-point core architecture has an internal 40-bit
accumulator which provides a high degree of overflow protection. While floatingpoint DSPs eliminate most of the problems associated with precision and overflow,
fixed-point processors are still popular for many applications, and therefore
overflow, underflow, and data scaling issues must be dealt with properly.
DUAL OPERAND FETCH
Regardless of the nature of a processor, performance limitations are generally based
on bus bandwidth. In the case of general purpose microprocessors or
microcontrollers, code is dominated by single memory fetch instructions, usually
addressed as a base plus offset value. This leads architects to embed fixed data into
the instruction set so that this class of memory access is fast and memory efficient.
DSPs, on the other hand, are dominated by instructions requiring two independent
memory fetches. This is driven by the basic form of the convolution (kernel or dot
product) Σh(i)x(i). The goal of fast dual operand fetches is to keep the MAC fully
loaded. We saw in the discussion on MACs that the performance of a DSP is first
limited by MAC time. Assuming an adequate MAC cycle time, two data values need
to be supplied at the same rate; increases in operand fetch cycle time will result in
corresponding increases in MAC cycle time. Ideally, the operand fetches occur
simultaneously with the MAC instruction so that the combination of the MAC and
memory addressing occurs in one cycle.
Dual operand fetch is implemented in DSPs by providing separate buses for
program memory data and data memory data. In addition, separate program
memory address and data memory address buses are also provided. The MAC can
7.4
DSP HARDWARE
therefore receive inputs from each data bus simultaneously. This architecture is
often referred to as the Harvard Architecture.
CIRCULAR BUFFERING
If we examine the kernel equation more carefully, the advantages of circular
buffering in DSP applications become apparent. A Finite Impulse Response (FIR)
filter is used to demonstrate the point. First, coefficients or tap values for FIR filters
are periodic in nature. Second, the FIR filter uses the newest real-world signal
value and discards the oldest value when calculating each output.
In the series of FIR filter equations, the N coefficient locations are always accessed
sequentially from h(0) to h(N–1). The associated data points circulate through the
memory as follows: new samples are stored replacing the oldest data each time a
filter output is computed. A fixed boundary RAM can be used to achieve this
circulating buffer effect. The oldest data sample is replaced by the newest with each
convolution. A "time history" of the N most recent samples is kept in RAM.
This delay line can be implemented in fixed boundary RAM in a DSP chip if new
data values are written into memory, overwriting the oldest value. To facilitate
memory addressing, old data values are read from memory starting with the value
one location after the value that was just written. In a 4-tap FIR filter, for example,
x(4) is written into memory location 0, and data values are then read from locations
1, 2, 3, and 0. This example can be expanded to accommodate any number of taps.
By addressing data memory locations in this manner, the address generator need
only supply sequential addresses regardless of whether the operation is a memory
read or write. This data memory buffer is called circular because when the last
location is reached, the memory pointer must be reset to the beginning of the buffer.
The coefficients are fetched simultaneously with the data. Due to the addressing
scheme chosen, the oldest data sample is fetched first. Therefore, the last coefficient
must be fetched first. The coefficients can be stored backwards in memory: h(N–1) is
the first location, and h(0) is the last, with the address generator providing
incremental addresses. Alternatively, coefficients can be stored in a normal manner
with the accessing of coefficients starting at the end of the buffer, and the address
generator being decremented.
This allows direct support of the FIR filter unit delay taps without software
overhead. These data characteristics are DSP algorithm-specific and must be
supported in hardware to achieve the best DSP performance. Implementing circular
buffers in hardware allows buffer parameters (i.e. start, length, etc.) to be set up
outside of the core instruction loop. This eliminates the need for extra instructions
within the loop body. Lack of a hardware implementation for circular buffering can
significantly impact MAC performance.
ZERO OVERHEAD LOOPING
Zero overhead looping is required by the repetitive nature of the kernel equation.
The multiply-accumulate function and the data fetches required are repeated N
times every time the kernel function is calculated. Traditional microprocessors
7.5
DSP HARDWARE
implement loops that have one instruction execution time or more of overhead
associated with repeating the loop. Analog Devices’ DSP architectures provide
hardware support that eliminates the need for looping instructions within the loop
body. For true DSP architectures, the difference of zero overhead body looping and
programmed looping can easily exceed 20% cycle time.
SUMMARY
Any processor can accomplish any software task, given enough time. However,
DSPs are optimized for the unique computational requirements of real-time, realworld signal processing. Traditional computers are better suited for tasks that can
be performed in non-real-time. In the following section, we will examine the
architecture of a high-performance 16-bit fixed-point DSP Microcomputer, the
Analog Devices’ ADSP-21xx-family.
ADSP-21XX 16-BIT FIXED-POINT DSP CORE
Traditional microprocessors use the Von Neumann architecture (named after the
American mathematician John Von Neumann) as shown in Figure 7.4A. The Von
Neumann architecture consists of a single memory which contains data and
instructions and a single bus for transferring data and instructions into and out of
the CPU. Multiplying two numbers requires at least three cycles: two cycles are
required to transfer the two numbers into the CPU, and one cycle to transfer the
instruction. This architecture is satisfactory when all the required tasks can be
executed serially. In fact, most general purpose computers today use the Von
Neumann architecture.
MICROPROCESSOR ARCHITECTURES
A:
VON NEUMANN
B:
HARVARD
C:
ADI MODIFIED
HARVARD
MEMORY:
Data and
Instructions
Address Bus
PROGRAM
MEMORY:
Instructions
Only
PM Address Bus
PROGRAM
MEMORY:
Instructions
and Data
PM Address Bus
CPU
Data Bus
DM Address Bus
CPU
PM Data Bus
PM Data Bus
DM Data Bus
CPU
Instruction
Cache
Figure 7.4
7.6
DM Address Bus
DM Data Bus
DATA
MEMORY:
Data Only
DATA
MEMORY:
Data Only
DSP HARDWARE
For faster processing, however, the Harvard architecture shown in Figure 7.4B is
more suitable. This is named for the work done at Harvard University under the
leadership of Howard Aiken. Data and program instructions each have separate
memories and buses as shown. Since the buses operate independently, program
instructions and data can be fetched at the same time, thereby improving the speed
over the single bus Von Neumann design. In order to perform a single FIR filter
multiply-accumulate, an instruction is fetched from the program memory, and
during the same cycle, a coefficient can be fetched from data memory. A second cycle
is required to fetch the data word from the data memory.
Figure 7.4C illustrates Analog Devices’ modified Harvard architecture where
instructions and data are allowed in the program memory. For example, in the case
of a digital filter, the coefficients are stored in the program memory, and the data
samples in the data memory. A coefficient and a data sample can thus be fetched in
a single cycle. In addition to fetching the coefficient from program memory and a
data sample from data memory, an instruction must also be fetched from program
memory. Analog Devices’ DSPs handle this in one of two ways. In the first method,
the program memory is accessed twice (double pumped) in an instruction cycle. The
ADSP-218x series uses this method. In the second method, a program memory
cache is provided. If an algorithm requires dual data fetches, the programmer places
one buffer in program memory and the other in data memory. The first time the
processor executes an instruction, there is a one-cycle stall because it must fetch the
instruction and the coefficient over the program memory data bus. However,
whenever this conflict occurs, the DSP “caches” the instruction in a cache memory.
The next time this instruction is required, the program sequencer obtains it from
the cache, while the coefficient is obtained over the program memory data bus. The
cache method if used in the ADSP-219x family as well as in the SHARC family.
DIGITAL FILTER EXAMPLE
Now that the fundamental architecture of the ADSP-21xx family has been
presented, a simple FIR filter design will illustrate the ease of programming the
family. Pseudocode for an FIR filter design is shown in Figure 7.5. For Analog
Devices’ DSPs, all operations within the filter loop are completed in one instruction
cycle, thereby greatly increasing efficiency. Extra instructions are not required to
repeat the loop. This is referred to as zero-overhead looping. The actual FIR filter
assembly code for the ADSP-21xx family of fixed point DSPs is shown in Figure 7.6.
The arrows in the diagram point to the actual executable instructions (7 lines), the
rest of the code are simply comments added for clarification. The first instruction
(labeled fir:) sets up the computation by clearing the MR register and loading the
MX0 and MY0 registers with the first data and coefficient values from data and
program memory. The multiply-accumulate with dual data fetch in the convolution
loop is then executed N–1 times in N–1 cycles to compute the sum of the first N–1
products. The final multiply-accumulate instruction is performed with the rounding
mode enabled to round the result to the upper 24 bits of the MR register. The MR1
register is then conditionally saturated to its most positive or negative value based
on the status of the overflow flag contained in the MV register. In this manner,
results are accumulated to the full 40-bit precision of the MR register, with
saturation of the output only if the final result overflowed beyond the least
significant 32 bits of the MR register.
7.7
DSP HARDWARE
PSEUDOCODE FOR FIR FILTER PROGRAM
USING A DSP WITH CIRCULAR BUFFERING
1. Obtain sample from ADC (typically interrupt driven)
2. Move sample into input signal’s circular buffer
3. Update the pointer for the input signal’s circular buffer
4. Zero the accumulator
5. Implement filter (control the loop through each of the coefficients)
6. Fetch the coefficient from the coefficient’s circular buffer
7. Update the pointer for the coefficient’s circular buffer
8. Fetch the sample from the input signal’s circular buffer
9. Update the pointer for the input signal’s circular buffer
10. Multiply the coefficient by the sample
11. Add the product to the accumulator
12. Move the filtered sample to the DAC
ADSP21xx Example code:
CNTR = N-1;
DO convolution UNTIL CE;
convolution:
MR = MR + MX0 * MY0(SS), MX0 = DM(I0,M1), MY0 = PM(I4,M5);
Figure 7.5
ADSP-21XX FIR FILTER ASSEMBLY CODE
(SINGLE PRECISION)
.MODULE
{
fir_sub;
FIR Filter Subroutine
Calling Parameters
I0 --> Oldest input data value in delay line
I4 --> Beginning of filter coefficient table
L0 = Filter length (N)
L4 = Filter length (N)
M1,M5 = 1
CNTR = Filter length - 1 (N-1)
Return Values
MR1 = Sum of products (rounded and saturated)
I0 --> Oldest input data value in delay line
I4 --> Beginning of filter coefficient table
Altered Registers
MX0,MY0,MR
Computation Time
(N - 1) + 6 cycles = N + 5 cycles
All coefficients are assumed to be in 1.15 format. }
.ENTRY
fir:
fir;
MR=0, MX0=DM(I0,M1), MY0=PM(I4,M5);
CNTR = N-1;
DO convolution UNTIL CE;
MR=MR+MX0*MY0(SS), MX0=DM(I0,M1), MY0=PM(I4,M5);
MR=MR+MX0*MY0(RND);
IF MV SAT MR;
RTS;
convolution:
.ENDMOD;
Figure 7.6
7.8
DSP HARDWARE
The ADSP-21xx family architecture (Figure 7.7) is optimized for digital signal
processing and other high-speed numeric processing applications. This family of
DSPs combine the complete ADSP-2100 core architecture (three computational
units, data address generators, and a program sequencer) with two serial ports, a
programmable timer, extensive interrupt capabilities and on-board program and
data memory RAM. ROM-based versions are also available.
ADSP-21XX CORE ARCHITECTURE
Data
Address
Generator 1
Data
Address
Generator 2
Cache
Memory
Instruction
Register
Program
Sequencer
PMA BUS
DMA BUS
PMD BUS
Bus
Exchange
DMD BUS
Input Regs
Input Regs
Input Regs
ALU
MAC
Shifter
Output Regs
Output Regs
Output Regs
R BUS
Figure 7.7
The ADSP-21xx's flexible architecture and comprehensive instruction set support a
high degree of operational parallelism. In one cycle the ADSP-21xx can generate the
next program address, fetch the next instruction, perform one or two data moves,
update one or two data address pointers, perform a computational operation, receive
and transmit data via the two serial ports, and update a timer.
7.9
DSP HARDWARE
ADSP-21XX CORE ARCHITECTURE
■ Buses
◆ Program Memory Address (PMA)
◆ Data Memory Address (DMA)
◆ Program Memory Data (PMD)
◆ Data Memory Data (DMD)
◆ Result (R)
■ Computational Units
◆ Arithmetic Logic Unit (ALU)
◆ Multiply-Accumulator (MAC)
◆ Shifter
■ Data Address Generators
■ Program Sequencer
■ On-Chip Peripheral Options
◆ Program Memory RAM or ROM
◆ Data Memory RAM
◆ Serial Ports
◆ Timer
◆ Host Interface Port
◆ DMA Port
Figure 7.8
BUSES
The ADSP-21xx processors have five internal buses to ensure efficient data transfer.
The program memory address (PMA) and data memory address (DMA) buses are
used internally for the addresses associated with program and data memory. The
program memory data (PMD) and data memory data (DMD) buses are used for the
data associated with the memory spaces. Off chip, the buses are multiplexed into a
single external address bus and a single external data bus; the address spaces are
selected by the appropriate control signals. The result (R) bus transfers
intermediate results directly between the various computational units.
The PMA bus is 14-bits wide allowing direct access of up to 16K words of data. The
data memory data (DMD) bus is 16-bits wide. The DMD bus provides a path for the
contents of any register in the processor to be transferred to any other register or to
any data memory location in a single cycle. The data memory address comes from
two sources: an absolute value specified in the instruction code (direct addressing)
or the output of a data address generator (indirect addressing). Only indirect
addressing is supported for data fetches from program memory.
The program memory data (PMD) bus can also be used to transfer data to and from
the computational units through direct paths or via the PMD-DMD bus exchange
unit. The PMD-DMD bus exchange unit permits data to be passed from one bus to
the other. It contains hardware to overcome the 8-bit width discrepancy between the
two buses, when necessary.
7.10
DSP HARDWARE
Program memory can store both instructions and data, permitting the ADSP-21xx
to fetch two data operands in a single cycle, one from program memory and one
from data memory. The corresponding instruction is obtained directly from program
memory by “double pumping” (ADSP-218x series) or from a cache memory (ADSP219x and SHARC series).
COMPUTATIONAL UNITS (ALU, MAC, SHIFTER)
The processor contains three independent computational units: the arithmetic logic
unit (ALU), the multiplier-accumulator (MAC), and the barrel shifter. The
computational units process 16-bit data directly and have provisions to support
multiprecision computations. The ALU has a carry-in (CI) bit which allows it to
support 32-bit arithmetic.
The ALU provides a standard set of arithmetic and logic functions: add, subtract,
negate, increment, decrement, absolute value, AND, OR, EXCLUSIVE OR and
NOT. Two divide primitives are also provided.
ARITHMETIC LOGIC UNIT (ALU) FEATURES
■ Add, Subtract, Negate, Increment, Decrement, Absolute
Value, AND, OR, Exclusive OR, NOT
■ Bitwise Operators, Constant Operators
■ Multi-Precision Math Capabilites
■ Divide Primitives
■ Saturation Mode for Overflow Support
■ Background Registers for Single-Cycle Context Switch
■ Example Instructions:
◆ IF EQ AR = AX0 + AY0;
◆ AF = MR1 XOR AY1;
◆ AR = TGLBIT 7 OF AX1;
Figure 7.9
The MAC performs single-cycle multiply, multiply/add, and multiply/subtract
operations. It also contains a 40-bit accumulator which provides 8-bits of overflow in
successive additions to ensure that no loss of data occurs; 256 overflows would have
to occur before any data is lost. Special instructions are provided for implementing
block floating-point scaling of data. A set of background registers is also available in
the MAC for interrupt service routines. If after a DSP routine is finished and the
MV flag has been set, this means that the register contains a word greater than 32
bits. The register can be “saturated” using the saturation routine which normalizes
the 40-bit word to either a negative fullscale or positive fullscale 32-bit word in 1.32
format.
7.11
DSP HARDWARE
MULTIPLY-ACCUMULATOR (MAC) FEATURES
■ Single-Cycle Multiply, Multiply-Add, Multiply-Subtract
■ 40-Bit Accumulator for Overflow Protection (219x Adds
Second 40-Bit Accumulator)
■ Saturation Instruction Performs Single Cycle Overflow
Cleanup
■ Background Registers for Single-Cycle Context Switch
■ Example MAC Instructions:
◆ MR = MX0 * MY0(US);
◆ IF MV SAT MR;
◆ MR = MR - AR * MY1(SS);
◆ MR = MR + MX1 * MY0(RND);
◆ IF LT MR = MX0 * MX0(UU);
Figure 7.10
The shifter performs logical and arithmetic shifts, normalization, denormalization,
and derive-exponent operations. The shifter can be used to efficiently implement
numeric format control including multiword floating-point representations.
SHIFTER FEATURES
■ Normalize (Fixed-Point to Floating-Point Conversion)
■ Denormalize (Floating-Point to Fixed-Point Conversion)
■ Arithmetic and Logical Shifts
■ Block Floating Point Support
■ Derive Exponent
■ Background Registers for Single-Cycle Context Switch
■ Example Shifter Instructions:
◆ SR = ASHIFT SI BY –6(LO);
{Arithmetic Shift}
◆ SR = SR OR LSHIFT SI BY 3(HI);
{Logical Shift}
◆ SR = NORM MR1(LO);
{Normalization}
Figure 7.11
7.12
DSP HARDWARE
The computational units are arranged side-by-side instead of serially so that the
output of any unit may be the input of any unit on the next cycle. The internal
result (R) bus directly connects the computational units to make this possible.
DATA ADDRESS GENERATORS AND PROGRAM SEQUENCER
Two dedicated data address generators and a powerful program sequencer ensure
efficient use of the computational units. The data address generators (DAGs)
provide memory addresses when memory data is transferred to or from the input or
output registers. Each DAG keeps track of up to four address pointers. Whenever
the pointer is used to access data (indirect addressing), it is post-modified by the
value of a specified modify register. A length value may be associated with each
pointer to implement automatic modulo addressing for circular buffers. With two
independent DAGs, the processor can generate two addresses simultaneously for
dual operand fetches.
DAG1 can supply addresses to data memory only; DAG2 can supply addresses to
either data memory or program memory. When the appropriate mode bit is set in
the mode status register (MSTAT), the output address of DAG1 is bit-reversed
before being driven onto the address bus. This feature facilitates addressing in
radix-2 FFT algorithms.
DATA ADDRESS GENERATOR FEATURES
■ Automatic Linear Addressing and Circular Buffering
■ Each DAG Manages Four Pointers
■ Supports Dual Operand Fetch
■ Bit-Reverser (DAG1) For FFTs
■ Background Registers in ADSP-219x
■ Example DAG Instructions:
◆ AX0 = DM(I0,M3);
◆ MODIFY(I1,M2);
◆ MR = MR + MX0 * MY0, MX0 = DM(I0,M1), MY0 = PM(I4,M4);
Fetch Data Memory Data Word Located
at Address I0, Increment Pointer by M1
Fetch Program Memory Data Word Located
at Address I4, Increment Pointer by M4
Figure 7.12
7.13
DSP HARDWARE
The program sequencer supplies instruction addresses to the program memory. The
sequencer is driven by the instruction register which holds the currently executing
instruction. The instruction register introduces a single level of pipelining into the
program flow. Instructions are fetched and loaded into the instruction register
during one processor cycle, and executed during the following cycle while the next
instruction is prefetched. To minimize overhead cycles, the sequencer supports
conditional jumps, subroutine calls and returns in a single cycle. With an internal
loop counter and loop stack, the processor executes looped code with zero overhead.
No explicit jump instructions are required to loop. The sequencer also efficiently
processes interrupts with its interrupt controller for fast interrupt response with
minimum latency. When an interrupt occurs, it causes a jump to a known specified
location in memory. Short interrupt service routines can be coded in place. For
interrupt service routines with more than four instructions, program control is
transferred to the service routine by means of a JUMP instruction placed at the
interrupt vector location.
PROGRAM SEQUENCER FEATURES
■
■
■
■
■
Generates Next Instruction Address
Low-Latency Interrupt Handling
Hardware Stacks
Single-Cycle Conditional Branch (218x)
Supports Zero-Overhead Looping
ADSP21xx Example code:
CNTR = 10;
DO endloop UNTIL CE;
IO(DACCONTROL) = AX0;
MR = MR + MX0 * MY0(SS), MX0 = DM(I0,M1), MY0 = PM(I4,M5);
endloop:
IF MV SET FL1;
IF EQ CALL mysubroutine;
Figure 7.13
ADSP-21XX-FAMILY ON-CHIP PERIPHERALS
The discussion so far has involved the core architecture of the fixed-point ADSP21xx DSPs which is common to all members of the family. This section discusses the
on-chip peripherals which have different configurations and options depending on
the particular processor in the family. The ADSP-218x architecture is shown in
Figure 7.14.
7.14
DSP HARDWARE
ADSP-218x FAMILY ARCHITECTURE
• Code Compatibility
• Pin-for-Pin Compatibility
POWERDOWN
CONTROL
FULL MEMORY
MODE
ADSP-2100 CORE
DATA ADDRESS
GENERATORS
DAG1
DAG2
MEMORY
PROGRAM
SEQUENCER
PROGRAM
MEMORY
EXTERNAL
ADDRESS
BUS
PROGRAMMABLE
EXTERNAL I / O
and FLAG PINS
DATA
MEMORY
EXTERNAL
DATA BUS
PMA
DMA
PMD
BYTE DMA
CONTROLLER
DMD
OR
EXTERNAL
DATA BUS
ARITHMETIC UNITS
SERIAL PORTS
ALU MAC SHIFTER
SPORT 0 SPORT 1
TIMER
INTERNAL
DMA PORT
HOST MODE
External bus features are multiplexed for 100-pin parts. All are available on 128-pin parts
Figure 7.14
ADSP-21xx ON-CHIP PERIPHERALS:
MEMORY INTERFACE
■ All Family Members use an Enhanced Harvard Architecture
◆ Separate Program Memory and Data Memory Spaces
◆ Can access Data Values in Program Memory
■ Different Family Members Have Different Memory Configurations
■ External Memory Interface Supports Both Fast and Slow Memories
with Programmable Wait States
■ Supports DSP Boot Loading from Byte Memory Space or from a
Host Processor
■ Supports Memory-Mapped Peripherals Through I / O Space
■ Bus Request / Grant Mechanism for Shared External Bus
Figure 7.15
7.15
DSP HARDWARE
The 21xx-family comes with a variety of on-chip memory options, and the newer
218x family has up to 48K words of program memory and 56K words of data
memory. All family members use the modified Harvard architecture which provides
separate program and data memory and allows data to be stored in program
memory. The external memory interface supports both fast and slow memories with
programmable wait states. The ADSP-218x family also supports memory-mapped
peripherals through I/O space.
All 21xx parts (except the ADSP-2105) have two double-buffered serial ports
(SPORTs) for transmitting or receiving serial data. Each SPORT is bi-directional,
full duplex, double-buffered, and has its own programmable serial clock. The
SPORT word length can be configured from 3 to 16 bits. Data can be framed or
unframed. Each SPORT generates an interrupt and supports A-law and u-law
companding.
ADSP-21xx ON-CHIP PERIPHERALS:
SERIAL PORTS (SPORTs)
■ ADSP-21xx SPORTs Are Used For Synchronous Communication
■ Full Duplex
■ Fully Programmable
■ Autobuffer/DMA Capability
■ TDM Multi-Channel Capability
■ A-law and u-law Companding
■ Data Rates of 25 Mbits/sec and Above
■ Glueless Interface to a Wide Range of Serial Peripherals or
Processors
■ 219x DSPs Add SPI and UART Serial Ports (With Boot Capability)
Figure 7.16
The ADSP-218x-family internal direct-memory-access (IDMA) port supports booting
from and runtime access by a host processor. This feature allows data to be
transferred to and from internal memory in the background while continuing
foreground processing. The IDMA port allows a host processor to access all of the
DSPs internal memory without using mailbox registers. The IDMA port supports 16
and 24-bit words, and 24-bit transfers take two cycles to complete.
7.16
DSP HARDWARE
ADSP-21xx ON-CHIP PERIPHERALS:
INTERNAL DMA (IDMA)
■ Allows an External System to Access DSP Internal Memory
■ External Device or DSP Can Specify Internal Starting Address
■ Address Automatically Increments to Speed Throughput
■ 16-Bit Bus Supports Both Data and Instruction Transfers (219x
has 8-Bit Bus Support)
■ Single DSP Processor-Cycle Transfers
■ Supports Power-On Booting
Figure 7.17
The ADSP-218x-family also has a byte memory interface which supports booting
from and runtime access to 8-bit memories. It can access up to 4MB. This memory
space takes the place of the boot memory space found on other ADSP-21xx family
processors. Byte memory consists of 256 pages of 16K x 8 locations. This memory
can be written and read in 24-bit, 16-bit, or 8-bit left or right justified transfers.
Transfers happen in the background to the DSP internal memory by stealing cycles.
ADSP-21xx ON-CHIP PERIPHERALS:
BYTE DMA PORT (BDMA)
■ Provides Bulk Storage for Both Data and Program Code
■ Can Access up to 4 Mbytes of External Code & Data
■ Supports Multiple Data Formats
◆ Automatic Data Packing/Unpacking to 16 and 24 bits
◆ 8-Bit Transfers, Left- or Right-Justified
■ Background Transfer to DSP Internal Memory
◆ One Cycle per Word
◆ DSP Specifies Destination/Source and Word Count
■ Supports Power-On Booting
■ Allows Multiple Code Segments
◆ DSP Can Overlay Code Sections
◆ Processor Can Run During Transfer, or Halt and Restart
Figure 7.18
7.17
DSP HARDWARE
The ADSP-217x, ADSP-218x, and ADSP-21msp5x devices provide a powerdown
feature that allows the processor to enter a very low power state (less than 1mW)
through hardware or software control. This feature is extremely useful for batterypowered applications. During some of the powerdown modes, the internal clocks are
disabled, but the processor registers and memory are maintained.
ADSP-21xx INTERNAL PERIPHERALS:
POWERDOWN
■ Non-Maskable Interrupt
◆ Hardware Pin (PWD), or Software Forced
■
Holds Processor in CMOS Standby
■
Rapid CLKIN-Cycle Recovery
■
Acknowledge Handshake (PWDACK)
■
Ideal for Battery-Powered Applications
■ 219x is Fully Static
Figure 7.19
From the above discussions it should be obvious that the ADI DSPs are designed for
maximum efficiency when performing typical DSP functions such as FFTs or digital
filtering. ADI DSPs can perform several operations in an instruction cycle as has
been shown in the above filter example. DSPs are often rated in terms of Millions of
Instructions per Second, or MIPS. However, the MIPS rating does not tell the entire
story. For example if processor A has a instruction rate of 50 MIPS and can perform
1 operation per instruction, it can perform 50 Million Operations per Second, or 50
MOPS. Now assume that processor B has an instruction rate of 20 MIPS, but can
perform 4 operations per instruction. Processor B can therefore perform 80 Million
Operations per Second, or 80 MOPS, and is actually more efficient than processor A.
A better way to evaluate DSP performance is to use well defined benchmarks such
an FIR filter with a prescribed number of taps or an FFT of known size. Benchmark
comparisons eliminate the confusion often associated with MIPS or MOPS ratings
alone and are discussed later in this section. Even this form of benchmarking does
not give a true comparison of performance between two processors. Analysis of the
target system requirements, processor architecture, memory needs and other
factors must be considered.
7.18
DSP HARDWARE
The ADSP-219x family maintains code compatibility with the ADSP-218x.
Streamlined for faster processing and improved C-compiler efficiency, this new
family will include DSPs with speeds between 100 and 300MIPS and power
consumption as low as 0.4mA/MIP. JTAG support is also included to provide a more
robust software emulation and test capability. A block diagram of the family is
shown in Figure 7.20.
ADSP-219x SERIES ARCHITECTURE
ADSP-219x
Processor Core
DAG1
Instr.
Register
Seq.
DAG2
Internal
Memory System
Cache
Program
Sequencer
PM
DM
JTAG &
EMULATION
24
PMA
24
DMA
24
PMD
DATA REGISTERS
ALU
MAC
Shift
Computational Units
PERIPHERAL AND
DMA INTERFACE
16
DMD
EXTERNAL
MEMORY
INTERFACE
■ Improved addressing (14-bit reach
increased to 24-bit reach)
■ Compiler-efficient data register
design
■ Program sequencer with Cache for
fast code execution
■ JTAG support for easy development
Figure 7.20
The address reach of the ADSP-219x series has been extended from the ADSP218x’s 14-bit reach to a 24-bit reach. This supports 64K word direct memory
addressing or 16M word paged memory addressing. All existing addressing modes
are supported, and five new DAG addressing modes have been added.
Many of the enhancements in the ADSP-219x are designed to improve compiler
efficiency. A global register allocator and support for register file-like operand
access reduces spills and reduces reliance on the local stack. The compiler features
DSP intrinsic support including fractional and complex. On-chip cache memory has
also been added.
The ADSP-219x core will serve as a key DSP technology for ADI’s 16-bit general
purpose DSP offerings, and embedded DSP solutions, where application-specific
circuitry and software are custom-designed to a customer’s precise requirements.
For performance-driven applications, multiple cores will be integrated together on a
single die. In the future, four cores will be combined in a family of devices capable of
delivering 1.2 billion MACs per second per square inch. Power-conscious designers
will appreciate operating currents of 0.15mA/MIPS.
7.19
DSP HARDWARE
ADSP-219x FAMILY KEY SPECIFICATIONS
■ Code Compatible
◆ Compatible with ADSP-218x Series
◆ Single Cycle Instruction Execution, Zero-Over Head Looping,
Single Cycle Context Switch
■ Performance
◆ Architectural Performance Beyond 300 MIPs.
◆ Fully Transparent Instruction Cache
■ Compiler-Friendly and Tool-Friendly
◆ 64K Word Direct and 16 Mword Paged Memory Support
◆ 5 New DAG Addressing Modes
◆ Register File-Like Operand Access
◆ JTAG Interface Support
Figure 7.21
The history of the Analog Devices 16-bit fixed-point DSP family is shown in Figure
7.22. Notice the migration in performance as well as improvements in power
dissipation and packaging while maintaining code compatibility between the
various devices. The newer families offer 3.3V (L-series) and 2.5V (M-series)
operation for further efficiency. Earlier DSPs were packaged in expensive pin-gridarray (PGA) packages or plastic-leaded-chip-carriers (PLCCs), but these have been
largely replaced by plastic-quad-flat-packs (PQFPs), and more recently by thin
(1.6mm) quad-flat-packs (TQFPs). Note: In 1998, JEDEC changed the specifications
for the TQFP package designation, assigning it to packages 1.0mm thick. Previously
labeled TQFP packages (1.6mm thick) are now designated as LQFP per the JEDEC
specification.
The 144-ball miniBGA package (see Figure 7.23) from ADI represents innovative
packaging combined with low power consumption (0.4mA per MIP), and allows
75MIPS operation and more than 2M bits of SRAM in a 1cm2 package which is
1.35mm thick. For example, the 75MIPS ADSP-2188M has 48K of 24-bit program
memory and 56K of 16-bit data memory for a total of 48K×24 + 56K×16 = 2028K
bits with a power dissipation of less than 100mW.
7.20
DSP HARDWARE
16-BIT DSP FAMILY TREE: HISTORY OF IMPROVEMENTS
IN PACKAGING - POWER - PERFORMANCE
Performance
miniBGA
1.6mm thick
150-300 MIPS
0.4 mA/MIPs
219x
LQFP
4.45mm thick
PLCC
21msp5x
20 MIPS
2K/1K RAM
216x
2111
b l e BGA
2188M
ati
m p all Mini
o
2189M
C
B
4
P i n FP & 14
LQ
n
i
P
“L” = 3.3V
100
2185/L/M
“M” = 2.5V
2186/L/M
“N” = 1.8V
33 - 75 MIPS
2184/L
0.4 mA/MIPS
2181/3
33 MIPS
2K/2K RAM
210x
PQFP
2.45mm thick
Code Compatible
Figure 7.22
‘M’ SERIES OFFERS THE LARGEST
MIPS/MEMORY DENSITY WITH miniBGA PACKAGING!!!
■ 75 MIPS and >2Mbit in 1cm2
1.35mm Thick
■ Small size with no compromises in
performance: 144-balls
m
10m
■ Up to 2 Mbits on-chip SRAM
■ Innovative packaging for highly
portable applications
mm
10
0.8mm
Pitch
miniBGA package
144-ball grid array (BGA)
■ Small package size and 0.4mA per MIP
targets power-sensitive applications
■ Includes ALL ‘M’ series members plus
the popular 2183, 2185L, 2186, & 2186L
derivatives
Figure 7.23
7.21
DSP HARDWARE
ADI 16-Bit DSP ROADMAP
16 Years of Code Compatibility & Beyond
1000
312 - 600 MIPS
8M - 16MBits
2100 Core
ple
Com
100
• As Low as 0.5mW/MIPS
• As High as 8MIPs/mm2
• As Low as 0.7mW/MIPS
• As High as 0.5MIP/mm2
500
MMACs
600 - 1200 MIPS
16M - 128MBits
218x Core
Multiple Core
Devices
te C
ibil
mpat
o
C
ode
ity
29-80 MIPS
160K - 2MBits
10 - 33 MIPS
32K - 80KBits
150-300 MIPS
640K - 32MBits
• As Low as 0.4mW/MIPS
• As High as 3MIPs/mm2
Single Core
Devices
• As Low as 0.6mW/MIPS
• As High as 0.8MIP/mm2
• As Low as 6mW/MIPS
• As High as 0.1MIP/mm2
219x Core
1985
2000+
Figure 7.24
ADSP-218x ROADMAP
1.8V
2.5V
500
5/3.3V
312 MIPS
8MBits
SRAM
MMACs
100
• 2.6mW/MIPS
• 0.3MIP/mm2
29-52 MIPS
160K- 1.2MBits
SRAM
• As Low as 2.6mW/MIPS
• As High as 0.5MIP/mm2
600 MIPS
16MBits
SRAM
600 MIPS
16MBits
SRAM
• 0.7mW/MIPS
• 0.5MIPs/mm2
• 1mW/MIPS
• 0.5MIP/mm2
Code
& Pin
s
ption
O
e
l
patib
Com
75 MIPS
320K - 2MBits
SRAM
•As Low as 1mW/MIPS
•As High as 0.75MIP/mm2
1994
80 MIPS
640K -2MBits
SRAM
•As Low as 0.7mW/MIPS
•As High as 0.8MIPs/mm2
Single Core
Devices
2000+
Figure 7.25
7.22
Multiple
Core
Devices
1.5V
40 MIPS
640K -2MBits
SRAM
•As Low as 0.6mW/MIPS
•As High as 0.4MIPs/mm2
DSP HARDWARE
ADSP-219x ROADMAP
ADSP-219x
1200 MIPS
re
-Co
e
l
p
lti
Mu
e
c
n
rma
o
f
r
e
st P
ADSP-219x
e
h
300 MIPS
Hig
1000
500
MMACs
100
ADSP-219x
600 MIPS
218x to 219x
16 Years of
continuing code
compatible
architectures!
ADSP-219x
320 MIPS
ADSP-219x
200 MIPS
ADSP-2191
160 MIPS
0.5 mW / MIPS
2.5V
h
Hig
ADSP-219x
0.1 mW / MIPS
1.5V
ADSP-219x
0.05 mW / MIPS
1.0V
Sin
fficient
E
r
e
w
Po
ADSP-218X
<2000
nce
rma
o
f
r
Pe
ore
e-C
l
g
Sin
2000
2001
2002
re
gle-Co
2003
Figure 7.26
FIXED-POINT VERSUS FLOATING POINT
DSP arithmetic can be divided into two catagories: fixed-point and floating-point.
These refer to the format used to store and manipulate the numbers within the
devices. The Analog Devices’ fixed-point DSPs such as those discussed so far
represent each number with 16 bits. There are four common ways that 216 = 65,536
possible bit patterns can represent a number. In unsigned integer format, the stored
number takes on any integer value from 0 to 65,536. In signed integer format, two’s
complement is used to make the range include negative numbers, from –32,768 to
+32,767. Using unsigned fractional format, the 65,536 levels are spread uniformly
between 0 and +1. Finally, signed fractional format allows negative numbers, with
65,536 levels equally spaced between –1 and +1.
The ADSP-21xx family arithmetic is optimized for the signed fractional format
denoted by 1.15 (“one dot fifteen”). In the 1.15 format, there is one sign bit (the
MSB) and fifteen fractional bits representing values from –1 up to 1 LSB less than
+1 as shown in Figure 7.27.
7.23
DSP HARDWARE
16-BIT FIXED POINT ARITHMETIC
FRACTIONAL 1.15 FORMAT
BIT WEIGHT
MSB
LSB
–20 2–1 2–2 2–3 2–4 2–5 2–6 2–7 2–8 2–9 2–10 2–11 2–12 2–13 2–14 2–15
HEX
BINARY
DECIMAL
7FFF
0111
1111
1111
1111
+0.999969
0001
0000
0000
0000
0001
+0.000031
0000
0000
0000
0000
0000
+0.000000
FFFF
1111
1111
1111
1111
–0.000031
8000
1000
0000
0000
0000
–1.000000
Figure 7.27
This convention can be generalized as “I.Q”, in which I is the number of bits to the
left of the radix point, and Q is the number of bits to the right. For example, full
unsigned integer number representation is 16.0 format. For most signal processing
applications, however, fractional numbers (1.15) are assumed. Fractional numbers
have the advantage that the product of two fractional numbers is smaller than
either of the numbers.
By comparison, floating point DSPs typically use a minimum of 32 bits to represent
each number. This results in many more possible numbers than for 16-bit fixed
point, 232 = 4,294,967,296 to be exact. More importantly, floating point greatly
increases the range of values that can be expressed. The most common floating
point standard is ANSI/IEEE Standard 754-1985, where the largest and smallest
numbers allowed by the standard are ±3.4×1038 and ±1.2×10–38, respectively. Note
that the 754 standard reserves some of the possible range to free up bit patterns
which allow other special classes of numbers such as ±0 and ±∞, for example.
The IEEE-754 floating point standard is described in more detail in Figure 7.28.
The 32-bit word is divided into a sign bit, S, an 8-bit exponent, E, and a 23-bit
mantissa, M. The relationship between the decimal and binary IEEE-754 floating
point equivalent is given by the equation:
NUMBER10 = ( −1)S × 1.M × 2(E −127) .
Notice that the “1.” is assumed to precede the “M”, and that a “bias” of 127 is
subtracted from the exponent “E” so that “E” is always a positive number.
7.24
DSP HARDWARE
SINGLE PRECISION IEEE-754
32-BIT FLOATING POINT FORMAT,
NUMBER10
= (–1)S × 1.M × 2(E–127)
Bias
Assumed
MSB
0/1
27
SIGN
BIT = S
LSB MSB
26
25
24
23
22
21
20
2–1
LSB
2–2
EXPONENT = E
8-BITS
0 00000111 1100…00
+
7
0.75
2–3
2–4
2–22
2–23
MANTISSA = M
23-BITS
Bias
+ 1.75 × 2(7–127)
= + 1.316554 × 10–36
Assumed
Bias
1
–
10000001 0110…00
129
0.375
– 1.375 × 2(129–127)
= – 5.500000
Assumed
Figure 7.28
In the case of extended precision floating point arithmetic, there is one sign bit, the
mantissa is 31-bits, the exponent is 11-bits, and the total word length is 43-bits.
Extended precision thus adds 8 bits of dynamic range to the mantissa and can
execute almost as fast as single precision, since accumulators can readily be
extended beyond 32 bits. On the other hand, true 64-bit double precision (52-bit
mantissa, 11-bit exponent, and 1 sign bit) requires extra processor cycles.
Requirements for double precision are rare in most DSP applications.
Many DSP applications can benefit from the extra dynamic range provided by 32-bit
floating point arithmetic. In addition, programming floating point processors is
generally easier, because fixed-point problems such as overflow, underflow, data
scaling, and round-off error are minimized, if not completely eliminated. Although
the floating point DSP may cost slightly more than the fixed point DSP,
development time may well be shorter with floating point.
While all floating point DSPs can also handle fixed point numbers (required to
implement counters, loops, and ADC/DAC signals), this doesn’t necessarily mean
that fixed point math will be carried out as quickly as the floating point operations;
it depends on the internal DSP architectures. For instance, the SHARC DSPs from
Analog Devices are optimized for both floating point and fixed point operations, and
execute them with equal efficiency. For this reason, the SHARC devices are often
referred to as “32-bit DSPs” rather than just “Floating Point.”
7.25
DSP HARDWARE
FIXED POINT VS. FLOATING POINT ARITHMETIC
■ 16-Bit Fixed-Point:
◆ 216 = 65,536 Possible Numbers
■ 32-Bit Floating Point:
◆ Biggest Number: ±6.8 × 1038
754 Std: ±3.4 × 1038
◆ Smallest Number: ±5.9 × 10–39
754 Std: ±1.2 × 10–38
■ Extended-Precision (40-Bits: Sign + 8-Bit Exponent + 31-Bit Mantissa)
■ Double-Precision (64-Bits: Sign + 11-Bit Exponent + 52-Bit Mantissa)
■ 32-Bit Floating Point
◆ More Precision
◆ Much Larger Dynamic Range
◆ Easier to Program
Figure 7.29
ADI SHARC® FLOATING POINT DSPS
The ADSP-2106x Super Harvard Architecture SHARC is a high performance 32-bit
DSP. The SHARC builds on the ADSP-21000 family DSP core to form a complete
system-on-a-chip, adding a dual-ported on-chip SRAM and integrated I/O
peripherals supported by a dedicated I/O bus. With its on-chip instruction cache, the
processor can execute every instruction in a single cycle. Four independent buses for
dual data, instructions, and I/O plus crossbar switch memory connections comprise
the Super Harvard Architecture of the ADSP-2106x shown in Figure 7.30.
A general purpose data register file is used for transferring data between the
computation units and the data buses, and for storing intermediate results. The
register file has two sets (primary and alternate) of 16 registers each, for fast
context switching. All of the registers are 40 bits wide. The register file, combined
with the core processor’s super Harvard architecture, allows unconstrained data
flow between computation units and internal memory.
The ADSP-2106x SHARC processors address the five central requirements for DSPs
established in the ADSP-21xx family of 16-bit fixed point DSPs: (1) fast, flexible
arithmetic computation units, (2) unconstrained data flow to and from the
computation units, (3) extended precision and dynamic range in the computation
units, (4) dual address generators, and (5) efficient program sequencing with zerooverhead looping.
The program sequencer includes a 32-word instruction cache that enables three-bus
operation for fetching an instruction and two data values. The cache is selective –
only instructions whose fetches conflict with program memory data accesses are
cached. This allows full-speed multiply-accumulates and FFT butterfly processing.
7.26
DSP HARDWARE
TIMER
INSTRUCTION
CACHE
32x48 BIT
TWO INDEPENDENT
DUAL-PORTED BLOCKS
PROCESSOR PORT
ADDR
DAG1
8x4x32
DAG2
8x4x24
BUS
Connect
(PX)
DATA
ADDR
PROGRAM
SEQUENCER
PMA
24
DMA
32
PMD
48
DMD
40
I/O PORT
DATA
DATA
DUAL PORTED SRAM
BLOCK 1
CORE
PROCESSOR
BLOCK 0
ADI SUPER HARVARD ARCHITECTURE (SHARC®)
32-BIT DSP ARCHITECTURE FOR ADSP-2106x FAMILY
ADDR
DATA
IOD
48
ADDR
JTAG
TEST &
EMULATION
IOA EXTERNAL PORT
17
ADDR BUS
MUX
7
32
MULTIPROCESSOR
INTERFACE
48
DATA BUS
MUX
HOST PORT
MULTIPLIER
DATA
REGISTER
FILE
16x40
BIT
BARREL
SHIFTER
ALU
P
M
D
D
M
D
E I
P O
D D
IOP REGISTERS
(Memory Mapped)
Control, Status,
and Data Buffers
DMA
CONTROLLER
SERIAL PORTS
(2)
LINK PORTS
(6)
4
6
6
36
I/O PROCESSOR
Figure 7.30
SHARC KEY FEATURES
■ 100MHz Core / 300 MFLOPS Peak
■ Parallel Operation of: Multiplier, ALU, 2 Address Generators &
Sequencer
◆ No Arithmetic Pipeline; All Computations Are Single-Cycle
■ High Precision and Extended Dynamic Range
◆ 32/40-Bit IEEE Floating-Point Math
◆ 32-Bit Fixed-Point MAC’s with 64-Bit Product & 80-Bit
Accumulation
■ Single-Cycle Transfers with Dual-Ported Memory Structures
◆ Supported by Cache Memory and Enhanced Harvard
Architecture
■ Glueless Multiprocessing Features
■ JTAG Test and Emulation Port
■ DMA Controller, Serial Ports, Link Ports, External Bus, SDRAM
Controller, Timers
Figure 7.31
7.27
DSP HARDWARE
SHARC® THE LEADER IN FLOATING POINT DSP
■ SHARC is the de facto standard in multiprocessing
■ ADSP-21160 continues SHARC leadership in multiprocessing
■ ADSP-21065L is the right choice for low-cost floating point
SUPER HARVARD ARCHITECTURE:
Balancing Memory, I/O and Computational
Power...
■ High Performance Computation Unit
■ Four Bus Performance
◆ Fetch Next Instruction
◆ Access 2 data values
◆ Perform DMA for I/O
■ Memory Architecture
■ Non Intrusive DMA
Figure 7.32
The ADSP-2106x family execute all instructions in a single cycle. They handle 32-bit
IEEE floating point format, 32-bit integer and fractional fixed point formats (twoscomplement and unsigned), and extended precision 40-bit IEEE floating point
format. The processors carry extended precision throughout their computation
units, minimizing intermediate data truncation errors. When working with data onchip, the extended precision 32-bit mantissa can be transferred to and from all
computation units. The 40-bit data bus may be extended off-chip if desired. The
fixed point formats have an 80-bit accumulator for true 32-bit fixed point
computations.
The ADSP-2106x has a super Harvard architecture combined with a 10-port data
register file. In every cycle, (1) two operands can be read or written to or from the
register file, (2) two operands can be supplied to the ALU, (3) two operands can be
supplied to the multiplier, and (4) two results can be received from the ALU and
multiplier.
The ADSP-2106x family instruction set provides a wide variety of programming
capabilities. Multifunction instructions enable computations in parallel with data
transfers, as well as simultaneous multiplier and ALU operations.
7.28
DSP HARDWARE
The ADSP-21060 contains 4 Mbits of on-chip SRAM, organized as two blocks of 2
Mbits each, which can be configured for different combinations of code and data
storage. The ADSP-21062, ADSP-21061 and ADSP-21065 each include 2 Mbit,
1Mbit and 544Kbits of on-chip SRAM, respectively. Each memory block is dualported for single-cycle, independent accesses by the core processor and I/O processor
or DMA controller. The dual-ported memory and separate on-chip buses allow two
data transfers from the core and one from I/O, all in a single cycle.
While each memory block can store combinations of code and data, accesses are
most efficient when one block stores instructions and data, using the DM bus for
transfers, and the other block stores instructions and data, using the PM bus for
transfers. Using the DM bus and PM bus in this way, with one dedicated to each
memory block, assures single-cycle execution with two data transfers. In this case,
the instruction must be available in the cache. Single-cycle execution is also
maintained when one of the data operands is transferred to or from off-chip, via the
ADSP-2106x’s external port.
The ADSP-2106x’s external port provides the processor’s interface to off-chip
memory and peripherals. The 4 Gword off-chip address space is included in the
ADSP-2106x’s unified address space. The separate on-chip buses – for PM
addresses, PM data, and DM addresses, DM data, I/O addresses, and I/O data – are
multiplexed at the external port to create an external system bus with a single 32bit address bus and a single 48-bit data bus. The ADSP-2106x provides
programmable memory wait states and external memory acknowledge controls to
allow interfacing to DRAM and peripherals with variable access, hold, and disable
time requirements.
The ADSP-2106x’s host interface allows easy connection to standard microprocessor
buses, both 16-bit and 32-bit, with little additional hardware required. Four
channels of DMA are available for the host interface; code and data transfers are
accomplished with low software overhead. The host can directly read and write the
internal memory of the ADSP-2106x, and can access the DMA channel setup and
mailbox registers. Vector interrupt support is provided for efficient execution of host
commands.
The ADS-2106x offers powerful features tailored to multiprocessing DSP systems.
The unified address space allows direct interprocessor accesses of each ADSP2106x’s internal memory. Distributed bus arbitration logic is included on-chip for
simple, glueless connection of systems containing up to six ADSP-2106xs and a host
processor. Master processor changeover incurs only one cycle of overhead. Bus
arbitration is selectable as either fixed or rotating priority. Maximum throughput
for interprocessor data transfer is 240 Mbytes/second (with a 40MHz clock) over the
link ports or external port.
The ADSP-2106x’s I/O Processor (IOP) includes two serial ports, six 4-bit link ports,
and a DMA controller. The ADSP-2106x features two synchronous serial ports that
provide an inexpensive interface to a wide variety of digital and mixed-signal
peripheral devices. The serial ports can operate at the full external clock rate of the
processor, providing each with a maximum data rate of 50 Mbit/second.
Independent transmit and receive functions provide greater flexibility for serial
communications. Serial port data can be automatically transferred to and from on7.29
DSP HARDWARE
chip memory via DMA. Each of the serial ports offers a TDM multichannel mode.
They offer optional µ-law or A-law companding. Serial port clocks and frame syncs
can be internally or externally generated.
The ADSP-21060 and ADSP-21062 feature six 4-bit link ports that provide
additional I/O capabilities. The link ports can be clocked twice per cycle, allowing
each to transfer 8 bits per cycle. Link port I/O is especially useful for point-to-point
interprocessor communication in multiprocessing systems. The link ports can
operate independently and simultaneously, with a maximum data throughput of
240 Mbytes/second. Link port data is packed into 32-bit or 48-bit words, and can be
directly read by the core processor or DMA-transferred to on-chip memory. Each
link port has its own double-buffered input and output registers. Clock/acknowledge
handshaking controls link port transfers. Transfers are programmable as either
transmit or receive. There are no link ports on the ADSP-21061 or ADSP-21065
devices.
The ADSP-2106x’s on-chip DMA controller allows zero-overhead data transfers
without processor intervention. The DMA controller operates independently and
invisibly to the processor core, allowing DMA operations to occur while the core is
simultaneously executing its program. Both code and data can be downloaded to the
ADSP-2106x using DMA transfers. DMA transfers can occur between the ADSP2106x’s internal memory and external memory, external peripherals, or a host
processor. DMA transfers can also occur between the ADSP-2106x’s internal
memory and its serial ports or link ports. DMA transfers between external memory
and external peripheral devices are another option.
The internal memory of the ADSP-2106x can be booted at system powerup from an
8-bit EPROM or a host processor. Additionally, the ADSP-21060 and the ADSP21062 can also be booted through one of the link ports. Both 32-bit and 16-bit host
processors can be used for booting.
The ADSP-2106x supports the IEEE standard P1149.1 Joint Test Action Group
(JTAG) standard for system test. This standard defines a method for serially
scanning the I/O status of each component in a system. In-circuit emulators also use
the JTAG serial port to access the processor’s on-chip emulation features. EZ-ICE
Emulators use the JTAG test access port to monitor and control the target board
processor during emulation. The EZ-ICE in-circuit emulator provides full-speed
emulation to enable inspection and modification of memory, registers, and processor
stacks. Use of the processor’s JTAG interface assures non-intrusive in-circuit
emulation – the emulator does not affect target system loading or timing.
The SHARC architecture avoids processor bottlenecks by balancing core, memory,
I/O processor, and peripherals as shown in Figure 7.30. The core supports 32-bit
fixed and floating point data. The memory contributes to the balance by offering
large size and dual ports. The core can access data from one port, and the other port
is used to move data to and from the I/O processor. The I/O processor moves data to
and from the peripherals to the internal memory using zero overhead DMAs. These
operations run simultaneously to the core operation.
7.30
DSP HARDWARE
ADSP-2116X SINGLE-INSTRUCTION, MULTIPLEDATA (SIMD) CORE ARCHITECTURE
The ADSP-21160 is the first member of the second-generation ADI 32-bit DSPs. Its
core architecture is shown in Figure 7.33. Notice that the core is similar to the
ADSP-2106x core except for the width of the buses and the addition of a second
computational unit complete with its own multiplier, ALU, shifter, and register file.
This architecture is called single-instruction, multiple-data (SIMD) as opposed to
single-instruction, single-data (SISD). The second computational unit allows the
DSP to process multiple data streams in parallel. The core operates at up to 100
MIPS. At 100MHz clock operation the core is capable of 400 MFLOPS (millions of
floating point operations per second) sustained and 600 MFLOPS peak operation.
SIMD is a natural next step in increasing performance for ADI DSPs. Because their
basic architecture already allows single instruction and multiple data access, adding
another computational unit lets the architecture process the multiple data. The
SIMD architectural extension allows code-compatible higher performance parts.
ADSP-2116x CORE PROCESSOR FEATURING
SINGLE-INSTRUCTION, MULTIPLE-DATA (SIMD)
TIMER
INSTRUCTION
CACHE
32x48 BIT
ADSP-2116x
CORE PROCESSOR
DAG1
8x4x32
DAG2
8x4x32
PMA
32
DMA
32
PMD
BUS
Connect
(PX)
MULTIPLIER
PROGRAM
SEQUENCER
16/32/40/48/64
DMD
DATA
REGISTER
FILE
(PEx)
16x40
BIT
BARREL
SHIFTER
32/40/64
ALU
MULTIPLIER
DATA
REGISTER
FILE
(PEy)
16x40
BIT
BARREL
SHIFTER
ALU
Figure 7.33
7.31
DSP HARDWARE
The SIMD features of the ADSP-2116x include two computational units (PEx, PEy)
and double data-word size buses (DMD and PMD). The primary processing element,
PEx, is always enabled. The secondary processing element, PEy, is mode control
enabled. The double wide data buses provide each computational unit with its own
data set in each cycle. With SIMD enabled, both processing elements execute the
same instruction each cycle (that’s the Single-Instruction), but they execute that
instruction using different data (that’s the Multiple-Data). The SIMD-based
performance increase appears in algorithms that can be optimized by splitting the
processing of data between the two computational units. By taking advantage of the
second computational unit the cycle time can be cut in half compared to the SISD
approach for many algorithms.
The ADSP-21160 has a complete set of integrated peripherals: I/O Processor, 4
Mbyte on-chip dual-ported SRAM, glueless multiprocessing features, and ports
(serial, link, external bus, host, and JTAG). Power dissipation is approximately 2W
at 100MHz in a 400-ball 27 × 27mm PBGA package. The complete SHARC family
roadmap is shown in Figure 7.35.
ADSP-21160 32-BIT SHARC KEY FEATURES
■ SIMD (Single-Instruction, Multiple-Data) Architecture
■ Code Compatible with ADSP-2106x Family
■ 100 MHz Core / 600 MFLOPS Peak
■ On-Chip Peripherals Similar to ADSP-2106x Family
■ Dual-Ported 4Mbit SRAM
■ Glueless Multiprocessing Features
■ 400-Ball, PBGA 27×27mm Package
Figure 7.34
Figure 7.36 shows some typical coding using the SHARC family of DSPs. Note the
algebraic syntax of the assembly language which facilitates coding of algorithms
and reading the code after it is written. In a single cycle, the SHARC performs
multiplication, addition, subtraction, memory read, memory write, and address
pointer updates. In the same cycle, the I/O processor can transfer data to and from
the serial port, the link ports, memory or DMA, and update the DMA pointer.
7.32
DSP HARDWARE
SHARC ROADMAP
Commitment to Code Compatibility Into Tomorrow
M
T
UL
IP
C
O
R
ADSP-21062
S
S
E
I
MP
HP SHARC
NG
ADSP-21160M
ADSP-21060
• 120-198 MFLOPS
• 0.5 - 4 Mbits Memory
ADSP-21161N
ADSP-21061
ADSP-21065
• 10 GFLOPs
• 64 Mbits
Low Cost
HP SHARC
• 1200 MFLOPs
• <<$10 SHARC
MASS MARKET
Figure 7.35
EXAMPLE: SHARC MULTI-FUNCTION INSTRUCTION
f11 = f1 * f7, f3 = f9 + f14, f9 = f9 – f14, dm (i2, m0) = f13, f7 = pm (i8, m8);
■ In this Single-Cycle Instruction the SHARC Performs:
◆ 1 (2) Multiply
◆ 1 (2) Addition
◆ 1 (2) Subtraction
( ) = ADSP-2116x SIMD DSP
◆ 1 (2) Memory Read
◆ 1 (2) Memory Write
◆ 2 Address Pointer Updates
■ Plus the I/O Processor Performs:
◆ Active Serial Port Channels: Transmit and Receive on all Ports
◆ 6 Active Link Ports if Present
◆ Memory DMA
◆ 2 (4) DMA Pointer Updates
The Algebraic Syntax of the Assembly Language Facilitates Coding of
DSP Algorithms
Figure 7.36
7.33
DSP HARDWARE
MULTIPROCESSING USING SHARCS
Analog Devices’ SHARC DSPs such as the ADSP-21160 are optimized for
multiprocessing applications such as telephony, medical imaging, radar/sonar,
communications, 3D graphics and imaging. Figure 7.37 shows SHARC benchmark
performance on common DSP algorithms.
DSP BENCHMARKS FOR SHARC FAMILY
ADSP-21065L
SHARC
Clock Cycle
66 MHz
Instruction Cycle Time
15 ns
MFLOPS Sustained
132 MFLOPS
MFLOPS Peak
198 MFLOPS
1024 Point Complex FFT 274 µs
(Radix 4, with reversal)
FIR Filter (per tap)
15 ns
IIR Filter (per biquad)
60 ns
Matrix Multiply
(pipelined)
[3x3] * [3x1]
135 ns
[4x4] * [4x1]
240 ns
Divide (y/x)
90 ns
Square Root
135 ns
ADSP-21160
SISD
100 MHz
10 ns
200 MFLOPS
300 MFLOPS
180 µs
ADSP-21160 SIMD/
multiple channels
100 MHz
10 ns
400 MFLOPS
600 MFLOPS
90 µs
10 ns
40 ns
5 ns
20 ns
90 ns
160 ns
60 ns
90 ns
45 ns
80 ns
30 ns
45 ns
Figure 7.37
Multiprocessor systems typically use one or both of two methods to communicate
between processor nodes. One method uses dedicated point-to-point communication
channels. This method is often called data-flow multiprocessing. In the other
method, nodes communicate through a single shared global memory via a parallel
bus. The SHARC family supports the implementation of point-to-point
communication through its six link ports called link port multiprocessing. It also
supports an enhanced version of shared parallel bus communication called cluster
multiprocessing.
For applications that require high computational bandwidth, but only limited
flexibility, data flow multiprocessing is the best solution. The DSP algorithm is
partitioned sequentially across several processors and data is passed directly across
them as shown on the right side of Figure 7.38. The SHARC is ideally suited for
data flow multiprocessing applications because it elminates the need for
interprocessor data FIFOs and external memory. Each SHARC has six link ports
allowing 2D and 3D arrays as well as traditional data flow. The internal memory of
the SHARC is usually large enough to contain both code and data for most
applications using this topology. All a data flow system requires are a number of
SHARC processors and point-to-point signals connecting them.
7.34
DSP HARDWARE
MULTIPROCESSOR COMMUNICATION
EXAMPLES FOR SHARC
Dataflow
Links
Links
Links
EP
EP
EP
SHARC Cluster
Links
Links
Links
Links
Links
Links
Links
Links
Links
2D Array
Links
Links
Links
EP
EP
EP
SHARC Cluster
Expanding Clusters
■ Link Ports for point-to-point communication
■ Cluster Bus for high bandwidth parallel
communication
◆ On-chip Bus Arbitration
◆ Unified Memory Map
Figure 7.38
EXTERNAL PORT VERSUS
LINK PORT COMMUNICATIONS
■ Advantages of External Ports (EP)
◆ Communications through the EP has the Highest Bandwidth
between two SHARCs (400 Mbytes/s)
◆ Allows up to 6 SHARCs and a Host to share the EP
◆ The EP offers flexible communication of data and control
information
◆ The Shared Memory model allows a simple software structure
■ Advantage of Link Ports
◆ Each Link Port provides independent, 100 MBytes/s
communication between two SHARCs
◆ Up to 6 Link Ports (600 MBytes/s)
◆ Easily scalable to any number of SHARCs
■ Both Link Port and EP communications can be used
simultaneously
Figure 7.39
7.35
DSP HARDWARE
Cluster multiprocessing is best suited for applications where a fair amount of
flexibility is required. This is especially true when a system must be able to support
a variety of different tasks, some of which may be running concurrently. SHARC
processors also have an on-chip host interface that allows a cluster to be easily
interfaced to a host processor or even to another cluster.
Cluster multiprocessing systems include multiple SHARC processors connected by a
parallel bus that allows interprocessor access of on-chip memory as well as access to
shared global memory. In a typical cluster of SHARCs, up to six ADSP-21160
processors and a host can arbitrate for the bus. The on-chip bus arbitration logic
allows these processors to share the common bus. The SHARC’s other on-chip
features help eliminate the need for any extra glue hardware in the cluster
multiprocessor configuration. External memory, both local and global, can
frequently be eliminated in this type of system.
TIGERSHARC™: THE ADSP-TS001 STATIC
SUPERSCALAR DSP
The ADSP-TS001 is the first DSP from Analog Devices to use the new
TigerSHARC™ static superscalar architecture. The TigerSHARC targets
telecommunications infrastructure equipment with a new level of integration and
the unique ability to process 8-, 16-, 32-bit fixed and floating-point data types on a
single chip. Each of these data types is critical to the next generation of
telecommunications protocols currently under development, including IMT-2000
(also known as 3G wireless) and xDSL (digital subscriber line). Unlike any other
DSP, the ADSP-TS001 has the unique ability to accelerate processing speed based
on the data type. Moreover, the chip delivers the highest performance floating-point
processing.
In telecommunications infrastructure equipment, voice coder and channel coder
protocols are developed around 16-bit data types. To improve signal quality, many
telecom applications employ line equalization and echo cancellation techniques that
boost overall signal quality and system performance. These algorithms benefit from
the added precision of 32-bit and floating-point data processing. The 8-bit native
support is well suited to the commonly used Viterbi channel decoder algorithm, as
well as image processing where it is more straightforward and cost-effective to
represent red, green, and blue components of the signal with 8-bit data types. Many
of these applications require high levels of performance and may require algorithms
to be executed consecutively or even concurrently. The end application determines
the exact requirements. The flexibility of the TigerSHARC architecture enables the
software engineer to match the application precision requirements without any loss
of system performance. In the TigerSHARC, performance is traded directly against
numerical precision.
The TigerSHARC architecture uses key elements from multiple microprocessor
types – RISC (reduced instruction set computer), VLIW (very long instruction word),
and DSP in order to provide the highest performance digital signal processing
engine. The new architecture leverages existing DSP product attributes such as fast
and deterministic execution cycles, highly responsive interrupts, and an excellent
peripheral interface to support large core computation rates and a high data rate
7.36
DSP HARDWARE
I/O. To achieve excellent core performance, RISC-like features such as load/store
operations, a deeply pipelined sequencer with branch prediction, and large
interlocked register files are introduced. Additionally, the VLIW (very long
instruction word) attributes offer more efficient use of code space, especially for
control code.
TigerSHARC®: ANALOG DEVICES’ NEW
STATIC SUPERSCALER DSP ARCHITECTURE
RISC
VLIW
• Load / store architecture
• Deeply pipelined for high
clock rates
• Branch prediction
• Large interlocked register
file
DSP
• Instruction level parallelism
determined prior to run time in
multi-instruction lines
• Determinism and real-time execution
• Fast and responsive interrupt system
• I/O and internal memory capable of
sustaining core rates
• Fast multiply accumulates, HW support for
circular buffers, bit reverse, zero overhead
looping
Figure 7.40
TigerSHARC KEY ARCHITECTURAL FEATURES
Core
■ 1200 MMACs/s @ 150 MHz -- 16-Bit Fixed Point
■ 300 MMACs/s @150 MHz -- 32-Bit Floating Point
■ 900 MFLOPS -- 32-Bit Floating Point
Memory
■ 6 Mbits of on-chip SRAM organized in a unified memory map as
opposed to the traditional Harvard architecture.
■
■
■
■
■
■
I/O, Peripherals, & Package
600 Mbytes/s transfer rate through external bus.
600 Mbytes/s aggregate transfer rate through 4 Link Ports
Glueless multiprocessor cluster support for up to 8 ADSP-TS001s
4 General Purpose I/O Ports
SDRAM Controller
360 Ball, SBGA Package 35×35mm
Figure 7.41
7.37
DSP HARDWARE
Finally, to supply all functional blocks with instructions, clever management of the
instruction word is necessary. Specifically, multiple instructions must be dispatched
to processing units simultaneously, and functional parallelism must be calculated
prior to runtime.
By incorporating the best of all worlds, the TigerSHARC architecture will provide a
state of the art platform for the most demanding signal processing applications.
The TigerSHARC core shown in Figure 7.42 consists of multiple functional blocks:
computation blocks, memory, integer ALUs, and a sequencer. There are two
computational blocks (X and Y) in the TigerSHARC architecture, each containing a
multiplier, ALU, and 64-bit shifter. With the resources in these blocks, it is possible
to execute eight 40-bit MACs on 16-bit data, two 40-bit MACs on 16-bit complex
data, or two 80-bit MACs on 32-bit data – all in a single cycle. TigerSHARC is a
register-based load/store architecture, where each computational block has access to
a fully orthogonal 32-word register file.
ADSP-TS001 TigerSHARC® ARCHITECTURE
SEQUENCER
J ALU
0
128-entry
BTB
0
J-RF
K-RF
31
31
128
128
K ALU
128
EXTERNAL
PORT
128
MD2
128
MD1
128
DMA
MD0
128
128
Comp Block X
ALU
0
R
F
31
128
128
128
128
Comp Block Y
ALU
0
R
F
Mult
Shift
PERIPHERALS
128
31
Mult
M0
M1
2Mbit 2Mbit
M2
2Mbit
Shift
ON-CHIP SRAM
Figure 7.42
7.38
DSP HARDWARE
The TigerSHARC DSP features a short-vector memory architecture organized in
three 128-bit wide banks. Quad, long, and normal word accesses move data from the
memory banks to the register files for operations. In a given cycle, four 32-bit
instruction words can be fetched, and 256 bits of data can be loaded to the register
files or stored into memory. Data in 8-, 16-, or 32-bit words can be stored in
contiguous, packed memory. Internal and external memories are organized in a
unified memory map which leaves specific partitioning to the programmer. The
internal memory bandwidth for data and instructions is 7.2 Gbytes/second when
operating on a 150MHz clock.
Two integer ALUs are available for data addressing and pointer updates. They
support circular buffering and bit reversal, and each has its own 32-word register
file. More than simple data address generations units, both integer ALUs support
general purpose integer computations. The general purpose nature of the integer
ALUs benefits the compiler efficiency and increases programming flexibility.
The TigerSHARC architecture is designated static superscalar, as it executes up to
four 32-bit instructions per clock cycle, and the programmer has the flexibility to
issue individual instructions to each of the computational units. The sequencer
supports predicted execution, where any individual instruction executes according
to the result of a previously defined condition. The same instruction can be executed
by the two computation blocks concurrently using different data values (this is
called SIMD – single-instruction multiple-data operation).
The TigerSHARC architecture enables native operation using 8-, 16-, or 32-bit data
values. The overall processor performance increases as the level of data precision
decreases.
The inclusion of a branch target buffer (BTB) and static branch prediction logic
eliminates the programming task of filling the instruction pipeline after branch
instructions. If seen before, the branch is taken in a single cycle.
Three internal 128-bit wide busses ensure a large data bandwidth between internal
functional blocks and external peripherals. The three-bus structure matches typical
mathematical instructions which require two inputs and compute one output. The
programming model is orthogonal and provides for deterministic interrupts.
The TigerSHARC architecture is free of hardware modes. This eliminates wasted
cycles and simplifies compiler operation. The instruction set directly supports all
DSP, image, and video processing arithmetic types including signed, unsigned,
fractional, and integer data types. There is optional saturation (clipping) arithmetic
for all cases.
At 150MHz, the ADSP-TS001 offers the highest integer and floating point
performance of any SHARC product. Additionally, at 6 Mbits of on-chip SRAM,
Analog Devices has increased its level of memory integration by 50% over previous
SHARC family members. The migration to smaller process geometries will enable
ADI to increase clock frequencies and integrate additional memory for future
product derivatives.
7.39
DSP HARDWARE
TigerSHARC KEY FEATURES
„ Execution of 1 to 4 32-Bit Instructions Per Clock Cycle
„ Single-Instruction Multiple Data (SIMD) Operations
Supported by Two Computation Blocks
„ Multiple Data Type Computation Blocks
◆ Each With Register File, MAC, ALU, Shifter
◆ 32/40-Bit Floating or 32-Bit Fixed Point Operations (6
Per Cycle)
◆ 16-Bit (24 Per Cycle) or 8-Bit (32 Per Cycle)
Operations
„ Static Branch Prediction Mechanism, with 128-Entry
Branch Target Buffer (BTB)
„ Internal Bandwidth of 7.2 Gbytes/second
„ Simple and Fully Interruptible Programming Model
Figure 7.43
The ADSP-TS001 reduces total material costs by integrating multiple I/O and
peripheral functions that reduce or eliminate the need for external glue logic and
support chips. Specifically, the ADSP-TS001 at 150MHz integrates four glueless
link ports with an aggregate transfer rate of 600 Mbytes/s, glueless multiprocessor
cluster interface support for up to 8 ADSP-TS001s, an SDRAM controller, and a
JTAG controller. This unprecedented functionality is packaged in a 35 by 35mm 360
ball SBGA package.
Typical computation rates and coding details of the TigerSHARC are shown in
Figure 7.44. Four 32-bit instructions are executed in parallel forming one 128-bit
instruction line. The entire instruction line is executed in one cycle. This example
assembly code is for a single line and is performing the following:
xR3:0=Q[j0+=4];//
load 4 registers (xR0,xR1,xR2,xR3) in the X register file from
memory
yR3:0=Q[k0+=4];//
load 4 registers in the Y register file from memory
FR5=R4*R4;
multiply 2 32-bit floats in X computational block and
2 more in Y (2 multiplies)
//
FR9:8=R6+/-R7;;//
7.40
add and subtract in both X and Y computational blocks
(4 ALU operations)
DSP HARDWARE
A single semicolon separates each 32-bit instruction, and a double semicolon
indicates the end of an instruction line. This particular example shows the syntax
for 32-bit floating point multiplies and ALU operations. Parallel 16-bit operands can
easily be specified by using the “S” for “short” prefix instead of the “F” for the “float”
prefix. J0 and K0 are IALU registers being used as indirect address pointers for the
memory reads.
TigerSHARC PEAK COMPUTATION RATES
■ 4 instructions per cycle accomplishes:
Ö
24 16-bit ops , or 6 32-bit ops
Ö 8 16-bit MACs, or 2 32-bit MACs
■ As well as 256-bit data moves, and 2 address
calculations
8 16-bit
or
2 32-bit
2 loads carrying 256 bits
multiplications
2 address calculations
in SIMD
xR3:0=Q[j0+=4];
yR3:0=Q[k0+=4];
FR5=R4*R4;
16 16-bit
or
4 32-bit
ALU ops
in SIMD
FR9:8=R6+/-R7;;
Figure 7.44
DSP programmers demand and require the capability to program in both high level
languages and low level assembly language. The determination of programming
language is dependent upon a number of factors including speed performance,
memory size, and time to market considerations. Ultimately, however, the whole
DSP product should incorporate features that enable user friendly coding in both
high level and low level languages. The TigerSHARC architecture does indeed meet
these requirements.
Specifically, the TigerSHARC core includes 128 ea. 32-bit general purpose registers.
This large number of registers allows C compilers sufficient flexibility to capitalize
on the full potential performance of the architecture. In order to ensure data
integrity, all registers are completely interlocked meaning that the programmer
does not have to be cognizant of architecture delays. The hardware ensures valid
data is used in computations. Additionally, all registers can be accessed via all
addressing modes (orthogonal) and a deterministic delay (2 clock cycles) is achieved
for all computational instructions. Lastly, the TigerSHARC architecture includes a
Branch Target Buffer which holds the effective address of the last 128 branches or
jumps. This feature alleviates the programming task of filling the instruction
pipeline after branch instructions. If seen before, the architecture jumps to the next
instruction in a single clock cycle.
7.41
DSP HARDWARE
ARCHITECTURAL FEATURES FOR HIGH-LEVEL
LANGUAGE SUPPORT
■ 128 General Purpose Registers
■ All Registers Fully Interlocked
■ General Purpose Integer ALUs for Addressing
■ Branch Prediction
■ No Hardware Modes
■ Orthogonal Addressing Modes
■ Assembly Language Support
Figure 7.45
Figure 7.46 depicts one possible configuration of a TigerSHARC design in a
multiprocessing implementation. Up to 8 ADSP-TS001 processors can communicate
directly via the high speed 64-bit wide external bus interface. In this type of
communication, a commonly used master-slave protocol is implemented which
enables any two processors to communicate directly at any one time.
In addition to the primary external bus, a limitless number of processors can be
connected via the ADSP-TS001 link ports. While offering more flexibility, link port
connectivity provides lower per port bandwidth than the primary external bus
interface. Again, all data transfers via the link ports are managed by a dedicated
I/O processor and require no CPU intervention.
To summarize, the data I/O bandwidth of the link port (600MBytes/s) and external
port (600MBytes/s) can be aggregated yielding an overall individual processor data
bandwidth of 1200 Mbytes/s with 150MHz clock operation. Additionally, both the
link port interface and the multiprocessor cluster interface are both completely
glueless.
The ADSP-TS001 is the first member of a planned family of TigerSHARC-based
products. Specifically, future members of the TigerSHARC family will contain
optimized mixes of memory and peripherals to meet the requirements of specific
target markets. These markets include third generation cellular base stations and
VOIP (Voice Over the Internet Protocol) servers/concentrators. Additionally, process
and design improvements will double the baseline performance of the general
purpose TigerSHARC family members.
7.42
DSP HARDWARE
MULTIPROCESSING COMMUNICATION VIA
LINK PORTS AND CLUSTER BUS
ADSP-TS001
TigerSHARC
SRAM
ADSP-TS001
TigerSHARC
DRAM
SDRAM
Link Port
connections
ADSP-TS001
TigerSHARC
I/O
ADSP-TS001
TigerSHARC
Host Interface
Bus connections
Figure 7.46
TigerSHARC ROADMAP
250 MHz
2 Mbit/sec
channel on a chip
150 MHz
Highest Density
Voice Over
The Internet
ADSP-TS001
2.4 Billion
MACs per
Second
Figure 7.47
7.43
DSP HARDWARE
Comparing DSPs based solely on MIPS, MOPS, or MFLOPS does not tell the entire
performance story. It is much more useful to compare DSPs based on their
performance with respect to specific algorithms. The FFT and the FIR filter are
popular benchmarks as well as the IIR biquad filter, matrix multiplications,
division, and square root.
Figure 7.48 shows the benchmark performance of the ADSP-TS001 TigerSHARC
operating on 16-bit fixed point data. Figure 7.49 shows its benchmark performance
operating on 32-bit floating point data.
ADSP-TS001 TigerSHARC BENCHMARKS @ 150MHz
16-BIT PERFORMANCE
■ 16-Bit performance -- 1200 MMACs/s peak performance
Execution
Time
Cycles to
Execute
256 Point Complex FFT (Radix 2)
7.3 µs
1100
50 Tap FIR on 1024 inputs
48 µs
7200
Single FIR MAC
0.93 ns
0.14
Single Complex FIR MAC
3.80 ns
0.57
6.7 ns
1.0
Algorithm
Single FFT Butterfly
Figure 7.48
ADSP-TS001 TigerSHARC BENCHMARKS @ 150MHz
32-BIT PERFORMANCE
■ 32-Bit performance -- 300 MMACs/s peak performance
Execution
Time
Algorithm
69 µs
10300
50 Tap FIR on 1024 Input
184 µs
27500
Single FIR MAC
3.7 ns
0.55
Single FFT Butterfly
13.3 ns
2.0
Single Complex FIR MAC
13.3 ns
2.0
20 ns
3.0
33.3 ns
5.0
3.3 ns
0.5
1024 Point Complex FFT (Radix 2)
Divide
Square Root
Viterbi Decode(per Add/Compare/Select)
Figure 7.49
7.44
Cycles to
Execute
DSP HARDWARE
DSP EVALUATION AND DEVELOPMENT TOOLS
The availability of a complete set of hardware and software development tools is
essential to any DSP-based design. A typical DSP system design cycle is described
below.
The first step in the process is to describe the system architecture. This would
include such things as the type of processor, the peripherals (external memory,
codecs, host processor, links), the configuration, etc. This information is placed in a
file known as the Link Descriptor File (or LDF).
The next step in the process is to generate the actual DSP code. This can be done
using a higher level language (usually C or C++), the DSP assembly language, or a
combination of both. DSP code developed in C must be compiled and assembled in
order to generate the assembly language code. While programming in C is easier,
the resulting assembly code which results after compiling is not as efficient as if the
coding had been done in assembly language originally. For this reason, many DSP
programmers do most of the programming in C, but use assembly language for the
critical loops in the program. The Analog Devices’ DSP assembly language is based
on algebraic syntax and is relatively easy to use directly. The linker then generates
an executable file.
The software must then be debugged using the software simulator in conjunction
with an evaluation board such as the EZ-LAB evaluation board or perhaps a thirdparty card which plugs into a slot in the PC.
After the software is debugged using the evaluation board, it must be tested with
the actual system target board (this is the board that you design with the DSP in
your system). An in-circuit emulator, such as the EZ-ICE, interfaces with the target
board, usually via a PCI or a JTAG interface port and connector.
The final step in the process is to generate the code required for booting the system
using the prom splitter.
A summary of the tools available from Analog Devices is shown in Figure 7.50. Each
one will be discussed in detail.
EZ-KIT Lites are basically DSP starter kit evaluation boards. In addition to the
processor itself, these boards contain an ADC and a DAC (codec) which interfaces to
the DSP over the DSP serial port. All necessary analog and digital support circuitry
is contained on the boards. The options on the board are controlled over an RS-232
port connection to a PC as well as jumpers on the board. Windows 95/98/NT
compatible software is supplied with the board. The software includes limited code
generation tools including a limited feature set compiler, assembler, linker, prom
splitter (loader), and Visual DSP debugger. Application examples such as DTMF
generator, echo cancellation, FFT, simple digital filters, etc., are included as part of
the software. The EZ-KIT Lite boards are primarily starter kit evaluation systems
(and “lite” on the wallet!).
7.45
DSP HARDWARE
ADI DSP DEVELOPMENT TOOLS
■ EZ-KIT Lite Evaluation Boards
■ Emulators
■ Integrated Development Environment (IDE) Software, VisualDSP®
and VisualDSP++

■
■
■
■
■
◆ Assembler, Linker, PROM Splitter, HIP Splitter, Simulator,
Compiler, Debugger
Extensive Algorithm Libraries
Factory, Field, and WWW Support
Seminars
ADI and Third Party DSP Workshops
ADI DSP Collaborative™ Third Party Support
Figure 7.50
EZ-KIT LITE’S™ FOR ANALOG DEVICES’ DSPS
■ The EZ-KIT Lite
 is a stand-alone (desktop) system that connects
to a PC running on Windows
■ The EZ-KIT Lites provide:
◆ A cost effective method for initial evaluation of the capabilities
of ADSP-series DSPs.
◆ A powerful development environment for a variety of general
purpose applications.
■ Target market:
◆ First time DSP users
◆ First time ADI DSP users
◆ Existing ADI DSP users implementing new designs
◆ Existing ADI DSP users upgrading to faster devices for current
designs
Figure 7.51
7.46
DSP HARDWARE
ADSP-2189M EZ-KIT LITE™
■ Hardware Features
◆
◆
◆
◆
◆
◆
◆
◆
ADSP-2189M 75 Mips processor
AD73322L Stereo codec
DSP-programmable CODEC gain
2 Mbit or greater boot protected Flash EPROM.
RS-232 PC to EZ-Kit Lite interface
Selectable Host vs. Full Memory mode implemented via dip switch
ADSP-218x EZ-ICE emulator port connector
Expansion connector includes all signal I/O plus 5V, 3.3V, 2.5V, and
GND connections
◆ LED indicators for master power, RS-232 interface, and one PF I/O
■ Software Features
◆ Windows 95/98/NT-4.0 PC host support
◆ VisualDSP: Limited feature set compiler, assembler, linker, prom
splitter (loader), VisualDSP debugger interface
◆ Application Examples: DTMF Generator, Echo Cancellation, FFT,
etc.(similar to 2181 EZ-KIT Lite)
◆ Email Support
Figure 7.52
ADSP-21160M EZ-KIT LITE™
■ Hardware Features
◆
◆
◆
◆
◆
◆
◆
◆
ADSP-21160M SHARC processor
AD1881 16-bit Stereo AC’97 SoundMAX Codec
EPROM flash memory (2 Mbit)
JTAG header
Support for ADSP-2116x family of processors
64K x 64 bit SBSRAM
Enhanced parallel port
CE compliant
■ Software Features
◆ Support for Win95, Win98 and WinNT
◆ Evaluation suite of VisualDSP++ : compiler, assembler,
linker, prom splitter (loader), VisualDSP debugger
interface. VisualDSP limited to use with EZ-KIT Lite
hardware
◆ DEMONSTRATIONS: DFT.dxe, BP.dxe, Pluck.dxe,
Primes.dxe, Tt.dxe
Figure 7.53
7.47
DSP HARDWARE
ADSP-21065L EZ-KIT LITE™
■ Hardware Features
◆ ADSP-21065L SHARC DSP running at 60MHz
◆ Full Duplex, 16-Bit Audio Codec
◆ RS-232 Interface with UART
◆ JTAG Emulation Connector
◆ Expansion via MAFE+ Connector
■ Software Features
◆ Support for Win95, Win98 and WinNT
◆ Evaluation suite of VisualDSP++ : compiler, assembler, linker,
prom splitter (loader), VisualDSP debugger interface.
VisualDSP limited to use with EZ-KIT Lite hardware
◆ Demonstrations: Fast Fourier Transform (FFT), Discrete
Fourier Transform (DFT), Band Pass Filter, Pluck String
Themes, Talk Through
Figure 7.54
The final step in the DSP system development is the debugging of the actual
system, or “target” board. The Analog Devices’ in-circuit emulator, EZ-ICE,
interfaces with a connector on the target board for use in final system hardware and
software debugging. Examples are shown in Figures 7.55 through 7.58. Figure 7.56
shows the Apex-ICE which interfaces to the target board via a JTAG connector
which in turn interfaces to the SHARC DSP. A USB port connector is used to
interface the emulator to a PC. Other in-circuit emulators are available which
interface to ISA, PCI, RS232, and Ethernet ports.
7.48
DSP HARDWARE
EZ-ICE® FOR THE ADSP-218x DSP FAMILY
■ Serial port interface, printed circuit board and 14-pin header
■ Controls equipment for testing, observing, and debugging a
target system
■ 6 foot cable
■ Hardware switch to accommodate
of 2.5V, 3.3V, and 5V
■ Shielded enclosure to cover bare
circuit board
■ Performance increase via faster
data transfer
Figure 7.55
APEX-ICE™ USB EMULATOR
■ Universal Serial Bus (USB)based emulator for Analog
Devices JTAG DSPs
■ First portable solution for
Analog Devices JTAG DSPs
■ Small hand-held unit
■ Small diameter cable, 5
meters in length, for hard
to reach targets
■ Power provided externally
JTAG CONNECTOR
USB PORT
CONNECTOR
Figure 7.56
7.49
DSP HARDWARE
TREK-ICE™ ETHERNET EMULATOR
■ Network hosted mini-tower emulator with 10-Base-T port
■ Installs to a LAN as easy as a laser printer
■ Remote debugging between either PC or SUN workstation debug
host (client) and the target DSP system
■ Rugged high-speed 3V/5V pod
■ Flexible emulator cable (1.5m)
Figure 7.57
SUMMIT-ICE™ PCI EMULATOR
■ 32-bit PCI interface add-in card
■ Four inch, flexible shielded target board cable for easy access to
a 14-pin JTAG header
■ Embedded ICEPAC technology provides a rugged and reliable
solution
■ Remote 3V/5V JTAG Pod with
extended, shielded cable (1.5 m)
■ Windows 95 & NT PNP
Figure 7.58
7.50
DSP HARDWARE
VISUALDSP® AND VISUALDSP++
New development software for Analog Devices’ DSPs is written in easy to use
VisualDSP® and VisualDSP ++ which is Windows 95/98/NT compatible. VisualDSP
is a completely integrated development environment which uses an algebraic syntax
assembler and an optimized C compiler. Multiprocessor environments can be
simulated and debugged as well. VisualDSP++ provides C++ language support.
VisualDSP versions currently exist for the ADSP-218x and ADSP-219x families as
well as the SHARC family of DSPs.
A “test drive” CDROM is available with a limited license for evaluation purposes.
In addition to the tools and support functions described thus far, Analog Devices’
DSP Collaborative consists of over 80 companies who provide a range of products
and services to make the DSP design task easier. Over 30 companies support the
16-bit ADSP-21xx family, and over 50 companies support the SHARC DSP family. A
directory of the collaborative can be found at:
http://www.analog.com/industry/dsp/3rdparty/index.html
Further information about Analog Devices’ DSP tools can be found at:
http://www.analog.com/dsp/tools
SOFTWARE DEVELOPMENT ENVIRONMENT
■ VisualDSP® and VisualDSP++
◆ Debugger front-end
◆ Integrated development
environment (IDE)
◆ Algebraic syntax assembler
◆ Cycle-accurate instruction
level simulator
◆ Optimizing ANSI C compiler
with inline assembly
◆ Math, DSP, and C runtime libraries
◆ Sophisticated multiprocessor linker
◆ Intelligent loader
◆ ADSP-218x, 219x: Windows 95, 98, NT, 2000 Compatible
◆ SHARC: Windows 95, 98, NT, 2000 Compatible
Figure 7.59
7.51
DSP HARDWARE
VisualDSP® 7.0 FOR ADSP-218x AND ADSP-219x
■ Hosted on Windows 95, Windows 98, Windows NT 4.0
with SP3 or later
■ ADSP-219x simulator target supports the 219x core
■ ELF/DWARF toolset, including compiler with classical and
processor-specific optimizations
■ Tcl command line scripting language
◆ Support automated test of DSP system
■ 21xx Object Translator
■ Peripheral Code Wizard supporting 218x and 219x peripherals
Figure 7.60
VisualDSP++™ FOR SHARC® DSPs
■ Supports Windows 95, 98, NT, 2000
■ ELF/Dwarf-2 file format enables effective debug operation
■ Pre-processor for linker/assembler separate from compiler
■ Fast ICE stepping - 0.9 sec. per step
■ MultiProcessor (MP) support
◆ Synchronous run, step, and halt
■ Tcl command line scripting language
◆ Support automated test of DSP system
■ Statistical Profiling
■ C++ capabilities
■ Graphical Plotting
Figure 7.61
7.52
DSP HARDWARE
TigerSHARC® DEVELOPMENT TOOLS
■ The TigerSHARC architecture is supported by a robust set of
simulation, code generation, and debug tools that includes:
◆ VisualDSP Integrated Development Environment
● Simulator, Assembler, Loader, Debugger, and Compiler
● DSP & Math Libraries
◆ Emulators – All present ADI JTAG emulators support the
TigerSHARC DSP family
◆ EZ-KIT Lite
■ Successive VisualDSP revisions will continuously increase
functionality.
Figure 7.62
VisualDSP TEST DRIVE
■ The test drive is a 30-day evaluation of VisualDSP's full package. It
does not include a tutorial. The new test drive is a full version of
VisualDSP and contains pdf.'s of the VisualDSP manuals.
■ The customer is presented with the test drive CD. The customer
than proceeds to the Analog Devices DSP Tools website, clicks on
Test Drive Registration
www.analog.com/industry/dsp/tools/test_drive.html and registers
online. After they complete the registration process they will
receive a serial number immediately that will allow them to use the
test drive. The test drive will expire 30-days from the install and
they will not be able to register the test drive again.
■ The SHARC VisualDSP test drive is now available, SAP part #
VDSP-SHARC-PC-TEST.
■ The VisualDSP TigerSHARC test drive will be available
Summer 2000
■ The VisualDSP ADSP-218x/219x test drive will be available in
September 2000
Figure 7.63
7.53
DSP HARDWARE
ADI DSP COLLABORATIVE - WHAT IS IT?
■ Over 80 Companies that provide a wide range of products and
services to make your design challenge easier
■ Architecture Coverage
◆ Over 30 Companies support 16-bit, ADSP-21xx Family
◆ Over 50 Companies support SHARC® DSP Family
■ Over 400 Products from the Following Categories:
◆ Algorithms
◆ Emulators
◆ Real-Time Operating Systems
◆ Hardware Development Boards
◆ Debuggers
◆ Graphical S/W Programs
®
◆ MATLAB DSP Support
◆ Consulting Services
■ Focused Applications :
◆ Audio
◆ Motor/Motion Control
◆ Digital Radio
◆ Radar/Sonar
◆ Industrial Inspection & Control ◆ Telecom
◆ Medical Instrumentation/Imaging ◆ Video/Sound Processing
◆ Military/Aerospace
■ http://www.analog.com/industry/dsp/3rdparty/index.html
Figure 7.64
7.54
DSP HARDWARE
REFERENCES
1.
Steven W. Smith, The Scientist and Engineer’s Guide to Digital Signal
Processing, Second Edition, 1999, California Technical Publishing,
P.O. Box 502407, San Diego, CA 92150. Also available for free download at:
http://www.dspguide.com or http://www.analog.com/industry/dsp/dsp_book
2.
C. Britton Rorabaugh, DSP Primer, McGraw-Hill, 1999.
3.
Richard J. Higgins, Digital Signal Processing in VLSI, Prentice-Hall,
1990.
4.
Ethan Bordeaux, Advanced DSP Performance Complicates Memory
Architecures in Wireless Designs, Wireless Systems Design, April 2000.
5.
DSP Designer’s Reference (DSP Solutions) CDROM, Analog Devices,
1999.
6.
DSP Navigators: Interactive Tutorials about Analog Devices’ DSP
Architectures (Available for ADSP-218x family and SHARC family):
http://www.analog.com/industry/dsp/training/index.html#Navigator
7.
General DSP Training and Workshops:
http://www.analog.com/industry/dsp/training
The following DSP Reference Manuals and documentation are available
for free download from: http://www.analog.com/industry/dsp/tech_docs.html
8.
ADSP-2100 Family Users Manual, 3rd Edition, Sept., 1995.
9.
ADSP-2100 Family EZ Tools Manual.
10.
ADSP-2100 EZ-KIT Lite Reference Manual.
11.
Using the ADSP-2100 Family, Vol. 1, Vol. 2.
12.
ADSP-2106x SHARC User’s Manual, 2nd Edition, July, 1996.
13.
ADSP-2106x SHARC EZ-KIT Lite Manual.
14.
ADSP-21065L SHARC User’s Manual, Sept. 1, 1998.
15.
ADSP-21065L SHARC EZ-LAB User’s Manual.
16.
ADSP-21160 SHARC DSP Hardware Reference.
7.55
DSP HARDWARE
7.56
INTERFACING TO DSPS
SECTION 8
INTERFACING TO DSPs
■
Parallel Interfacing to DSP Processors: Reading Data
From Memory-Mapped Peripheral ADCs
■
Parallel Interfacing to DSP Processors: Writing Data to
Memory-Mapped DACs
■
Serial Interfacing to DSP Processors
■
Interfacing I/O Ports, Analog Front Ends, and Codecs
to DSPs
■
DSP System Interface
8.a
INTERFACING TO DSPS
8.b
INTERFACING TO DSPS
SECTION 8
INTERFACING TO DSPs
Walt Kester, Dan King
INTRODUCTION
As the technology in the rapidly growing field of mixed signal processing evolves,
more highly integrated DSP products are being introduced (such as the ADSP21ESP202) which contain on-chip ADCs and DACs as well as the DSP, thereby
eliminating most component-level interface problems. Stand-alone ADCs and DACs
are now available with interfaces especially designed for DSP chips, thereby
minimizing or eliminating external interface support or glue logic. High
performance sigma-delta ADCs and DACs are currently available in the same
package (called a CODEC or COder/DECcoder) such as the AD73311 and AD73322.
These products are also designed to require minimum glue logic when interfacing to
the most common DSP chips. This section discusses the various data transfer and
timing issues associated with the various interfaces.
PARALLEL INTERFACING TO DSP PROCESSORS:
READING DATA FROM MEMORY-MAPPED PERIPHERAL
ADCS
Interfacing an ADC or a DAC to a fast DSP parallel requires an understanding of
how the DSP processor reads data from a memory-mapped peripheral (the ADC)
and how the DSP processor writes data to a memory-mapped peripheral (the DAC).
We will first consider some general timing requirements for reading and writing
data. It should be noted that the same concepts presented here regarding ADCs and
DACs apply equally when reading and writing from/to external memory.
A block diagram of a typical parallel DSP interface to an external ADC is shown in
Figure 8.1. This diagram has been greatly simplified to show only those signals
associated with reading data from an external memory-mapped peripheral device.
The timing diagram for the ADSP-21xx read-cycle is shown in Figure 8.2.
In this example it is assumed that the ADC is sampling at a continuous rate which
is controlled by the external sampling clock, not the internal DSP clock. Using a
separate clock for the ADC is the preferred method, since the DSP clock may be
noisy and introduce jitter in the ADC sampling process, thereby increasing the noise
level.
Assertion of the sampling clock at the ADC convert start input initiates the
conversion process (step 1). The leading (or trailing) edge of this pulse causes the
internal ADC sample-and-hold to switch from the sampling mode to the hold mode
so that the conversion process can take place. When the conversion is complete, the
conversion complete output of the ADC is asserted (step 2). The read process thus
begins when this signal is applied to the processor interrupt request line ( IRQ ) of
the DSP. The processor then places the address of the peripheral initiating the
8.1
INTERFACING TO DSPS
interrupt request (the ADC) on the memory address bus (A0 - A13) (step 3). At the
same time, the processor asserts a memory select line ( DMS is shown here) (step 4).
The two internal address buses of the ADSP-21xx (program memory address bus
and data memory address bus) share a single external address bus, and the two
internal data buses (program memory data bus and data memory data bus) share a
single external data bus. The boot memory select ( BMS ), data memory select
( DMS ), program memory select ( PMS ) and input/output memory select ( IOMS )
signals indicate which memory space the external buses are being used for. These
signals are typically used to enable an external address decoder as shown in Figure
8.1. The output of the address decoder drives the chip select input of the peripheral
device (step 5).
The memory read ( RD ) is asserted tASR ns after the DMS line is asserted (step 6).
The sum of the address decode delay plus the peripheral chip select setup time
should be less than tASR in order to take full advantage of the RD low-time. The RD
line remains low for tRP ns. The memory read signal is used to enable the three-state
parallel data outputs of the peripheral device (step 7). The RD line is connected to
the appropriate pin on the peripheral device usually called output enable or read.
The rising edge of the RD signal is used to clock the data on the data bus into the
DSP processor (step 8). After the rising edge of the RD signal, the data on the data
bus must remain valid for tRDH ns, the data hold time. In the case of most members
of the ADSP-21xx family, this specification value is 0ns.
The key timing requirements for the peripheral device are shown in Figure 8.3.
Values are given for the ADSP-2189M DSP operating at 75MHz.
ADC TO ADSP-21xx-FAMILY PARALLEL INTERFACE
ADSP-21xx
3
SAMPLING
CLOCK
MEMORY A0- A13
ADDRESS BUS
DATA MEMORY
DMS
SELECT
PROCESSOR
INTERRUPT
REQUEST
4
ADC
CONVERT START
5
ADDRESS
DECODE
CHIP SELECT
2
CONVERSION
COMPLETE
IRQ
6
MEMORY READ
RD
MEMORY DATA
BUS
D
8
OUTPUT ENABLE
7
OUTPUT DATA
Figure 8.1
8.2
1
INTERFACING TO DSPS
ADSP-21xx FAMILY MEMORY READ TIMING
DSP
CLKOUT
A0- A13
DMS, PMS,
IOMS, CMS
RD
tASR
tCRD
tRP
tRDA
tRWR
D
tRDH
tAA
tRDD
Figure 8.2
PARALLEL PERIPHERAL DEVICE READ INTERFACE
KEY REQUIREMENTS
■ Peripheral Device Data Outputs Must Be Three-State Compatible
■ Address Decode Delay + Peripheral Chip Select Setup Time Must Be
Less Than Address and Memory Select Setup Time tASR (0.325ns min
for ADSP-2189M)
■ For Zero Wait-State Access, the Time from a Negative-Going Edge of
Read Signal (RD) to Output Data Valid Must be Less than tRDD (1.65ns
max for ADSP-2189M Operating at 75MHz) or Software Wait-States
Must be Added, or Processor Clock Frequency Reduced
■ Output Data from Peripheral must Remain Valid for tRDH from the
Rising Edge of Read Signal (RD) (0ns for ADSP-2189M)
■ Peripheral Device Must Accept Minimum Output Enable Pulse Width of
tRP (3.65ns for ADSP-2189M Operating at 75MHz) or Software WaitStates Must be Added, or Processor Clock Frequency Reduced
Figure 8.3
8.3
INTERFACING TO DSPS
The DSP tRDD specification determines the peripheral device data access time
requirement. In the case of the ADSP-2189M, tRDD = 1.65ns minimum at 75MHz. If
the access time of the peripheral is greater than this, wait states must be added or
the processor speed reduced. This is a relatively common situation when interfacing
external memory or ADCs to fast DSPs. The relationship between these timing
parameters for the ADSP-2189M is given by the equations shown in Figure 8.4.
Note that these specifications are dependent on the DSP clock frequency.
ADSP-2189M PARALLEL READ TIMING AT 75MHz
■ tCK = Processor Clock Period (13.3ns)
■ tASR = Address and Memory Select Setup Before Read Low
= 0.25tCK – 3ns Minimum
■ tRDD = Read Low to Data Valid = 0.5tCK – 5ns + # wait states
× tCK Maximum
■ tRDH = Data Hold from Read High = 0ns Minimum
■ tRP = Read Pulse Width = 0.5tCK – 3ns + # wait states × tCK
Minimum
Figure 8.4
The ADSP-2189M can easily be interfaced to slow peripheral devices using its
programmable wait state generation capability. Three registers control wait state
generation for boot, program, data and I/O memory spaces. You can specify 0 to 15
wait states for each parallel memory interface. Each wait state added increases the
allowable external data memory access time by an amount equal to the processor
clock period (13.3ns for the ADSP-2189M operating at 75MHz). In this example, the
data memory address, DMS , and RD lines are all held stable for an additional
amount of time equal to the duration of the wait states.
The AD7854/AD7854L is a 12 bit, 200/100kSPS ADC which operates in the parallel
mode. It operates on a single +3V to +5.5V supply and dissipates only 5.5mW (+3V
supply, AD7854L). An automatic power-down after conversion feature reduces this
to 650µW.
8.4
INTERFACING TO DSPS
A functional block diagram of the AD7854/AD7854L is shown in Figure 8.5. The
AD7854/AD7854L uses a successive approximation architecture which is based on a
charge redistribution (switched capacitor) DAC. A calibration mode removes offset
and gain errors. The key interface timing specifications for the AD7854/AD7854L
and the ADSP-2189M are compared in Figure 8.6. Specifications for the ADSP2189M are given for a clock frequency of 75MHz.
Examining the timing specifications shown in Figure 8.6 reveals that for the timing
between the devices to be compatible, 5 software wait-states must be programmed
into the ADSP-2189M. This increases tRDD to 68.15ns which is greater than the
data access time of the AD7854/AD7854L (t8 = 50ns max.). The read pulse, tRP, is
likewise increased to 70.15ns which meets the ADC’s read pulse width requirement
(t7 = 70ns min.). Unless the memory-mapped peripheral has an extremely short
access time, wait states are generally required, whether interfacing to ADCs, DACs,
or external memory.
AD7854/AD7854L, +3V SINGLE SUPPLY, 12-BIT,
200/100kSPS PARALLEL OUTPUT ADC
AVDD
AIN(+)
AGND
AD7854/AD7854L
T/H
DVDD
AIN(–)
2.5 V
REFERENCE
REFIN /
REFOUT
CREF1
DGND
COMP
BUF
CHARGE
REDISTRIBUTION
DAC
CLKIN
SAR + ADC
CONTROL
CONVST
BUSY
CREF2
CALIBRATION
MEMORY
AND CONTROLLER
PARALLEL INTERFACE/CONTROL REGISTER
DB11 - DB0
CS
RD
WR
HBEN
Figure 8.5
8.5
INTERFACING TO DSPS
ADSP-2189M AND AD7854/AD7854L PARALLEL READ
INTERFACE TIMING SPECIFICATION COMPARISON
ADSP-2189M Processor (75MHz)
AD7854/AD7854L ADC
tASR (Data Address, Memory
Select Setup Time Before RD Low)
= 0.325ns min
t5 (CS to RD Setup Time =
0ns min (Must Add Address
Decode Time to this Value)
tRP (RD Pulse Width) =
3.65ns + # wait states × 13.3ns min
= 70.15ns min
t7 (RD Pulse Width) = 70ns min
tRDD (RD Low to Data Valid) =
1.65ns + # wait states × 13.3ns min
= 68.15ns min
t8 (Data Access Time After RD) =
50ns max
tRDH (Data Hold from RD High) =
0ns min
t9 (Bus Relinquish Time After RD) =
5ns min / 40ns max
NOTES:
(1) Adding 5 wait-states to the ADSP-2189M increases tRP to 70.15ns
which is greater than t7 (70ns) and meets the t8 (50ns) requirement.
(2) t9 max (40ns) may cause bus contention if a write cycle immediately follows
the read cycle.
Figure 8.6
A simplified interface diagram for the two devices is shown in Figure 8.7. The
conversion complete signal from the AD7854/AD7854L corresponds to the BUSY
output pin. Notice that the configuration allows the DSP to write data to the
AD7854/AD7854L parallel interface control register. This is needed in order to set
various options in the AD7854/AD7854L and perform the calibration routines. In
normal operation, however, data is read from the AD7854/AD7854L as described
above. Writing to external parallel memory-mapped peripherals is discussed later in
this section.
Parallel interfaces between other DSP processors and external peripherals can be
designed in a similar manner by carefully examining the timing specifications for all
appropriate signals for each device. The data sheets for most ADCs contain
sufficient information in the application section to interface them to the DSPs.
8.6
INTERFACING TO DSPS
AD7854/AD7854L ADC PARALLEL INTERFACE
TO ADSP-2189M
SAMPLING
CLOCK
CONVST
ADSP-2189M
75MHz
AD7854/AD7854L
ADC
DMS
A0
CS
(LOW = READ DB11 - DB0)
HBEN
IRQ
BUSY
WR
WR
RD
RD
D23 - D8
Notes:
DATA
DB11 - DB0
5 Software Wait-States
HBEN and WR required for writing to ADC
Sampling clock may come from DSP
Figure 8.7
PARALLEL INTERFACING TO DSP PROCESSORS:
WRITING DATA TO MEMORY-MAPPED DACS
A simplified block diagram of a typical DSP interface to a parallel peripheral device
(such as a DAC is shown in Figure 8.8. The memory-write cycle timing diagram for
the ADSP-21xx-family is shown in Figure 8.9.
In most real-time applications, the DAC is operated continuously from a stable
sampling clock. Most DACs for these applications have double buffering: an input
latch to handle the asynchronous DSP interface, followed by a second latch (called
the DAC latch) which drives the DAC current switches. The DAC latch strobe is
derived from an external stable sampling clock. In addition to clocking the DAC
latch, the DAC latch strobe is also used to generate a processor interrupt to the DSP
which indicates that the DAC is ready for a new input data word.
8.7
INTERFACING TO DSPS
DAC TO ADSP-21xx-FAMILY PARALLEL INTERFACE
ADSP-21xx
SAMPLING
CLOCK
MEMORY A0- A13
ADDRESS BUS
3
DATA MEMORY
DMS
SELECT
PROCESSOR
INTERRUPT
REQUEST
MEMORY DATA
BUS
DAC LATCH
STROBE
5
ADDRESS
DECODE
CHIP SELECT
1
IRQ
7
4
MEMORY WRITE
DAC
1
2
INPUT LATCH
STROBE
WR
6
PARALLEL DATA
INPUT
D
Figure 8.8
ADSP-21xx FAMILY MEMORY WRITE TIMING
DSP
CLKOUT
A0- A13
DMS, PMS,
BMS, CMS
tWP
WR
tWRA
tAW
tWWR
tASW
tDH
tCWR
D
tWDE
Figure 8.9
8.8
tDW
tDDR
INTERFACING TO DSPS
The write process is thus initiated by the peripheral device asserting the DSP
interrupt request line indicating that the peripheral is ready to accept a new parallel
data word (step 1). The DSP then places the address of the peripheral device on the
address bus (step 2) and asserts a memory select line ( DMS is shown here) (step 3).
This causes the output of the address decoder to assert the chip select input to the
peripheral (step 5). The write ( WR ) output of the DSP is asserted tASW ns after the
negative-going edge of the DMS signal (step 4). The width of the WR pulse is tWP
ns. Data is placed on the data bus (D) and is valid tDW ns before the WR line goes
high (step 6). The positive-going transition of the WR line is used to clock the data
on the data bus (D) into the external parallel memory (step 7). The data on the data
bus remains valid for tDH ns after the positive-going edge of the WR signal.
The key timing requirements for writing to the peripheral device are shown in
Figure 8.10. The key specification is tWP , the write pulse width. All but the fastest
peripheral devices will require wait states to be added due to their longer data
access times. Figure 8.11 shows the key timing specifications for the ADSP-2189M.
Note that they are all related to the processor clock frequency.
PARALLEL PERIPHERAL DEVICES WRITE INTERFACE
KEY REQUIREMENTS
■ Address Decode Delay + Peripheral Chip Select Setup Time Must
Be Less Than Address and Memory Select Setup Time tASW
(0.325ns for ADSP-2189M Operating at 75MHz)
■ For Zero Wait-State Access, Input Data Setup Time Must be Less
Than tDW (2.65ns for ADSP-2189M Operating at 75MHz) or Software
Wait-States Must be Added, or Processor Clock Frequency
Reduced
■ Input Data Hold Time Must be Less Than tDH (2.325ns for ADSP2189M Operating at 75MHz)
■ Peripheral Device Must Accept Input Write Clock Pulse of Width
tWP (3.65ns min for ADSP-2189M Operating at 75MHz) or SoftwareWait States Must be Added, or Processor Clock Frequency
Reduced
Figure 8.10
8.9
INTERFACING TO DSPS
ADSP-2189M PARALLEL WRITE TIMING
■ tCK = Processor Clock Period (13.3ns)
■ tASW = Address and Memory Select Before WR Low
= 0.25tCK – 3ns Minimum
■ tDW = Data Setup Before WR High = 0.5tCK – 4ns + # Wait States × tCK
■ tDH = Data Hold After WR High = 0.25tCK – 1ns
■ tWP = WR Pulse Width = 0.5tCK – 3ns + # Wait States × tCK Minimum
Figure 8.11
The AD5340 is a 12 bit 100kSPS DAC which has a parallel data interface. It
operates on a single +2.5V to +5.5V supply and dissipates only 345µW (+3V supply).
A power-down mode further reduces the power to 0.24µW. The part incorporates an
on-chip output buffer which can drive the output to both supply rails. The AD5340
allows the choice of a buffered or unbuffered reference input. The device has a
power-on reset circuit that ensures that the DAC output powers on at 0V and
remains there until valid data is written to the part. A block diagram is shown in
Figure 8.12. The input is double buffered. The key interface timing specifications for
the two devices are compared in Figure 8.13. Specifications for the ADSP-2189M are
given for a clock frequency of 75MHz.
AD5340 12-BIT, 100kSPS PARALLEL INPUT DAC
VREF
POWER-ON
RESET
DB0
BUF
GAIN
CS
WR
INTERFACE LOGIC
DB11
INPUT
REGISTER
DAC
REGISTER
12-BIT
DAC
RESET
CLR
POWER-DOWN
LOGIC
LDAC
PD
Figure 8.12
8.10
VOUT
BUF
GND
INTERFACING TO DSPS
ADSP-2189M AND AD5340 PARALLEL WRITE INTERFACE
TIMING SPECIFICATIONS
ADSP-2189M PROCESSOR (75MHz)
AD5340 DAC
tASW (Address and Data Memory
Select Setup Before WR Low)
= 0.325ns min
t1 (CS to WR Setup Time)
= 0ns min
tWP (WR Pulse Width) =
3.65ns + # Wait States × 13.3ns min
= 30.25ns min
t3 (WR Pulse Width)
= 20ns min
tDW (Data Setup Before WR High) =
2.65ns + # Wait States × 13.3ns min
= 29.25ns min
t4 (Data Valid to WR Setup Time)
= 5ns min
tDH (Data Hold After WR High)
= 2.325ns min
t5 (Data Valid to WR Hold Time)
= 4.5ns min
NOTE: Adding 2 wait states to the ADSP-2189M increases tWP to 30.25ns
and tDW to 29.25ns which is greater than t3 (20ns) and t4 (5ns) respectively.
Figure 8.13
Examining the timing specifications shown in Figure 8.13 reveals that for the
timing between the devices to be compatible, two software wait states must be
programmed into the ADSP-2189M. This increases the width of WR to 30.25ns
which is greater than the minimum required AD5340 write pulse width (20ns). The
data setup time of 5ns for the AD5340 is also met by adding two wait states. A
simplified interface diagram for the two devices is shown in Figure 8.14.
Parallel interfaces with other DSP processors can be designed in a similar manner
by carefully examining the timing specifications for all appropriate signals for each
device.
8.11
INTERFACING TO DSPS
AD5340 DAC PARALLEL INTERFACE TO ADSP-2189M
SAMPLING
CLOCK
LDAC
AD5340
DAC
ADSP-2189M
75MHz
DMS
CS
IRQ
WR
WR
D
Notes:
DB0 - DB11
2 Software Wait-States
Sampling clock may come from DSP
Figure 8.14
SERIAL INTERFACING TO DSP PROCESSORS
DSP processors which have serial ports (such as the ADSP-21xx family) provide a
simple interface to peripheral ADCs and DACs. Use of the serial port eliminates the
need for using large parallel buses to connect the ADCs and DACs to the DSP. In
order to understand serial data transfer better, we will first examine the serial port
operation of the ADSP-21xx series.
A block diagram of one of the two serial ports of the ADSP-21xx is shown in Figure
8.15. The Transmit (TX) and Receive (RX) registers are identified by name in the
ADSP-21xx assembly language, and are not memory mapped.
8.12
INTERFACING TO DSPS
ADSP-21xx FAMILY SERIAL PORT BLOCK DIAGRAM
DMD BUS
16
16
16
TXn
TRANSMIT DATA
REGISTER
COMPANDING
HARDWARE
(µ-LAW or A-LAW)
RXn
RECEIVE DATA
REGISTER
16
16
TRANSMIT SHIFT
REGISTER
SERIAL
CONTROL
RECEIVE SHIFT
REGISTER
INTERNAL
SERIAL
CLOCK
GEN.
DT
TFS SCLK
Polarity of TFS and RFS
is software programmable
RFS
DR
Figure 8.15
ADSP-21xx FAMILY SERIAL PORT FEATURES
■ Separate Transmit and Receive Sections for Each Port
■ Double-Buffered Transmit and Receive Registers
■ Serial Clock Can be Internally or Externally Generated
■ Transmit and Receive Frame Sync Signals Can be Internally or
Externally Generated
■ Serial Data Words of 3 to 16 Bits Supported
■ Automatically Generated Processor Interrupts
■ Hardware Companding Requires no Software Overhead
Figure 8.16
8.13
INTERFACING TO DSPS
In the receiving portion of the serial port, the receive frame sync (RFS) signal
initiates reception. The serial receive data (DR) from the external device (ADC) is
transferred into the receive shift register one bit at a time. The negative-going edge
of the serial clock (SCLK) is used to clock the serial data from the external device
into the receive shift register. When a complete word has been received, it is written
to the receive data register (RX), and the receive interrupt for that serial port is
generated. The receive data register is then read by the processor.
Writing to the transmit data register readies the serial port for transmission. The
transmit frame sync (TFS) signal initiates transmission. The value in the transmit
data register (TX) is then written to the internal transmit shift register. The data in
the transmit shift register is sent to the peripheral device (DAC) one bit at a time,
and the positive-going edge of the serial clock (SCLK) is used to clock the serial
transmit data (DT) into the external device. When the first bit has been transferred,
the serial port generates the transmit interrupt. The transmit data register can then
be written with new data, even though the transmission of the previous data is not
complete.
In the normal framing mode, the frame sync signal (RFS or TFS) is checked at the
falling edge of SCLK. If the framing signal is asserted, data is available (transmit
mode) or latched (receive mode) on the next falling edge of SCLK. The framing
signal is not checked again until the word has been transmitted or received. In the
alternate framing mode, the framing signal is asserted in the same SCLK cycle as
the first bit of a word. The data bits are latched on the falling edge of SCLK, but the
framing signal is checked only on the first bit. Internally-generated framing signals
remain asserted for the length of the serial word. The alternate framing mode of the
serial port in the ADSP-21xx is normally used to receive data from ADCs and
transmit data to DACs.
The serial ports of the ADSP-21xx family are extremely versatile. The TFS, RFS, or
SCLK signals can be generated from the ADSP-21xx clock (master mode) or
generated externally (slave mode). The polarity of these signals can be reversed
with software, thereby allowing more interface flexibility. The port also contains µlaw and A-law companding hardware for voiceband telecommunications
applications.
SERIAL ADC TO DSP INTERFACE
A timing diagram of the ADSP-2189M serial port operating in the receive mode
(alternate framing) is shown in Figure 8.17. The first negative-going edge of the
SCLK to occur after the negative-going edge of the RFS , clocks the MSB data from
the ADC into the serial input latch. The process continues until all serial bits have
been transferred into the serial input latch. The key timing specifications of concern
are the serial data setup (tSCS) and hold times (tSCH) with respect to the negativegoing edge of the SCLK. In the case of the ADSP-2189M, these values are 4ns and
7ns, respectively . The latest generation ADCs with high speed serial clocks will
have no trouble meeting these specifications, even at the maximum serial data
transfer rate.
8.14
INTERFACING TO DSPS
ADSP-2189M SERIAL PORT RECEIVE TIMING
ADC
SCLK
RFS IN
tSCS
≥ 4ns
DR IN
tSCS
tSCH
≥ 7ns
MSB
BIT n
LSB
ALTERNATE FRAMING MODE, ADC IS MASTER
Figure 8.17
The AD7853/AD7853L is a 12 bit, 200/100kSPS ADC which operates on a single
+3V to +5.5V supply and dissipates only 4.5mW (+3V supply, AD7853L). After each
conversion, the device automatically powers down to 25µW. The AD7853/AD7853L
is based on a successive approximation architecture and uses a charge
redistribution (switched capacitor) DAC. A calibration feature removes gain and
offset errors. A block diagram of the device is shown in Figure 8.18.
The AD7853 operates on a 4MHz maximum external clock frequency. The AD7853L
operates on a 1.8MHz maximum external clock frequency. The timing diagram for
AD7853L is shown in Figure 8.19. The AD7853/AD7853L has modes which
configure the SYNC and SCLK as inputs or outputs. In the example shown here
they are generated by the AD7853L. The AD7853L serial clock operates at a
maximum frequency of 1.8MHz (556ns period). The data bits are valid 330ns after
the positive-going edges of SCLK. This allows a setup time of approximately 330ns
minimum before the negative-going edges of SCLK, easily meeting the ADSP2189M 4ns tSCS requirement. The hold-time after the negative-going edge of SCLK
is approximately 226ns, again easily meeting the ADSP-2189M 7ns tSCH timing
requirement. These simple calculations show that the data and RFS setup and hold
requirements of the ADSP-2189M are met with considerable margin.
8.15
INTERFACING TO DSPS
AD7853/AD7853L +3V SINGLE-SUPPLY 12-BIT
200/100kSPS SERIAL OUTPUT ADC
AVDD
AIN(+)
AGND AGND
DVDD
AD7853L
T/H
AIN(–)
2.5 V
REFERENCE
REFIN /
REFOUT
DGND
COMP
BUF
CHARGE
REDISTRIBUTION
DAC
CREF1
CLKIN
CONVST
SAR + ADC
CONTROL
CREF2
BUSY
CALIBRATION
MEMORY
AND CONTROLLER
CAL
SLEEP
SERIAL INTERFACE/CONTROL REGISTER
SM1
SM2
SYNC
DIN
DOUT
SCLK
POLARITY
Figure 8.18
AD7853L SERIAL ADC OUTPUT TIMING
+3V SUPPLY, SCLK = 1.8MHz
SYNC (O/P)
SCLK
(O/P)
1
5
6
16
330ns min
THREE-STATE
DB15
DOUT (O/P)
226ns
DB11
556ns
Figure 8.19
8.16
DB0
THREE-STATE
INTERFACING TO DSPS
Figure 8.20 shows the AD7853L interfaced to the ADSP-2189M connected in a mode
to transmit data from the ADC to the DSP (alternate/master mode). The
AD7853/AD7853L contains internal registers which can be accessed by writing from
the DSP to the ADC via the serial port. These registers are used to set various
modes in the AD7853/AD7853L as well as to initiate the calibration routines. These
connections are not shown in the diagram.
AD7853/AD7853L SERIAL ADC
INTERFACE TO ADSP-2189M
ADSP-2189M
75MHz
DSP
CLOCK
INPUT
4MHz / 1.8MHz max
CLKIN
SAMPLING
CLOCK
(OPTIONAL)
CONVST
SCLK
SCLK
RFS
SYNC
DR
DOUT
AD7853/
AD7853L
ADC
SERIAL
PORT
Figure 8.20
SERIAL DAC TO DSP INTERFACE
Interfacing serial input DACs to the serial ports of DSPs such as the ADSP-21xx
family is also relatively straightforward and similar to the previous discussion
regarding serial output ADCs. The details will not be repeated here, but a simple
interface example will be shown.
The AD5322 is a 12-bit, 100kSPS dual DAC with a serial input interface. It operates
on a single +2.5V to +5.5V supply, and a block diagram is shown in Figure 8.21.
Power dissipation on a +3V supply is 690µW. A power-down feature reduces this to
0.15µW. Total harmonic distortion is greater than 70dB below full scale for a 10kHz
output. The references for the two DACs are derived from two reference pins (one
per DAC). The reference inputs may be configured as buffered or unbuffered inputs.
The outputs of both DACs may be updated simultaneously using the asynchronous
LDAC input. The device contains a power-on reset circuit that ensures that the
DAC outputs power up to 0V and remain there until a valid write takes place to the
device.
8.17
INTERFACING TO DSPS
AD5322 12-BIT, 100kSPS DUAL DAC
VREFA
VDD
POWERON
RESET
AD5322
INPUT
REGISTER A
DAC
REGISTER A
STRING
DAC A
VOUTA
SCLK
INTERFACE LOGIC
SYNC
DIN
RESISTOR
NETWORK
POWER-DOWN
LOGIC
INPUT
REGISTER B
DAC
REGISTER B
STRING
DAC B
VOUTB
RESISTOR
NETWORK
LDAC
GND
VREFB
Figure 8.21
Data is normally input to the AD5322 via the SCLK, DIN, and SYNC pins from the
serial port of the DSP. When the SYNC signal goes low, the input shift register is
enabled. Data is transferred into the AD5322 on the falling edges of the following 16
clocks. A typical interface between the ADSP-2189M and the AD5322 is shown in
Figure 8.22. Notice that the clocks to the AD5322 are generated from the ADSP2189M clock. It is also possible to generate the SCLK and SYNC signals externally
to the AD5322 and use them to drive the ADSP-2189M. The serial interface of the
AD5322 is not fast enough to handle the ADSP-2189M maximum master clock
frequency. However, the serial interface clocks are programmable and can be set to
generate the proper timing for fast or slow DACs.
The input shift register in the AD5322 is 16-bits wide. The 16-bit word consists of
four control bits followed by 12 bits of DAC data. The first bit loaded determines
whether the data is for DAC A or DAC B. The second bit determines if the reference
input will be buffered or unbuffered. The next two bits control the operating modes
of the DAC (normal, power-down with 1kΩ to ground, power-down with 100kΩ to
ground, or power-down with a high impedance output).
8.18
INTERFACING TO DSPS
AD5322 DAC SERIAL INTERFACE TO ADSP-2189M
ADSP-2189M
75MHz
AD5322
DAC
SCLK
SCLK
TFS
SYNC
DIN
DT
SERIAL
PORT
Figure 8.22
INTERFACING I/O PORTS, ANALOG FRONT ENDS,
AND CODECS TO DSPS
Since most DSP applications require both an ADC and a DAC, I/O Ports and
CODECs have been developed which integrate the two functions on a single chip as
well as provide easy-to-use interfaces to standard DSPs. These devices also go by
the name of analog front ends.
A functional block diagram of the AD73322 is shown in Figure 8.23. This device is a
dual analog front end (AFE) with two 16-bit ADCs and two 16-bit DACs capable of
sampling at 64KSPS. It is designed for general purpose applications including
speech and telephony using sigma-delta ADCs and sigma-delta DACs. Each channel
provides 77dB signal-to-noise ratio over a voiceband signal bandwidth.
The ADC and DAC channels feature programmable input/output gains with ranges
of 38dB and 21dB, respectively. An on-chip voltage reference is included to allow
single supply operation on +2.7V to +5.5V. Power dissipation is 73mW with a +3V
supply.
8.19
INTERFACING TO DSPS
AD73322 SINGLE SUPPLY 16-BIT, 64kSPS
CODEC WITH SERIAL INTERFACE
AVDD1
VFBP1
VINP1
VINN1
VFBN1
VOUTP1
VOUTN1
SDI
Σ∆ DAC
CHANNEL 1
SCLK
REFERENCE
VFBP2
VINP2
VINN2
VFBN2
Σ∆ ADC
CHANNEL 2
VOUTN2
DVDD
Σ∆ ADC
CHANNEL 1
REFOUT
REFCAP
VOUTP2
AVDD2
SDIFS
SE
SERIAL
PORT
MCLK
SDOFS
Σ∆ DAC
CHANNEL 2
AGND1
RESET
SDO
AGND2
DGND
Figure 8.23
The sampling rate of the CODECs is programmable with four separate settings of
64kHz, 32kHz, 16kHz, and 8kHz when operating from a master clock of
16.384MHz. The serial port allows easy interfacing of single or cascaded devices to
industry standard DSP engines, such as the ADSP-21xx family. The SPORT
transfer rate is programmable to allow interfacing to both fast and slow DSP
engines. The interface to the ADSP-218x family is shown in Figure 8.24. The SE pin
(SPORT enable) may be controlled from a parallel output pin or a flag pin such as
FL1, or where SPORT power down is not required, it can be permanently strapped
high using a suitable pull-up resistor. The RESET pin may be connected to the
system hardware reset, or it may be controlled with another flag bit.
In the program mode, data is transferred from the DSP to the AD73322 control
registers to set up the device for desired operation. Once the device has been
configured by programming the correct settings to the various control registers, the
device may exit the program mode and enter the data mode. The dual ADC data is
transmitted to the DSP in two blocks of 16-bit words. Similarly, the dual DAC data
is transmitted from the DSP to the AD73322 in two blocks of 16-bit words.
Simplified interface timing is also shown in Figure 8.24.
8.20
INTERFACING TO DSPS
AD73322 INTERFACE TO ADSP-218x SERIES
(DATA TRANSFER MODE)
TFS
SDIFS
DT
SDI
ADSP-218x
DSP
SCLK
SCLK
DR
AD73322
CODEC
CLOCK
16.384MHz
SDO
RFS
SDOFS
FL0
RESET
FL1
SE
SE
SCLK
SDOFS
SDO
ADC SAMPLE WORD, DEVICE 2 ADC SAMPLE WORD, DEVICE 1
SDIFS
SDI
DAC DATA WORD, DEVICE 2
DAC DATA WORD, DEVICE 1
Figure 8.24
The AD73422 is the first product in the dspConverter™ family of products which
integrate a dual analog front end (AD73322) and a DSP (52MIPS ADSP-2185L/86L).
The entire functionality of the dual-channel CODEC and the DSP fits into a small,
119-ball 14mm by 22mm plastic ball grid array (PBGA) package. The obvious
advantage is the saving of circuit board real estate. ADC and DAC signal-to-noise
ratios are approximately 77dB over voiceband frequencies.
The AD74222-80 integrates 80K bytes of on-chip memory configured as 16K words
(24-bit) of program RAM, and 16K words (16-bit) of data RAM. The AD73422-40
integrates 40K bytes of on-chip memory configured as 8K words (24-bit) of program
RAM, and 8K words (16-bit) of data RAM. Power-down circuitry is also provided to
meet the low power needs of battery operated portable equipment. The AD73422
operates on a +3V supply and dissipates approximately 120mW with all functions
operational.
8.21
INTERFACING TO DSPS
AD73422 dspConverter™
■
■
■
■
■
Complete Dual CODEC (AD73322) and DSP (ADSP-2185L/86L)
14mm by 22mm BGA package
+3V Single-Supply Operation, 73mW Power Dissipation
Power-Down Mode
CODEC:
◆ Dual 16-bit Sigma-Delta ADCs and DACs
◆ Data Rates: 8, 16, 32, and 64KSPS
◆ 77dB SNR
■ DSP:
◆ 52MIPS
◆ ADSP-218x Code Compatible
◆ 80K Byte and 40K Byte On-Chip Memory Options
Figure 8.25
HIGH-SPEED INTERFACING
With the advent of ever faster DSP clock rates and newer architectures it has
become possible to acquire and process high speed signals. The programmability of
DSPs makes it possible to run different algorithms on the same hardware, while
providing different system functionality. Figure 8.26 shows a simplified ADSP21065L system connected to a high-speed ADC and high-speed DAC. The ADC and
DAC both have parallel interfaces connected to the External Port of the DSP. With
the SHARC family of DSPs there are several ways of connecting the converters to
this port. The access to the converters can be done using the direct memory access
(DMA) controller of the DSP or it can be done under program control using the core
of the DSP. Using the DMA places no load on the DSP core so it can continue
processing (executing program instructions) while the data is transferred to from
the on-chip memory.
The AD9201 is a dual-channel, 10-bit, 20MSPS ADC which operates on a single
+2.7V to +5.5V supply and dissipates only 215mW (+3V supply). The AD9201 offers
closely matched ADCs needed for many applications such as I/Q communications.
Input buffers, an internal voltage reference and multiplexed digital outputs buffers
make interfacing to the AD9201 very simple.
The companion part to the AD9201 ADC is the AD9761 DAC. The AD9761 is a dual
10-bit, 20MSPS per channel DAC operating on a single +2.7V to +5.5V supply and
dissipating only 200mW (+3V supply). A voltage reference, digital latches and 2x
interpolation make the AD9761 useful for I/Q transmitter applications.
8.22
INTERFACING TO DSPS
AD9201 ADC AND AD9761 DAC
INTERFACE TO ADSP-21065L
ADSP21065L
A0
D0 - D9
RD
WR
TCLK
DUAL 10-BIT
20MSPS ADCs
AD9201
ADC
AD9761
DAC
D0 - D9
D0 - D9
CLOCK
CLOCK
SELECT
SELECT
CHIP-SELECT
DUAL 10-BIT
20MSPS DACs
WRITE
Figure 8.26
DSP SYSTEM INTERFACE
Figure 8.26 shows a simplified ADSP-2189M system using the full memory mode
configuration with two serial devices, a byte-wide EPROM, and optional external
program and data overlay memories. Programmable wait state generation allows
the fast processor to connect easily to slower peripheral devices. The ADSP-2189M
also provides four external interrupts, seven general-purpose input/output pins and
two serial ports. One of the serial ports can be configured as two additional
interrupts, a general-purpose input and a general-purpose output pin for a total of
six external interrupts, 9 IOs and one serial port. The ADSP-2189M can also be
operated in the host memory mode which allows access to the full external data bus,
but limits addressing to a single address bit. Additional system peripherals can be
added in the host memory mode through the use of external hardware to generate
and latch address signals.
8.23
INTERFACING TO DSPS
ADSP-2189M SYSTEM INTERFACE:
FULL MEMORY MODE
ADSP-2189M
CLKIN
1/2 X CLOCK
OR
CRYSTAL
INTERRUPTS
GENERAL
PURPOSE IO
XTAL
4
7
14
14
ADDR
8
24
MODE x
5
SPORT1
DATA
11
WR
16
IOMS
SPORT0
PMS
ADDR
DATA
I/O SPACE
(PERIPHERALS)
2048 LOCATIONS
OVERLAY
MEMORY
TWO 8K PM SEGMENTS
TWO 8K DM SEGMENTS
DMS
CMS
BR
BG
BGH
PWD
PWDACK
BUS REQUEST/
GRANT/HUNG
POWER DOWN INPUT
POWER DOWN OUTPUT
Figure 8.27
8.24
DATA
CS
24
SERIAL
DEVICE
ADDR
14
5
BYTE
MEMORY
CS
BMS
RD
SERIAL
DEVICE
ADDR
8
DATA
IRQ x
22
INTERFACING TO DSPS
REFERENCES
1.
Steven W. Smith, The Scientist and Engineer’s Guide to Digital Signal
Processing, Second Edition, 1999, California Technical Publishing,
P.O. Box 502407, San Diego, CA 92150. Also available for free download at:
http://www.dspguide.com or http://www.analog.com
2.
C. Britton Rorabaugh, DSP Primer, McGraw-Hill, 1999.
3.
Richard J. Higgins, Digital Signal Processing in VLSI, Prentice-Hall,
1990.
4.
DSP Designer’s Reference (DSP Solutions) CDROM, Analog Devices,
1999.
5.
DSP Navigators: Interactive Tutorials about Analog Devices’ DSP
Architectures (Available for ADSP-218x family and SHARC family):
http://www.analog.com/industry/dsp/training/index.html#Navigator
6.
General DSP Training and Workshops:
http://www.analog.com/industry/dsp/training
The following DSP Reference Manuals and documentation are available
for free download from: http://www.analog.com/industry/dsp/tech_docs.html
7.
ADSP-2100 Family Users Manual, 3rd Edition, Sept., 1995.
8.
ADSP-2100 Family EZ Tools Manual.
9.
ADSP-2100 EZ-KIT Lite Reference Manual.
10.
Using the ADSP-2100 Family, Vol. 1, Vol. 2.
11.
ADSP-2106x SHARC User’s Manual, 2nd Edition, July, 1996.
12.
ADSP-2106x SHARC EZ-KIT Lite Manual.
13.
ADSP-21065L SHARC User’s Manual, Sept. 1, 1998.
14.
ADSP-21065L SHARC EZ-LAB User’s Manual.
15.
ADSP-21160 SHARC DSP Hardware Reference.
8.25
INTERFACING TO DSPS
8.26
DSP APPLICATIONS
SECTION 9
DSP APPLICATIONS
■
High Performance Modems for Plain Old Telephone
Service (POTS)
■
Remote Access Server (RAS) Modems
■
ADSL (Assymetric Digital Subscriber Line)
■
Digital Cellular Telephones
■
GSM Handset Using SoftFone™ Baseband Processor
and Othello™ Radio
■
Analog Cellular Basestations
■
Digital Cellular Basestations
■
Motor Control
■
Codecs and DSPs in Voiceband and Audio Applications
■
A Sigma-Delta ADC with Programmable Digital Filter
9.a
DSP APPLICATIONS
9.b
DSP APPLICATIONS
SECTION 9
DSP APPLICATIONS
Walt Kester
HIGH PERFORMANCE MODEMS FOR PLAIN OLD
TELEPHONE SERVICE (POTS)
Modems (Modulator/Demodulator) are widely used to transmit and receive digital
data using analog modulation over the Plain Old Telephone Service (POTS) network
as well as private lines. Although the data to be transmitted is digital, the telephone
channel is designed to carry voice signals having a bandwidth of approximately 300
to 3300Hz. The telephone transmission channel suffers from delay distortion, noise,
crosstalk, impedance mismatches, near-end and far-end echoes, and other
imperfections. While certain levels of these signal degradations are perfectly
acceptable for voice communication, they can cause high error rates in digital data
transmission. The fundamental purpose of the transmitter portion of the modem is
to prepare the digital data for transmission over the analog voice line. The purpose
of the receiver portion of the modem is to receive the signal which contains the
analog representation of the data , and reconstruct the original digital data at an
acceptable error rate. High performance modems make use of digital techniques to
perform such functions as modulation, demodulation, error detection and correction,
equalization, and echo cancellation.
A block diagram of an ordinary telephone channel (often referred to as plain old
telephone service – or POTS) is shown in Figure 9.1. Most voiceband telephone
connections involve several connections through the telephone network. The 2-wire
twisted pair subscriber line available at most sites is generally converted to a 4-wire
signal at the telephone central office: two wires for transmit, and two wires for
receive. The signal is converted back to a 2-wire signal at the far-end subscriber
line. The 2- to 4-wire interface is implemented with a circuit called a hybrid. The
hybrid intentionally inserts impedance mismatches to prevent oscillations on the 4wire trunk line. The mismatch forces a portion of the transmitted signal to be
reflected or echoed back to the transmitter. This echo can corrupt data the
transmitter receives from the far-end modem.
Half-duplex modems are capable of passing signals in either direction on a 2-wire
line, but not simultaneously. Full-duplex modems operate on a 2-wire line and can
transmit and receive data simultaneously. Full-duplex operation requires the ability
to separate a receive signal from the reflection (echo) of the transmitted signal. This
is accomplished by assigning the signals in the two directions different frequency
bands separated by filtering, or by echo cancelling in which a locally synthesized
replica of the reflected transmitted signal is subtracted from the composite receive
signal.
9.1
DSP APPLICATIONS
ANALOG MODEM USING PLAIN OLD TELEPHONE
SERVICE (POTS) ANALOG CHANNEL
NOISE
FAR-END
MODEM
NEAR-END
MODEM
TRANSMIT
CHANNEL
×
+
RECEIVER
2
TRANSMITTER
2
2
HYBRID
TRANSMITTER
FREQUENCY
SHIFT
FAR-END
ECHO
2
HYBRID
2
2
RECEIVE
CHANNEL
×
NEAR-END
ECHO
RECEIVER
+
FREQUENCY
NOISE
SHIFT
FOUR-WIRE TRUNK
Figure 9.1
There are two types of echo in a typical voiceband telephone connection. The first
echo is the reflection from the near-end hybrid, and the second echo is from the farend hybrid. In long distance telephone transmissions, the transmitted signal is
hetrodyned to and from a carrier frequency. Since local oscillators in the network
are not exactly matched, the carrier frequency of the far-end echo may be offset
from the frequency of the transmitted carrier signal. In modern applications this
shift can affect the degree to which the echo signal can be canceled. It is therefore
desirable for the echo canceller to compensate for this frequency offset.
For transmission over the telephone voice network, the digital signal is modulated
onto an audio sinewave carrier, producing a modulated tone signal. The frequency of
the carrier is chosen to be well within the telephone band. The transmitting modem
modulates the audio carrier with the transmit data signal, and the receiving modem
demodulates the tone to recover the receive data signal.
The baseband data signal may be used to modulate the amplitude, the frequency, or
the phase of the audio carrier, depending on the data rate required. These three
types of modulation are known as amplitude shift keying (ASK), frequency shift
keying (FSK), and phase shift keying (PSK). In its simplest form the modulated
carrier takes on one of two states - that is, one of two amplitudes, one of two
frequencies, or one of two phases. The two states represent a logic 0 or a logic 1.
Low- to medium-speed data links usually use FSK up to 1,200 bits/s. Multiphase
PSK are used for 2,400 bits/s and 4,800 bits/s links. PSK utilizes bandwidth more
efficiently than FSK but is more costly to implement. ASK is least efficient and is
9.2
DSP APPLICATIONS
used only for very low speed links (less than 100 bits/s). For 9,600 bits/s up to
33,600 bits/s a combination of PSK and ASK is used, known as Quadrature
Amplitude Modulation (QAM).
The International Telegraph and Telephone Consultative Committee (CCITT in
France) has established standards and recommendations for modems which are
given in Figure 9.2.
SOME MODEM STANDARDS
CCITT
Rec.
V.21
V.22
V.22bis
V.23
V.26 bis
V.26 ter
V.27 ter
V.32
V.32 bis
V.34
V.90
V.92
Approximate
Date
1964
1998
2001
Speed
(bits/s)
maximum
300
1200
2400
1200
2400
2400
4800
9600
14400
33600
56000*
56000**
Half Duplex/
Full Duplex/
Echo
Cancel
FDX
FDX
FDX
HDX
HDX
FDX(EC)
HDX
FDX(EC)
FDX(EC)
FDX(EC)
FDX(EC)
FDX(EC)
Modulation
Method
FSK
PSK
16QAM
FSK
PSK
PSK
8PSK
32QAM
QAM
QAM
PCM
PCM
* DOWNSTREAM ONLY, UPSTREAM IS V.34
** UPSTREAM AND DOWNSTREAM
Figure 9.2
The goal in designing high performance modems is to achieve the highest data
transfer rate possible over the POTS network and avoid the expense of using
dedicated conditioned private telephone lines. The V.90 recommendation describes a
full-duplex (simultaneous transmission and reception) modem that operates on the
POTS network. The V.90 modem communicates downstream from the central office
to the subscriber modem at a rate of 56,000 bits/s using Pulse Code Modulation
(PCM). Upstream communication from the subscriber to the central office is at the
V.34 rate of up to 33,600 bits/s (QAM).
A simplified block diagram for a V.90 analog modem is shown in Figure 9.3. The
diagram shows that the bulk of the signal processing is done digitally. Both the
transmit and receive portions of the modem subject the digital signals to a number
of DSP algorithms which can be efficiently run on modern processors.
9.3
DSP APPLICATIONS
V.90 ANALOG MODEM SIMPLIFIED BLOCK DIAGRAM
MIXED-SIGNAL FUNCTIONS
fs
×
TX
DATA
SIGNAL
ENCODING
MODULATION
sinωt
AND
FILTERING
SIGNAL
ENCODING
Σ
ANALOG
LPF
DAC
×
cosωt
ECHO CANCELLING
ADAPTIVE
FILTER
DSP FUNCTIONS
fs
×
RX
DATA
SIGNAL
DECODING
DEMODULATION
AND
FILTERING
HYBRID
POTS
LINE
–
Σ
sinωt
SIGNAL
DECODING
Σ
ADC
ANALOG
LPF
×
cosωt
Figure 9.3
The TX input serial bit stream is first scrambled and encoded. Scrambling takes the
input bit stream and produces a pseudo-random sequence. The purpose of the
scrambler is to whiten the spectrum of the transmitted data. Without the
scrambler, a long series of identical symbols could cause the receiver to lose carrier
lock. Scrambling makes the transmitted spectrum resemble white noise, to utilize
the bandwidth of the channel more efficiently, makes carrier recovery and timing
synchronization easy, and makes adaptive equalization and echo cancellation
possible.
The scrambled bit stream is divided into groups of bits, and the groups of bits are
first differentially encoded and then convolutionally encoded.
The symbols are then mapped into the signal space using QAM as defined in the
V.34 standard. The signal space mapping produces two coordinates, one for the real
part of the QAM modulator and one for the imaginary part. As an example, a
diagram of a 16-QAM signal constellation is shown in Figure 9.4. Larger
constellations are used in V.90 modems, and the actual size of the constellation is
adaptive and determined during the training, or “handshake” interval when the
modems synchronize with each other for upstream or downstream signaling.
9.4
DSP APPLICATIONS
QUADRATURE AMPLITUDE MODULATED (QAM) SIGNAL
TRANSMITS 4 BITS PER SYMBOL (16-QAM)
Q
1111
4-BITS/SYMBOL
I
0000
I OR Q CHANNEL
SAMPLING
CLOCK
t
Figure 9.4
Used prior to modulation, digital pulse shaping filters attenuate frequencies above
the Nyquist frequency that are generated in the signal mapping process. These
filters are designed to have zero crossings at the appropriate frequencies to cancel
intersymbol interference.
QAM is easily implemented in modern DSP processors. The process of modulation
requires the access of a sine or cosine value, the access of an input symbol (x or y
coordinate) and a multiplication. The parallel architecture of the ADSP-21xx-family
permits all three operations to be performed in a single instruction cycle.
The output of the digital modulator drives a DAC. The output of the DAC is passed
through an analog lowpass filter and to the 2-wire telephone line for transmission
over the POTS network.
The receiver is made up of several functional blocks: the input antialiasing filter
and ADC, a demodulator, an adaptive equalizer, a Viterbi decoder, an echo
canceller, a differential decoder, and a descrambler. The receiver DSP algorithms
are both memory-intensive and computation-intensive. The ADSP-218x-family
addresses both needs, providing sufficient program memory RAM (for both code and
data) on chip, data memory RAM on chip, and an instruction execution rate of up to
75MIPS.
The antialiasing filter and ADC in the receiver need to have a dynamic range from
the largest echo signal to the smallest. The received signal can be as low as -40dBm,
9.5
DSP APPLICATIONS
while the near-end echo can be as high as -6dBm. In order to ensure that the analog
front end of the receiver does not contribute any significant impairment to the
channel under these conditions, an instantaneous dynamic range of 84dB and an
SNR of 72dB is required.
In order to compensate for amplitude and phase distortion in the telephone channel,
equalization is required to recover the transmitted data at an acceptably low bit
error rate. In order to respond to rapidly changing conditions on the telephone line,
adaptive equalization is required for the V.90 modem receiver. An adaptive
equalizer can be implemented digitally in an FIR filter whose coefficients are
continuously updated based on current line conditions.
Separation between the transmit and receive signal in the V.90 modem is
accomplished using echo cancellation. Both near-end and far-end echo must be
cancelled in order to yield reliable communication. Echo cancellation is achieved by
subtracting an estimate of the echo return signal from the actual received signal.
The predicted echo is determined by feeding the transmitted signal into an adaptive
filter with a transfer function that approximates the telephone channel. The
adaptive filter commonly used in echo cancellers is the FIR filter (chosen for its
stability and linear phase response). The taps are determined using the least-meansquare (LMS) algorithm during a training sequence executed prior to full-duplex
communications.
The most common technique for decoding the received data is Viterbi decoding.
Named after its inventor, the Viterbi algorithm is a general-purpose technique for
making an error-corrected decision. Viterbi decoding provides a certain degree of
error correction by examining the received bit pattern over time to deduce the value
that was the most likely to have been transmitted at a particular time. Viterbi
decoding is computation-intensive. A history for each of the possible symbols sent at
each symbol interval has to be maintained. At each symbol interval, the length of
the path backward in time from each possible received symbol to a symbol sent
some time ago is calculated. The symbol that has the shortest path back to the
original signal is chosen to be the current decoded symbol. A complete description of
Viterbi decoding and its implementation on the ADSP-21xx- family of processors is
given in documentation available from Analog Devices (Reference 2).
Figure 9.5 shows a comparison between V.34 and V.90 modems. Note that in the
case of V.34 (Figure 9.5A), the communication is between two analog modems. This
requires and ADC/DAC in both the transmit and receive path as shown in the
diagram. The V.90 system requires an all digital network and a V.90 digital modem
as shown in Figure 9.5B. Note that the second ADC/DAC combination is eliminated,
thereby allowing the faster downstream data rate of 56Kbits/s. Downstream
communication to the V.90 analog modem uses 64Kbits/s PCM data, which is
standard for all digital telephone networks. This serial data is converted into a
pulse amplitude modulated (PAM) signal (8-bits, 8kSPS) using an 8-bit DAC. The
signal from the DAC to the analog modem is therefore a 256K constellation with no
imaginary component, i.e., the analog modem receiver must detect which one of the
256 levels is being sent during the symbol interval.
9.6
DSP APPLICATIONS
The V.90 standard allows downstream data rates of up to slightly less than
56Kbits/s, and upstream data rates of up to 33.6Kbits/s (V.34). The V.92 standard
will allow 56Kbits/s transmission in both directions.
V.34 VERSUS V.90 MODEMS
33.6K bps
A
V.34
ANALOG
MODEM
QAM
DATA
DAC
PSTNANALOG OR
DIGITAL
ADC
ADC
QAM
DATA
DAC
V.34
ANALOG
MODEM
33.6K bps
B
8-BIT, 8KSPS
PAM DATA
56K bps
DAC
V.90
ANALOG
MODEM
8-BIT, 8KSPS PCM = 64K bps
PSTNDIGITAL
ADC
QAM
DATA
(V.34)
V.90
DIGITAL
MODEM
33.6K bps
Figure 9.5
REMOTE ACCESS SERVER (RAS) MODEMS
Rapid growth and use of the Internet has created a problem in that there are more
users trying to get on the Internet than there is equipment to accommodate all
these users. Internet Service Providers (ISPs) like America On Line purchase
modem equipment so their customers (referred to as subscribers) can remotely
access a network (like the Internet from home). This application of accessing a
network from a remote location is called Remote Network Access. The equipment
used in this application is called a Remote Access Server (RAS) as shown in Figure
9.6. The Remote Access Server is made up of many modem ports; each modem port
can connect to a different user. The RAS can use analog modems, which connects to
a POTS line, or digital modems which are compatible with T1, E1, PRI, or BRI lines.
Digital modems are used in most RAS systems since they are more efficient for 8
ports or more.
Network access equipment enables individuals, small offices, and traveling
employees to connect to corporate networks (Intranets) and the Internet. Internet
Service Providers use devices called concentrators to connect their telephone access
lines to their networks. These concentrators are also referred to as Remote Access
Servers. The rapid growth in the use of the Internet and Intranets has created a
tremendous demand for modem equipment.
9.7
DSP APPLICATIONS
INTERNET GATEWAY USING
REMOTE ACCESS SERVER (RAS) MODEM
INTERNET GATEWAY
(REMOTE ACCESS SERVER)
POTS
DATA
DATA
MODEM
VOICE
CENTRAL
OFFICE
FAX
MODEM
INTERNET
PROVIDER (IP)
DATA
ROUTER
ISDN
FAX
VIDEO
VOIP
Network
DATA
ROUTER
Figure 9.6
Remote access servers accommodate employees and small offices/home offices
(SOHO) wishing to connect individual computers to LANs (Local Area Networks) or
Intranets. If a remote access server is installed in a corporate LAN, remote users
can access the network in such a way that their computers appear to be directly
connected to the LAN. This allows them to work from a remote location as if they
were sitting at their desk in their office.
The ADSP-21mod870 acts as a bridge between the voice-based continuous
connection switched network and the data based IP network as shown in Figure 9.7.
The high speed DMA interface and large on-chip RAM of the ADSP-21mod870
allows it to be configured to handle a large variety of tasks. The software with the
ADSP-21mod870-100 can be configured for modem calls or HDLC (high bit-rate
digital subscriber line) processing of digital ISDN (integrated services digital
network) based calls. Since the ADSP-21mod870 is an open platform, other
functions can be loaded by users. Examples of other functions include voice over
internet and FAX over internet. In these applications, the ADSP-21mod870 is a
gateway for voice network users to save toll charges by routing their call over the IP
network. The ADSP-21mod870 DSP uses the ADSP-218x 16-bit fixed-point core and
is code compatible with other members of the ADSP-21xx family.
As the number of remote network users has grown, capacity over the telephone
central office switched network has often become strained. These bottlenecks often
occur when thousands of calls from a metropolitan area are switched to a single
point of presence (POP). To eliminate these bottle necks, RAS equipment can be
pushed outward from the POP toward the edges of the switched network as shown
in Figure 9.8. When RAS equipment is located in the switching center for local
exchanges, data calls can be separated from voice calls, eliminating the strain on
9.8
DSP APPLICATIONS
the voice based switched network. RAS equipment integrated into switching
equipment is referred to on-switch based RAS. This differs for standalone RAS
systems in that on-switch RAS can separate data calls before they are connected to
trunk lines exiting the switch.
DSP-BASED RAS MODEM USING THE
ADSP-21modXXX - FAMILY
TRADITIONAL RAS MODEM
DSP-BASED RAS MODEM
DATA
MODEM
DATA MODEM
FAX MODEM,
ISDN,
VOIP
FAX
MODEM
ISDN
DATA MODEM
FAX MODEM,
ISDN,
VOIP
VoIP
ADSP-21modXXX
FAMILY
DATA
ROUTER
DATA
ROUTER
Figure 9.7
EXPANDING CENTRAL OFFICE
CAPABILITY WITH ADSP-21modXXX FAMILY
CENTRAL OFFICE SWITCH
LINE CARDS
DATA
VOICE
ADSP-21modXXX
Modem Pool
DATA
IP
Network
DATA
ROUTER
LINE CARD
LINE CARD
Voice
Network
FAX
VIDEO
VOICE
VOICE
SWITCH
LINE CARD
Figure 9.8
9.9
DSP APPLICATIONS
Several types of RAS equipment have evolved to satisfy the needs of the different
remote network access users. RAS equipment can take several forms. RAS
concentrators combine a modem pool with a router in a standalone chassis. NT
server RAS uses off the shelf Windows NT PCs to perform routing functions with a
modem pool connected to its PCI or ISA expansion bus. On-switch RAS integrates
modem pools into the line card chassis in switching systems. These various forms of
RAS equipment serve the different needs of end users. Local Exchange Companies
(LECs) or the new Competing Local Exchange Companies (CLECs) can take
advantage of On-switch RAS equipment to eliminate bottle necks in their switched
networks. ISPs (Internet Service Providers) use RAS concentrators to collect calls
from a large area for connection to the Internet. Large business users also use RAS
concentrators to connect their employees to their LAN or Intranets. Small office and
home office (SOHO) users can use low-cost NT Server based RAS to support all of
their RAS, LAN, and other telecommunication needs.
The ADSP-21mod870 digital modem processor is the first complete digital RAS
modem on a single chip. The device is V.34/56K and V.42/V.42bis compatible, has a
16-bit DMA port for software downloads, provides TDM serial port interfaces
directly to T1/E1, has 160Kbytes on-chip SRAM, dissipates 140mW at +3.3V, and is
packaged in a 16mm TQFP package. The ADSP-21mod870’s small footprint and
power efficiency will enable Internet Service Providers (ISPs) and central office
access providers to quadruple the number of ports within the existing modem bank
chassis. Moreover, the chip’s unique capability to support any protocol on any port
will improve ISP customer service and reduce business operating costs.
Part of a family of digital modem solutions from ADI, the ADSP-21mod870 includes
silicon, software and service. ADI is singular in its ability to provide the complete
solution. Also available are the ADSP-21mod970 (six modem channels, 31mm BGA)
and the ADSP-21mod980 (eight modem channels, 35mm BGA).
MULTI-CHANNEL VOICE-OVER-INTERNET-PROVIDER
(VOIP) SERVER
The ADSP-218x family of devices can efficiently implement multi-channel telephony
applications (such as RAS/VOIP servers and gateways) due to their high
performance and large on-chip memories. A typical system is shown in Figure 9.9.
The programmable nature of the DSP architecture gives the system architect the
flexibility to implement various speech coding algorithms in addition to baseline
telephony functionality.
The ADSP-2188M is the most highly integrated member of the 218x family (with
over 2MBit of on-chip SRAM). This high level of integration combined with 75MIPS
performance can support up to 6 voice channels per DSP (depending on the voice
coder chosen).
9.10
DSP APPLICATIONS
ADSP-218x IMPLEMENTATION OF A
MULTI-CHANNEL VOIP SERVER
Multi-Channel VoIP Server
G.723 annex A
G.728
G.729B
G.711
DTMF Gen & Detect
Call Discrimination
G.165
OS Kernal
Total
PCM
MIPS
18
28
11.4
.4
1.9
1.3
8
.2
Program Memory
8192
7200
10240
512
300
1536
800
512
Data Memory
12288
2000
5120
512
400
300
1024
N/A
29292
21644
Power
Management
T1/E1
(PCM)
Interface
RISC
Processor
> 50MIPS
or
POTS
ADC
DAC
Telephony
Interface
ADSP-218x
DSP
ADSP-2188M
48K PM
56K DM
Source: ADI Estimates
Based on ADSP-218x
10,100,
10/100
Base T/
Ethernet
and/or
ATM/
Frame
Relay
That’s Support
from 4 to 6
channels with
room to grow!
DRAM
or
SRAM
SRAM
Figure 9.9
ADSL (ASYMMETRIC DIGITAL SUBSCRIBER LINE)
Thanks to the widespread popularity of the World Wide Web, Internet traffic is at
an all-time high. A study recently conducted by the Wall Street Journal reported 58
million Internet users in the United States and Canada alone. Research firms all
predict heavier Internet traffic as more and more people buy PCs and use the
Internet for business, academia and recreational purposes.
Unless something is done to improve the way we access the Internet, user traffic
will ultimately burden the public switched telephone network (PSTN) beyond its
original design limits. Internet users are frustrated by the amount of time it takes
to view simple text-based Web sites, especially from 8:00 am to 6:00 pm on business
days when congestion and bottlenecks are most prevalent. The problem will only get
worse as users attempt to view graphically complex sites, download new video and
audio clips, and access other types of multimedia services becoming available over
the Internet.
Today's analog modem and telephone switch technology is simply inadequate.
Assuming little or no network delays, a 10 Megabyte data transmission - the
equivalent of a four-minute audio/video clip - takes about 95 minutes to download
when using a 14.4 Kbps analog modem, 45 minutes when using a 28.8 Kbps modem,
and 25 minutes when using a 56Kbps modem. Lengthy on-line calls are tying up
telephone systems originally designed to handle short (three-minute) voice calls and
9.11
DSP APPLICATIONS
switches built to accommodate approximately nine minutes per line during peak
hours. How often do you browse the Internet for 10 minutes or less? It now appears
that relief, in the form of ADSL (asymmetric digital subscriber line) technology, is
on the way.
ADSL is a new high-speed digital switching/routing and signal processing
technology. It promises to relieve network bottlenecks and provide enough user
bandwidth for the Internet explosion. Originally conceived in 1994, ADSL delivers
the huge amounts of bandwidth needed for interactive gaming, multimedia services
and video-on-demand. These applications, along with videoconferencing, remote
schooling and home shopping are still attractive applications. As more and more
people use the Internet for electronic commerce in homes around the world, the
need for high-speed network access becomes ever more paramount.
ADSL can transfer data over ordinary telephone lines nearly 200 times faster than
today's contemporary modems and 90 times faster than ISDN. Early tests and trials
have shown promising results all over the world. And while GTE and other large
telephone companies begin to deploy ADSL systems in select regions of the US and
abroad, others seek to deploy ADSL-based equipment in 1999 as standards-based
systems and modems become commercially available.
Due to ADSL's technical complexity, only a few semiconductor manufacturers are
presently developing ADSL silicon chips. Analog Devices is one of them, and is
considered a pioneer in ADSL after having produced the first complete ADSL
chipset back in 1997. Early adopters of ADSL rallied behind the AD20msp910
chipset after having successfully evaluated its high-speed, long reach capability.
Soon thereafter, the world's most influential industry standards committees (ANSI,
ETSI and ITU) approved an advanced discrete multi-tone (DMT) signal processing
technique used by the AD20msp910. Today, Analog Devices enjoys the advantages
of having the industry's first standards-based solution, the largest installed
customer base, and the most design wins of any semiconductor manufacturer to
date.
ADSL is attractive for the following reasons:
•
•
•
•
9.12
ADSL is fast. The same 10 MB video clip that takes 90 minutes to download
using a conventional modem would be downloaded in just 10 seconds using
an ADSL modem. Superfast ADSL modems can transmit data at rates as
high as 8 megabits per second.
ADSL is easy to install. It uses existing copper twisted-pair telephone lines
from a local exchange carrier (central office) to the subscriber's home or
office. Little or no rewiring is necessary.
ADSL is cost-effective. It requires no major upgrades in the existing
telephone network infrastructure.
ADSL is viable. Issues that have slowed the deployment of high-speed fiber
networks to the home (e.g., prohibitive cost and installation) do not apply.
ADSL works with existing POTS (plain-old telephone service). High-speed
data transport can occur simultaneously with voice calls and fax
transmissions.
DSP APPLICATIONS
Unlike other high-speed data transmission technologies, ADSL requires no rewiring
over the "last mile" of the network. Although commonly referred to as the last mile,
transmission length is typically 12,000 to 18,000 feet. That "last mile" (the leg from
the central office to a user's home or office) operates over existing copper loops of
twisted-pair telephone wires. But ADSL does require the installation of new
equipment at local exchange central offices and major switch offices. However, the
technology developed for central office equipment is the same as that used in PC
modems and home splitter boxes, thus ensuring interoperability throughout the
entire network. A simplified block diagram of an ADSL system is shown in Figure
9.10.
ASYMMETRIC DIGITAL SUBSCRIBER LINE (ADSL)
SYSTEM CHARACTERISTICS
HOME
CENTRAL OFFICE
8 Mbps
TELEPHONE
FAX
POTS
SPLITTER
PC OR
NETWORK
COMPUTER
ADSL MODEM
VOICE
SWITCH
640 Kbps
Single Pair
Twisted Wire
e.g. 12,000 ft
Voice
Network
POTS
SPLITTER
THE
“LAST MILE”
1MHz BANDWIDTH
■ 8Mbps Downstream Channel
■ 640Kbps Upstream/Downstream Channel
■ 1 Telephone Channel
ADSL
MODEM
BANK
DATA
ROUTER
Internet
Figure 9.10
Using sophisticated digital signal processing techniques, ADSL modems push as
much data as possible across copper wires by making the best use of available
bandwidth. Many perceived the telephone network is as having only 4 kHz of
bandwidth, but 4 kHz only represents the band used for the analog (voice)
transmission. Using ADSL, the physical connection between the home and local
exchange carrier (LEC) over conventional copper wire is capable of carrying 1 MHz.
ADSL takes advantage of the portion of the bandwidth not used for voice calls.
Essentially, it splits the 1 MHz bandwidth into three information channels: one
high-speed downstream channel, one medium-speed duplex (upstream/downstream)
channel, and one conventional voice channel. (Downstream refers to data
transmitted from the telephone network to the customer's premises; upstream is
data routed from the customer to the network.)
In addition to providing high bandwidth data transmission, ADSL preserves
"lifeline" services (i.e., the voice network that people depend upon for day-to-day
communications and emergency situations). This three-channel approach allows
9.13
DSP APPLICATIONS
subscribers to send an email, download a video for viewing, and talk on the
telephone-all at the same time. Telecommuters can access their corporate local area
network, and simultaneously videoconference with a customer. In fact, using
ADSL's full capability provides enough bandwidth for running four channels of
MPEG compressed video with no interruption to normal telephone service.
Most Internet applications require uneven amounts of upstream and downstream
bandwidth. In other words, users tend to move greater amounts of data in one
direction than they do in the other direction. Generally speaking, Internet users
access more information and receive more data than they transmit. That's the
nature of the Internet. They read more electronic mail than they send...they
download more video information than they produce. A user's upstream capacity is
typically limited to sending commands or transmitting small data files to a server.
Much more information travels downstream. ADSL was designed to take advantage
of this trend in uneven bandwidth. It provides data rates in excess of 8 Mbps from
the network to the subscriber (downstream), and up to 640 Kbps from the
subscriber to the network (upstream).
For telephone companies, ADSL may be the key to overcoming the congestion
created by explosive Internet growth. If telephone companies hope to capture the
Internet data carrier market, they can't use equipment designed for short voice
traffic. The average voice call lasts about three minutes. According to a 1997 study
by Bellcore, the research laboratory funded by US local telephone companies, (one of
five submitted to the Federal Communications Commission) the average length of
time for a typical Internet call is now in excess of 20 minutes. Telephony magazine
reported as early as 1994 that nearly 20 percent of all on-line data service
connections lasted more than an hour. Cable companies like MediaOne, based in the
Northeast, charge one flat rate for Internet use. "Broadband" subscribers can stay
on line 24-hours a day for roughly $40 a month.
ADSL was designed to take advantage of this natural unevenness in bandwidth
requirements and make the most of available bandwidth, a limited resource. It
provides for rates in excess of 8 Mbps from the network to the subscriber, and up to
640 Kbps from the subscriber to the network.
While some other telephony-based technologies offer relief for the PSTN, many
carry high price tags and long-term implementation time frames. Replacing the
copper network with fiber to the curb, for example, is a costly proposition. Industry
experts estimate that creating a fiber-to-the-curb network will cost about $1,500 per
telephone customer. Replacing the existing 560 million copper subscriber lines
worldwide would cost more than three-quarters of a trillion dollars.
ADSL, by contrast, is easily installed. To complete an ADSL circuit, two modems
are required, one at each end of a twisted-pair telephone line. One modem is located
at the subscriber's premise; the other modem (usually a rack of modems with line
cards) is located at the local telephone company's central office. Figure 9.10 shows a
simplified connection.
At the customer premise, an ADSL modem is inside a PC or connected to the
network computer, as well as the telephone and/or facsimile machine. A copper
twisted pair telephone line links the POTS splitter to the central office. At the
9.14
DSP APPLICATIONS
central office, where the switch and rack of modems (line cards) are located, a POTS
splitter separates voice from data. Voice calls are forwarded to the central office
switch for relay over the public telephone network. Data transmissions are sent
through an Ethernet switch and a router and eventually over a high-speed
connection (for example, 155 Mbps OC-3) to the Internet service provider. The
connection is made to the Internet backbone for access to the World Wide Web.
The key to ADSL's high-speed data transmission is advanced digital signal
processing (DSP). Drawing on its analog innovation and advanced digital signal
processing capabilities, Analog Devices developed the first generation AD20msp910
chipset (see Figure 9.11). The AD20msp910 offers three competitive advantages to
modem manufacturers:
•
•
•
It is the most functionally complete chipset solution on the market.
It is fully compliant to ANSI, ETSI and ITU industry standards.
It is compatible with nearly every DLC (digital loop carrier) and modem
manufacturer.
The AD20msp910 solution includes both hardware and software. It integrates a
DSP host processor, line driver, and control software, plus DMT technology. Other
supplier’s chips offer only portions of the modem. The AD20msp910 speeds and
simplifies development of ADSL modems for high-speed Internet access and
multimedia services and was released in 1997. The second generation AD20msp918
adds ATM functionality, improved performance and support for ADSL over ISDN
for the European market. Both chipsets are fully compliant to all standards (ANSI
T1.413 Issue 2, ETSI TR328, ITU G.dmt and ITU G.lite for splitterless) and both
are complete solutions, including all microcontroller hardware and software
functions.
ADSL MODEM BLOCK DIAGRAM
CONTROL
MESSAGES
FROM HOST
DSP
ADSP-2183
BOOT EPROM
SERIAL
DATA
RAM
DATA BUS
ADDRESS BUS
SERIAL
DATA
FROM PC
DIGITAL
INTERFACE
CONTROLLER
DISCRETE
MULTITONE
ENGINE
ADC, DAC,
AND
ANALOG
FRONT END
TO POTS
SPLITTER
TRANSCEIVER
Figure 9.11
9.15
DSP APPLICATIONS
Now on our third generation chipset, the AD20msp930, Analog Devices enables
modem manufacturers to reduce their development time and costs. To further speed
and streamline design efforts, Analog Devices will also provide a PC board layout,
schematics, and a reference design. As a result, modem manufacturers can focus
their development efforts on adding value to the core ADSL modem technology.
DIGITAL CELLULAR TELEPHONES
In the early 1990s, the GSM (Global System for Mobile Communications aka
Groupe Speciale Mobile) was introduced in Europe heralding the introduction of
digital cellular radio. Out of necessity, other countries, including the US, have
adopted various digital standards in order to relieve the congestion and other
problems faced by analog cellular systems such as AMPS (Advanced Mobile Phone
Service). The limitations of analog cellular radio are well known and include call
blocking during busy hours, misconnects and disconnects due to rapidly fading
signals, lack or privacy and security, and limited data transmission rates.
The basic cellular system is shown in Figure 9.12. A region is broken up into cells,
with each cell having its own base station and its own group of assigned
frequencies. Because the radius of each cell is small (10 miles, for example) low
power transmitters and receivers can be used.
CELLULAR PHONE FREQUENCY REUSE
FREQ
GROUP
G
FREQ
GROUP
F
FREQ
GROUP
C
FREQ
GROUP
B
FREQ
GROUP
A
FREQ
GROUP
E
FREQ
GROUP
G
FREQ
GROUP
C
FREQ
GROUP
D
FREQ
GROUP
B
FREQ
GROUP
E
FREQ
GROUP
G
FREQ
GROUP
F
FREQ
GROUP
C
FREQ
GROUP
B
FREQ
GROUP
A
FREQ
GROUP
E
■ Each Cell Covers a Radius of 5 to 10 miles
■ Each Cell Requires its own Basestation for Reception and
Re-Transmission
■ Each Cell Must Handle Multiple Callers Simultaneously
■ Callers May Be Handed Off from One Cell to Another
Figure 9.12
9.16
DSP APPLICATIONS
The cellular system lends itself to frequency reuse, since cells which are far enough
apart can utilize the same band of frequencies without interference. The base
stations must be linked together with an elaborate central control network so that a
call may be handed-off to another cell when the signal strength from the mobile unit
becomes too low for the current cell to handle.
The frequency spectrum allocation for analog cellular radio (AMPS) in the United
States is approximately 825 to 850MHz and 870 to 895MHz. Conventional
architectures (both analog and digital) are channelized. The total spectrum is
divided up into a large number of relatively narrow channels, defined by a carrier
frequency. The carrier frequency is frequency-modulated with the voice signal using
analog techniques. Each full-duplex channel requires a pair of frequencies, each
with a bandwidth of approximately 30kHz. A user is assigned both frequencies for
the duration of the call. The forward and reverse channel are widely separated, to
help the radio keep the transmit and receive functions separated. The signal
bandwidth for the “A” or “B” carriers serving a particular geographical area is
12.5MHz (416 channels, each 30kHz wide). There can only be one caller at a time
per channel.
Time Division Multiple Access (TDMA) allocates bandwidth on a time-slot basis. In
the United States TDMA system, the entire 30kHz channel is assigned to a
particular transmission, but only for a short period of time. A 3:1 multiplexing
scheme means that three conversations can take place with TDMA using the same
amount of bandwidth as one analog cellular conversation does. Each
transmit/receive sequence occurs on time slots lasting 6.7ms. The TDMA system
relies on an extensive amount of DSP technology to reduce the coded speech bit-rate
as well as to prepare the digital data for transmission over the analog medium. The
TDMA approach was chosen for the GSM system and will be discussed later in more
detail.
The second digital approach used in the United States is called Code Division
Multiple Access (CDMA). This technique has been used in secure military
communications for a number of years under the name of spread spectrum. In
spread spectrum, the transmitter transmits in a pseudo-random sequence of
frequency hops over a relatively wide frequency range. The receiver has access to
the same random sequence and can decode the transmission. The effect of adding
additional users on the system is to decrease the overall signal to noise ratio for all
the users. With this technique, the effect of allowing more calls than the normal
capacity is to increase the bit-error rate for all users. New callers can keep coming
in, interference levels will rise gradually, until at some point the process will
become self-regulating: the quality of the voice link will become so bad that users
will cut short or refrain from making additional calls. No one is ever blocked in the
conventional sense, as they are in FDMA or TDMA systems when all channels or
slots are full.
Both TDMA and CDMA digital systems make extensive use of DSP algorithms in
both speech encoding and in preparing the signal for transmission. In the receiver,
DSP techniques are used for demodulation and decoding the speech signal.
At the present time, both the analog and digital systems are in use in the US. In
many cases, analog and digital systems must co-exist in the same geographical area
9.17
DSP APPLICATIONS
and within the same service area. This implies that the cellular basestations must
handle both analog and digital formats, and therefore there is extensive use of
digital techniques in basestation design to simplify the hardware.
The remainder of this section will concentrate on speech processing and channel
coding as they relate to the GSM system. This will serve to illustrate the
fundamental principles which are applicable to all digital mobile radio systems.
DIGITAL MOBILE RADIO STANDARDS
■ Frequency Division Multiple Access (FDMA) - User Allocation
based on Frequency Slots (AMPS System)
■ Time Division Multiple Access (TDMA) - User Allocation Based on
Time Slots (At least 3X More Capacity than FDMA) GSM is an
Example of TDMA
■ Code Division Multiple Access (CDMA) - Based on Spread
Spectrum Technology - More Users Cause Graceful Degradation
in Bit-Error Rate
■ Both TDMA and CDMA make Extensive Use of DSP in Speech
Encoding and Channel Coding for Transmission and Reception
Figure 9.13
The GSM System
Figure 9.14 shows a simplified block diagram of the GSM Digital Cellular Telephone
System. The speech encoder and decoder and discontinuous transmission function
will be described in detail. Up conversion and downconversion portions of the
system contain a digital modem similar to the ones previously discussed. Similar
functions are performed digitally such as equalization, convolutional coding, Viterbi
decoding, modulation and demodulation.
The standard for encoding voice has been set in the T-Carrier digital transmission
system. In this system, speech is logarithmically encoded to 8 bits at a sampling
rate of 8kSPS. The logarithmic encoding and decoding to 8 bits is equivalent to
linear encoding and decoding to 13 bits of resolution. This produces a bit-rate of
104kb/s. In most handsets, a 16-bit sigma-delta ADC is used, so the effective bitrate is 128kb/s. The Speech Encoder portion of the GSM system compresses the
speech signal to 13kb/s, and the decoder expands the compressed signal at the
receiver. The speech encoder is based on an enhanced version of linear predictive
coding (LPC). The LPC algorithm uses a model of the human vocal tract that
represents the throat as a series of concentric cylinders of various diameters. An
excitation (breath) is forced into the cylinders. This model can be mathematically
represented by a series of simultaneous equations which describe the cylinders.
9.18
DSP APPLICATIONS
GSM HANDSET BLOCK DIAGRAM
VOICE ACTIVITY
DETECTOR
MICROPHONE
ADC
LPC AND
SHORT TERM
ANALYSIS
LONG TERM
PREDICTION
VOICE
FLAG
DISCONTINUOUS
TRANSMISSION
CONTROL
SPEECH ENCODER
128Kbits/s
TRANSMIT FUNCTIONS
PHONE NUMBER
DISPLAY
DAC
RECEIVE FUNCTIONS
LONG TERM
PREDICTION
RF
INPUT
13Kbits/s
SPEECH DECODER
SHORT TERM
SYNTHESIS
RF AMP
×
COMFORT NOISE INSERTION
PSEUDO RANDOM EXCITATION
128Kbits/s
UP
CONVERT
13Kbits/s
COEFFICIENT AVERAGING FOR
COMFORT NOISE INSERTION
DTMF
DIALING
ADD ERROR
CORRECTION &
REDUNDANCY BITS
DISCONTINUOUS
TRANSMISSION
CONTROL
REMOVE ERROR
CORRECTION &
REDUNDANCY BITS
DOWN
CONVERT
SPEAKER
Figure 9.14
The excitation signal is passed through the cylinders, producing an output signal. In
the human body, the excitation signal is air moving over the vocal cords or through
a constriction in the vocal tract. In a digital system, the excitation signal is a series
of pulses for vocal excitation, or noise for a constriction. The signal is input to a
digital lattice filter. Each filter coefficient represents the size of a cylinder.
An LPC system is characterized by the number of cylinders it uses in the model.
Eight cylinders are used in the GSM system, and eight reflection coefficients must
be generated.
Early LPC systems worked well enough to understand the encoded speech, but often
the quality was too poor to recognize the voice of the speaker. The GSM LPC system
employs two advanced techniques that improve the quality of the encoded speech.
These techniques are regular pulse excitation (RPE) and long term prediction (LTP).
When these techniques are used, the resulting quality of encoded speech is nearly
equal to that of logarithmic pulse code modulation (companded PCM as in the TCarrier system).
9.19
DSP APPLICATIONS
The actual input to the speech encoder is a series of 16-bit samples of uniform PCM
speech data. The sampling rate is 8kHz. The speech encoder operates on a 20ms
window (160 samples) and reduces it to 76 coefficients (260 bits total), resulting in
an encoded data rate of 13kb/s.
Discontinuous transmission (DTX) allows the system to shut off transmission
during the pauses between words. This reduces transmitter power consumption and
increases the overall GSM system's capacity.
Low power consumption prolongs battery life in the handset and is an important
consideration for hand-held portable phones. Call capacity is increased by reducing
the interference between channels, leading to better spectral efficiency. In a typical
conversation each speaker talks for less than 40% of the time, and it has been
estimated that DTX can approximately double the call capacity of the radio system.
The voice activity detector (VAD) is located at the transmitter. Its job is to
distinguish between speech superimposed on the background noise and noise with
no speech present. The input to the voice activity detector is a set of parameters
computed by the speech encoder. The VAD uses this information to decide whether
or not each 20ms frame of the encoder contains speech.
Comfort noise insertion (CNI) is performed at the receiver. The comfort noise is
generated when the DTX has switched off the transmitter; it is similar in amplitude
and spectrum to the background noise at the transmitter. The purpose of the CNI is
to eliminate the unpleasant effect of switching between speech with noise, and
silence. If you were listening to a transmission without CNI, you would hear rapid
alternating between speech in a high-noise background (i.e. in a car), and silence.
This effect greatly reduces the intelligibility of the conversation.
When DTX is in operation, each burst of speech is transmitted followed by a silence
descriptor (SID) frame before the transmission is switched off. The SID serves as an
end of speech marker for the receive side. It contains characteristic parameters of
the background noise at the transmitter, such as spectrum information derived
through the use of linear predictive coding.
The SID frame is used by the receiver's comfort noise generator to obtain a digital
filter which, when excited by pseudo-random noise, will produce noise similar to the
background noise at the transmitter. This comfort noise is inserted into the gaps
between received speech bursts. The comfort noise characteristics are updated at
regular intervals by the transmission of SID frames during speech pauses.
Redundant bits are then added by the processor for error detection and correction at
the receiver, increasing the final encoded bit rate to 22.8kb/s. The bits within one
window, and their redundant bits, are interleaved and spread across several
windows for robustness.
9.20
DSP APPLICATIONS
GSM HANDSET USING SOFTFONE™ BASEBAND
PROCESSOR AND OTHELLO™ RADIO
Analog Devices recently announced two chipsets which comprise the majority of a
GSM handset. The SoftFone™ chipset performs the baseband and DSP functions,
while the Othello™ radio chipset handles the RF functions.
The frequency bands originally allocated to GSM were 890MHz to 915MHz for
mobile transmitting and 935MHz to 960MHz for mobile receiving. Another
frequency allocation was made to further expand GSM capacity. This band,
allocated to digital communications services (DCS), was 1710MHz to 1785MHz and
1805MHz to 1880MHz. All countries adopting GSM use one of these two pairs of
frequencies. In the United States, these bands were already allocated by the FCC.
In the mid-1990s, a set of bands were made available for GSM in the United States:
1850MHz to 1910MHz, and 1930MHz to 1990MHz.
Because of the frequency allocations in GSM countries (other than the U.S.), most
GSM handsets must be dual-band: capable of handling both GSM and DCS
frequencies. The SoftFone and Othello chipsets supply the main functions necessary
for implementing dual- or triple-band radios for GSM cellular phones. The
AD20msp430 SoftFone™ chipset comprises the baseband portion of the GSM
handset. The AD20msp430 baseband processing chipset uses a combination of GSM
system knowledge and advanced analog and digital signal processing technology to
provide a new benchmark in GSM/GPRS terminal design.
The SoftFone architecture is entirely RAM-based. The software is loaded from
FLASH memory and is executed from the on-chip RAM. This allows fast
development cycles since no ROM-code turns are required. Furthermore, the
handset software can be updated in the field to enable new features. Combined with
the Analog Devices “Othello” RF chipset, a complete multiband handset design
contains less than 200 components, fits in a 20 cm2 single sided PCB layout, and
has a total bill-of-materials cost 20-30% lower than previous solutions. A simplified
block diagram of the handset is shown in Figure 9.15.
The AD20msp430 chipset is comprised of two chips, the AD6522 DSP-based
baseband processor and the AD6521 voiceband/baseband mixed-signal codec.
Together with the “Othello ” radio chipset, the AD20msp430 allows a significant
reduction in the component count and bill-of-materials (BOM) cost of GSM voice
handsets and data terminals. The software and hardware foundations of the
AD20msp430 chipset enjoy a long history of successful integration into GSM
handsets.
9.21
DSP APPLICATIONS
OTHELLO™ RADIO AND SOFTPHONE™ CHIPSETS
MAKE COMPLETE GSM/DCS HANDSET
900MHz/
1.8GHz
AD20msp430 SOFTFONE™ CHIPSET
FLASH
MEMORY
AD6522
AD6521
DSP
MEMORY
ADSP-218x
DSP
DMA
I/O
1 Mbit
SRAM
ARM7®
CONTROLLER
DISPLAY
SIM
AUXILIARY
ADC, DAC
OTHELLO™ RADIO
AD6523
DIRECT
CONVERSION
RADIO
BASEBAND I/Q
ADC, DAC
AD6524
VOICEBAND
CODEC
MULTIBAND
SYNTHESIZER
KEYPAD
SPEAKER MIC
PA
POWER
CONTROL
ADP33xx
ADP34xx
POWER MANAGEMENT
BATTERY CHARGING
Figure 9.15
This is Analog Devices’ 4th generation of GSM chipsets, each of which has passed
numerous type- approvals and network operator approvals in OEM handsets. In
each generation, additional features have been added, while cost and power have
been reduced. Numerous power-saving features have been included in the
AD20msp430 chipset to reduce the total power consumption. A programmable state
machine allows events to be controlled with a resolution of one-quarter of a bit
period. This reduces current in standby mode to 1mA. This allows a handset to
operate in standby mode for up to 1000 hours on a battery charge. The AD20msp430
chipset uses the SoftFone™ architecture, where all software resides in RAM or
FLASH memory. Since ROM is not used, development time is reduced and
additional features can be field-installed easily. A basic dual-band GSM terminal
typically requires a single 8Mb FLASH memory chip.
There are two processors in the AD20msp430 chipset. The DSP processor is the
ADSP-218x core, proven in previous generations of GSM chipsets, and operated at
65MIPS in the AD20msp430. This DSP performs the voiceband and channel coding
functions previously discussed. The AD6521 voiceband/baseband codec chip contains
all analog and mixed-signal functions. These include the I/Q channel ADCs and
DACs, high performance multichannel voiceband codec, and several auxiliary ADCs
and DACs for AGC (automatic gain control), AFC (automatic frequency control),and
power- amplifier ramp control. The microcontroller is an ARM7 TDMI, running at
39MIPS. The ARM7 handles the protocol stack and the man-machine interface
functions. Both processors are field-proven in digital wireless applications.
9.22
DSP APPLICATIONS
The AD20msp430 chipset is fully supported by a suite of development tools and
software. The development tools allow easy customization of the DSP and/or ARM®
controller software to allow handset and terminal manufacturers to optimize the
feature set and user interface of the end equipment. Software is available from
Analog Devices’ software partner TTPCom for all layers, including both voice and
data applications, and is updated as new features become available. The system
DMA and interrupt controllers are designed to allow easy upgrades to future
generations of DSP and controller cores. The display interface can be used with
either parallel or serial-interface displays. System development can be shortened by
the use of the debugging features in the AD20msp430. Most critical signals can be
routed under software control to the Universal System Connector. This allows
system debugging to take place in the final form factor. In addition, the architecture
includes high speed logger and address trace functions in the DSP and single-wire
trace/debug in the ARM controller.
Analog Devices recently announced the revolutionary Othello™ direct-conversion
radio for mobile applications. By eliminating intermediate-frequency (IF) stages,
this chip set will permit the mobile electronics industry to reduce the size and cost
of radio sections and enable flexible, multi-standard, multi-mode operation. The
radio consists of two integrated circuits, the AD6523 Zero-IF Transceiver and the
AD6524 Multi-Band Synthesizer.
The AD6523 contains the main functions necessary for both a direct-conversion
receiver and a direct VCO transmitter, known as the Virtual-IF™ transmitter. It
also includes the local-oscillator generation block and a complete on-chip regulator
that supplies power to all active circuitry for the radio. The AD6524 is a fractionalN synthesizer that features extremely fast lock times to enable advanced data
services over cellular telephones-such as high-speed circuit-switched data (HSCSD)
and general packet radio services (GPRS).
Most digital cellular phones today include at least one "downconversion" in their
signal chain. This frequency conversion shifts the desired signal from the allocated
RF band for the standard (say, at 900 MHz) to some lower intermediate frequency
(IF), where channel selection is performed with a narrow channel-select filter
(usually a surface acoustic-wave (SAW) or a ceramic type). The now-filtered signal
is then further down-converted to either a second IF or directly to baseband, where
it is digitized and demodulated in a digital signal processor (DSP). Figure 9.16
shows the comparison between this superheterodyne architecture and the
superhomodyne™ architecture of the Othello radio receiver.
The idea of using direct-conversion for receivers has long been of interest in RF
design. The reason is obvious: in consumer equipment conversion stages add cost,
bulk, and weight. Each conversion stage requires a local oscillator, (often including
a frequency synthesizer to lock the LO onto a given frequency), a mixer, a filter, and
(possibly) an amplifier. No wonder, then, that direct conversion receivers would be
attractive. All intermediate stages are eliminated, reducing the cost, volume, and
weight of the receiver.
9.23
DSP APPLICATIONS
DIRECT CONVERSION RECEIVER
ARCHITECTURE ELIMINATES COMPONENTS
GSM
900
MHz
BPF
CERAMIC
I
ADC
BPF
LNA
IF
CERAMIC
SW
DCS
1.8
GHz
BPF
CERAMIC
BPF
VCO
I
ADC
BPF
SAW
sin
LNA
SW
DCS
1.8
GHz
VCO
PLL
cos
BPF
SAW
Q
ADC
PLL
DIRECT CONVERSION (SUPERHOMODYNE™)
PLL
cos
CERAMIC
SUPERHETERODYNE
GSM
900
MHz
VCO
SAW
BPF
LNA
sin
LNA
Q
ADC
Figure 9.16
The Othello™ radio reduces the component count even more by integrating the
front-end GSM low-noise amplifier (LNA). This eliminates an RF filter (the "image"
filter) that is necessary to eliminate the image, or unwanted mixing product of a
mixer and the off chip LNA. This stage, normally implemented with a discrete
transistor, plus biasing and matching networks, accounts for a total of about 12
components. Integrating the LNA saves a total of about 15 to 17 components,
depending on the amount of matching called for by the (now-eliminated) filter.
A functional block diagram of the Othello™ dual band GSM radio's architecture is
shown in Figure 9.17. The receive section is at the top of the figure. From the
antenna connector, the desired signal enters the transmit/receive switch and exits
on the appropriate path, either 925-960 MHz for the GSM band or 1805-1880 MHz
for DCS. The signal then passes through an RF band filter (a so-called "roofing
filter") that serves to pass the entire desired frequency band while attenuating all
other out-of-band frequencies (blockers-including frequencies in the transmission
band) to prevent them from saturating the active components in the radio front end.
The roofing filter is followed by the low-noise amplifier (LNA). This is the first gain
element in the system, effectively reducing the contribution of all following stages to
system noise. After the LNA, the direct-conversion mixer translates the desired
signal from radio frequency (RF) all the way to baseband by multiplying the desired
signal with a local oscillator (LO) output at the same frequency.
9.24
DSP APPLICATIONS
SUPERHOMODYNE™ DIRECT CONVERSION DUAL-BAND
TRANSCEIVER USING THE AD6523/AD6524 CHIPSET
AD6523
BPF
(SAW)
GSM/
DCS
VGA
SW
VGA
BPF
(SAW)
BPF
VCO
TX/RX
SWITCH
I
ADC
BPF
LNA
Q
ADC
LO
GENERATOR
PLL
AD6524
DIV N
I
DAC
VCO
PA
PHASE
DETECT
DIV M
Σ
Q
DAC
Figure 9.17
The output of the mixer stage is then sent in quadrature (I and Q channels) to the
variable-gain baseband amplifier stage. The VGA also provides some filtering of
adjacent channels, and attenuation of in-band blockers. These blocking signals are
other GSM channels that are some distance from the desired channel, say 3 MHz
and beyond. The baseband amplifiers filter these signals so that they will not
saturate the Receive ADCs. After the amplifier stage, the desired signal is digitized
by the Receive ADCs.
The Transmit section begins on the right, at the multiplexed I and Q
inputs/outputs. Because the GSM system is a time division duplex (TDD) system,
the transmitter and receiver are never on at the same time. The Othello™ radio
architecture takes advantage of this fact to save four pins on the transceiver IC's
package. The quadrature transmit signals enter the transmitter through the
multiplexed I/Os. These I and Q signals are then modulated onto a carrier at an
intermediate frequency greater than 100 MHz.
The output of the modulator goes to a phase-frequency detector (PFD), where it is
compared to a reference frequency that is generated from the external channel
selecting LO. The output of the PFD is a charge pump, operating at above 100 MHz,
whose output is filtered by a fairly wide (1MHz) loop filter. The output of the loop
filter drives the tuning port of a voltage-controlled oscillator (VCO), with frequency
ranges that cover the GSM and DCS transmit bands.
9.25
DSP APPLICATIONS
The output of the transmit VCO is sent to two places. The main path is to the
transmit power amplifier (PA), which amplifies the transmit signal from about
+3dBm to +35dBm, sending it to the transmit/receive switch and low pass filter
(which attenuates power-amplifier harmonics). The power amplifiers are dual band,
with a simple CMOS control voltage for the band switch. The VCO output also goes
to the transmit feedback mixer by means of a coupler, which is either a printed
circuit, built with discrete inductors and capacitors, or a monolithic (normally
ceramic) coupling device. The feedback mixer downconverts the transmit signal to
the transmit IF, and uses it as the local oscillator signal for the transmit modulator.
This type of modulator has several names, but the most descriptive is probably
"translation loop". The translation loop modulator takes advantage of one key
aspect of the GSM standard: the modulation scheme is Gaussian-filtered minimumshift keying (GMSK). This type of modulation does not affect the envelope
amplitude, which means that a power amplifier can be saturated and still not
distort the GMSK signal sent through it.
GMSK can be generated in several different ways. In another European standard
(for cordless telephones), GMSK is created by directly modulating a free runningVCO with the Gaussian filtered data stream. In GSM, the method of choice has
been quadrature modulation. Quadrature modulation creates accurate phase
GMSK, but imperfections in the modulator circuit (or up-conversion stages) can
produce envelope fluctuations, which can in turn degrade the phase trajectory when
amplified by a saturated power amplifier. To avoid such degradations, GSM phone
makers have been forced to use amplifiers with somewhat higher linearity, at the
cost of reduced efficiency and talk time per battery charge cycle.
The translation loop modulator combines the advantages of directly modulating the
VCO and the inherently more accurate quadrature modulation. In effect, the
scheme creates a phase locked loop (PLL), comprising the modulator, the LO signal,
and the VCO output and feedback mixer. The result is a directly modulated VCO
output with a perfectly constant envelope and almost perfect phase trajectory.
Phase trajectory errors as low as 1.5 degrees have been measured in the AD6523
transceiver IC, using a signal generator as the LO signal to provide a reference for
the loop.
Because Othello™ radios can be so compact, they enable GSM radio technology to
be incorporated in many products from which it has been excluded, such as very
compact phones or PCMCIA cards. However, the real power of direct conversion will
be seen when versatile third-generation phones are designed to handle multiple
standards. With direct-conversion, hardware channel-selection filters will be
unnecessary, because channel selection is performed in the digital signal-processing
section, which can be programmed to handle multiple standards. Contrast this with
the superheterodyne architecture, where multiple radio circuits are required to
handle the different standards (because each will require different channel-selection
filters), and all the circuits will have to be crowded into a small space. With direct
conversion, the same radio chain could in concept be used for several different
standards, bandwidths, and modulation types. Thus, Web-browsing and voice
services could, in concept, occur over the GSM network using the same radio in the
handset.
9.26
DSP APPLICATIONS
ANALOG CELLULAR BASESTATIONS
Consider the analog superheterodyne receiver invented in 1917 by Major Edwin H.
Armstrong (see Figure 9.18). This architecture represented a significant
improvement over single-stage direct conversion (homodyne) analog receivers which
had previously been constructed using tuned RF amplifiers, a single detector, and
an audio gain stage. (Note that the homodyne technique has now gained favor in
DSP-based receivers as explained above). A significant advantage of the
superhetrodyne analog receiver is that it is much easier and more economical to
have the gain and selectivity of a receiver at fixed intermediate frequencies (IF)
than to have the gain and frequency-selective circuits "tune" over a band of
frequencies.
U.S. ADVANCED MOBILE PHONE SERVICE (AMPS)
SUPERHETERODYNE ANALOG BASESTATION RECEIVER
AMPS: 416 CHANNELS ("A" OR "B" CARRIER)
30kHz WIDE, FM
12.5MHz TOTAL BANDWIDTH
1 CALLER/CHANNEL
RF
BPF
LNA
LO1
TUNED
70MHz
LO2
FIXED
10.7MHz
LO3
FIXED
455kHz
ANALOG
DEMOD,
FILTER
1ST IF
2ND IF
SAME AS ABOVE
CHANNEL 1
30 kHz
3RD IF
CHANNEL n
30 kHz
Figure 9.18
The frequencies shown in Figure 9.18 correspond to the AMPS (Advanced Mobile
Phone Service) analog cellular phone system currently used in the U.S. The receiver
is designed for AMPS signals at 900MHz RF. The signal bandwidth for the "A" or
"B" carriers serving a particular geographical area is 12.5MHz (416 channels, each
30kHz wide). The receiver shown uses triple conversion, with a first IF frequency of
70MHz, a second IF of 10.7MHz, and a third of 455kHz. The image frequency at the
receiver input is separated from the RF carrier frequency by an amount equal to
twice the first IF frequency (illustrating the point that using relatively high first IF
frequencies makes the design of the image rejection filter easier).
The output of the third IF stage is demodulated using analog techniques
(discriminators, envelope detectors, synchronous detectors, etc.). In the case of
AMPS the modulation is FM. An important point to notice about the above scheme
9.27
DSP APPLICATIONS
is that there is one receiver required per channel, and only the antenna, prefilter,
and LNA can be shared.
It should be noted that in order to make the receiver diagrams more manageable,
the interstage amplifiers are not shown. They are, however, an important part of
the receiver, and the reader should be aware that they must be present.
It should be seen by now that analog receiver design is a complicated art, and there
are many tradeoffs that can be made between IF frequencies, single-conversion vs.
double-conversion or triple conversion, filter cost and complexity at each stage in
the receiver, demodulation schemes, etc. There are many excellent references on the
subject, and the purpose of this discussion is to form an historical frame of reference
for the following discussions on the application of digital techniques in the design of
advanced communications and basestation receivers.
DIGITAL CELLULAR BASESTATIONS
Cellular telephone basestations form the backbone of the modern wireless cellular
infrastructure. They must receive multiple calls, process the calls, and re-transmit
them. Handoffs to basestations in adjacent cells must be done seamlessly with
respect to the mobile customer. In addition, the basestation must often handle
multiple standards simultaneously. Many carriers in certain areas of the United
States use multiple technologies in the same geographical area; AMPS and CDMA,
for example.
Flexibility, performance, and low per-channel costs are the keys to modern
basestations. Maximizing the use of DSP in the tranceiver allows multiple
standards to be handled without the necessity for changing hardware. This has led
to the widespread proliferation of so-called software radios which dominate the
current basestation market.
As in the case of the cell phone handset, direct conversion techniques are used in
basestations. The signal is digitizied by a high performance wideband ADC after
only one stage of downconversion. Figure 9.19 shows two fundamental approaches
to a digital receiver: narrowband and wideband.
By narrowband, we mean that sufficient prefiltering has been done such that all
undesired signals have been eliminated and that only the channel of interest is
presented to the ADC input. Wideband simply means that a number of channels are
presented to the input of the ADC and further filtering, tuning, and processing is
performed digitally. Usually, a wideband receiver is designed to receive an entire
band; cellular or other similar wireless service such as PCS or CDMA. In fact, one
wideband digital receiver can be used to receive all channels within the band
simultaneously, allowing almost all of the analog hardware (including the ADC) to
be shared among all the channels.
9.28
DSP APPLICATIONS
NARROWBAND AND WIDEBAND DIGITAL RECEIVERS
FOR CELLULAR BASESTATIONS
NARROWBAND
TUNED
LO
BPF
RF
FRONT
END
ADC
DIGITAL
DECIMATION
FILTER
DSP
CHANNEL
1
ADC
DIGITAL
DECIMATION
FILTER
DSP
CHANNEL
n
DIGITAL
CHANNELIZER
DSP
CHANNEL
1
DIGITAL
CHANNELIZER
DSP
CHANNEL
n
TUNED
LO
IF
BPF
BW: 30-200kHZ
WIDEBAND
FIXED
LO
RF
FRONT
END
IF
BPF
BW: 5-25MHz
ADC
Figure 9.19
The wideband approach places severe constraints on the ADC and requires high
spurious free dynamic range (SFDR) and signal-to-noise ratio (SNR), especially in
cellular applications where signal strengths of channels can differ by more than
100dB. This requires ADCs with bandwidths of greater than 100MHz and sampling
frequencies greater than 50MSPS (required to handle a multicarrier bandwidth of
25MHz, for example). On the other hand, the narrowband approach provides more
processing gain because each channel is highly oversampled, but this approach also
requires more ADCs to process the same number of channels.
ADI's SoftCell™ chipset addresses key issues for wireless operators including the
cost of coverage, flexibility and size, as well as quality of service. Basestations
containing the SoftCell chip set are easy to modify in terms of adding services,
additional channels, and changing wireless standards incrementally. In effect,
operators will have the ability to use and move between any air interface standard
(e.g., GSM, PHS, D-AMPS), deploy a higher number of channels, and offer
frequency plans with greater efficiency. The new architecture also eliminates
redundant channel radios for both transmitters and receivers.
The SoftCell chip set is optimized for four RF carrier channels and is easily
expandable. This solution enables equipment manufacturers to create highly
scalable multicarrier, multimode base stations at a fraction of the cost of traditional
multichannel base stations using analog techniques. A block diagram of a system
using the SoftCell chipset is shown in Figure 9.20.
9.29
DSP APPLICATIONS
BASESTATION BLOCK DIAGRAM
USING SoftCell™ MULTICARRIER WIDEBAND CHIPSET
RX
fs
= 65MSPS
SFDR = 100dB
BW = 200MHz
SNR = 77dB
RF
FRONT
END
BPF
LO
CHANNELS
WIDEBAND
14-BIT ADC
4 CHANNEL
RX DIGITAL
PROCESSOR
AD6644
AD6624
TX
fs
= 65MSPS
SFDR = 80dB
RF
OUTPUT
BPF
LO
TigerSHARC
DSPs
WIDEBAND
14-BIT DAC
4 CHANNEL
TX DIGITAL
PROCESSOR
AD9772
AD6622
Figure 9.20
The decrease in size and cost savings associated with SoftCell means basestations
can be deployed in higher numbers and in tighter locations. The result is better
coverage, better quality, and fewer busy signals for users. The deployment
capability also makes the SoftCell chip set well-suited for wireless office
applications. In addition, the software radio technology implemented in the chipset
enables new applications such as Smart Antenna or phased-array antenna services
that enable power efficiencies and cost savings, and small picocell installations for
additional coverage capacity and in-building wireless systems.
The SoftCell chip set consists of ADI's AD6644 14-bit ADC, AD6624 quad digital
receive signal processor (RSP), AD9772 TxDAC® 14-bit DAC, and the AD6622 quad
digital transmit signal processor (TSP). The solution leverages the benefits of digital
signal processors to enable channel separation, equalization, error correction and
decoding with greater flexibility and efficiencies. This new chip set is optimized to
work with ADI's TigerSharc™ multiprocessor digital signal processor (DSP).
The TigerSHARC DSP is a ground breaking processor optimized for
communications applications capable of performing approximately 1 billion 16-bit
multiply/accumulates per second at 150 MHz. Additionally, the TigerSHARC is
unique in its ability to support 8-, 16-, and 32-bit word sizes on a single chip.
Modulation/demodulation, channel encode/decoding, and other radio functions can
be multiplexed to support multiple carriers in a single DSP.
SoftCell complements ADI's recently introduced AD6600 diversity receiver ADC.
The AD6600 addresses narrowband applications where a multicarrier architecture
is still not feasible but allows direct IF sampling of signals at frequencies up to
250MHz. Configured with the appropriate digital receiver signal processor, the
AD6600 can address a variety of air interface standards, including GSM Macrocell.
9.30
DSP APPLICATIONS
Classic basestation architectures require a complete transceiver for every RF carrier
processed (from 4 to 80 channels for digital and analog systems respectively). These
radios must be duplicated for diversity and sectorized antennas as well. It is easy to
see why base station electronics consume so much space, power, and cost. The
advantage of a multicarrier software radio is the elimination of redundant radios in
favor of a single, high-performance radio per antenna, where each RF carrier is
processed in the digital domain. Deployment of true software radios has been
limited by the performance of analog-to-digital converters (ADC) which must
digitize the enormous dynamic range demanded in a spectrum composed of multiple
carriers, blockers, and adjacent channel interference.
Multicarrier transmitters have similar challenges to meet the performance
demands of the newest air interface specifications. Digital-to-analog converters
(DAC) and multicarrier-power-amplifiers (MCPAs) must preserve the spectrum of
several digitally generated carriers without corruption or spurious signal generation
in adjacent channels. The AD9772 is a 14-bit interpolating DAC optimized to
accurately convert multiple carriers to a single IF frequency. The AD9772 is the
latest in ADI's TxDAC® family of high-performance converters.
The heart of ADI's SoftCell chip set is the AD6644, a 14-bit, 65-MSPS ADC that
provides up to 100dB of spurious-free-dynamic-range (SFDR) and 77dB signal-tonoise ratio (SNR). This provides the receiver performance needed to implement a
multicarrier digitizing radio for many applications. Shifting the channel tuning,
filtering, and demodulation to the digital domain allows the flexibility to support
different air interface standards, number of channels, and frequency plans with a
single radio design.
Following the ADC, a digital receive signal processor (RSP) performs the channel
tuning, filtering, and decimation required to provide baseband I and Q signals to a
digital signal processor (DSP). The AD6624 is an 65MSPS Quad Digital RSP
developed to support GSM, IS136, and other narrowband standards. The AD6624's
four channels are independently programmable to change air interface
characteristics on demand. It is a simple matter to add AD6624s in parallel for
additional channel capacity. The AD6624 can also be configured to support EDGE
extensions of GSM and IS136.
The AD6622 is a four channel Digital Transmit Signal Processor that takes
baseband I and Q inputs from a DSP and provides all the signal processing
functions required to drive the AD9772 DAC. Each independent channel can be
programmed to provide the desired channel filtering for most air interface
standards. The AD6622 supports IS95 and WCDMA standards and can be daisychained to combine an arbitrary number of channels onto a single 18-bit digital
output.
9.31
DSP APPLICATIONS
MOTOR CONTROL
Long known for its simplicity of construction, low-cost, high efficiency and long-term
dependability, the AC induction motor has been limited by the inability to control
its dynamic performance in all but the crudest fashion. This has severely restricted
the application of AC induction motors where dynamic control of speed, torque and
response to changing load is required. However, recent advances in digital signal
processing (DSP) and mixed-signal integrated circuit technology are providing the
AC induction motor with performance never before thought possible. Manufacturers
anxious to harness the power and economy of Vector Control can reduce R&D costs
and time to market for applications ranging from industrial drives to electric
automobiles and locomotives by utilizing a standard chipset/development system.
It is unlikely that Nikola Tesla (1856-1943), the inventor of the induction motor,
could have envisaged that this workhorse of industry could be rejuvenated into a
new class of motor that is competitive in most industrial applications.
Before discussing the advantages of Vector Control it is necessary to have a basic
understanding of the fundamental operation of the different types of electric motors
in common use.
Until recently, motor applications requiring servo-control tasks such as tuned
response to dynamic loads, constant torque and speed control over a wide range
were almost exclusively the domain of DC brush and DC permanent magnet
synchronous motors. The fundamental reason for this preference was the
availability of well understood and proven control schemes. Although easily
controlled, DC brush motors suffer from several disadvantages; brushes wear and
must be replaced at regular intervals, commutators wear and can be permanently
damaged by inadequate brush maintenance, brush/commutator assemblies are a
source of particulate contaminants, and the arcing of mechanical commutation can
be a serious fire hazard is some environments.
The availability of power inverters capable of controlling high-horsepower motors
allowed practical implementation of alternate motor architectures such as the DC
permanent magnet synchronous motor (PMSM) in servo control applications.
Although eliminating many of the mechanical problems associated with DC brush
motors, these motors required more complex control schemes and suffered from
several drawbacks of their own. Aside from being costly, DC PMSMs in larger, highhorsepower configurations suffer from high rotor moment-of-inertia as well as
limited use in high speed applications due to mechanical constraints of rotor
construction and the need to implement field weakening to exceed baseplate speed.
In the 1960's, advances in control theory, in particular the development of indirect
field-oriented control, provided the theoretical basis for dynamic control of AC
induction motors. Indirect field-oriented control makes use of reference frame
theory. Using these techniques, it is possible to transform the phase variable
machine description of a motor to another reference frame. By judicious choice of
the reference frame, it is possible to simplify considerably the complexity of the
mathematical machine model. While these techniques were initially developed for
the analysis and simulation of AC motors, they are now invaluable tools in the
9.32
DSP APPLICATIONS
digital control of such machines. As digital control techniques are extended to the
control of the currents, torque, and flux of such machines, the need for compact,
accurate motor models is obvious.
Fortunately, the theory of reference frames is equally applicable to the synchronous
machines, such as the Permanent Magnet Synchronous Machine (PMSM). This
motor is sometimes known as the sinusoidal brushless motor or the brushless AC
machine, and is very popular as a high performance servo drive.
Because of the intensive mathematical computations required by indirect fieldoriented control, now commonly referred to as vector control or reference frame
theory, practical implementation was not possible for many years. Available
hardware could not perform the high-speed precision sensing of rotor position and
near real-time computation of dynamic flux vectors. The current availability of
precision optical encoders, isolated gate bipolar transistors (IGBTs), high-speed
resolver-to-digital converters and high-speed digital signal processors (DSPs) has
pushed vector control to the forefront of motor development due to the advantages
inherent in the AC induction motor.
A simplified block diagram of an AC induction motor control system is shown in
Figure 9.21. The inputs to the controller are the motor currents (normally threephase) and the motor rotor position and velocity. Hall-effect sensors are often used
to monitor the currents, and a resolver and a resolver-to-digital converter (RDC)
monitor the rotor position and velocity. The DSP is used to perform the real time
vector-type calculations necessary to generate the control outputs to the inverter
processors. The transformations required for reference frame transformations and
vector control are also accomplished with the DSP.
MOTOR CONTROL BLOCK DIAGRAM
CONTROLLER
VECTOR
TRANSFORM
PROCESSOR
ADSP-21xx
DSP
PWM
ADMC200/
ADMC201
POWER
STAGE
(INVERTER)
ADCs
AD7861/
2/3/4/5
MOTOR CURRENTS
POSITION, VELOCITY
HOST
COMPUTER
AC
MOTOR
RESOLVER TO
DIGITAL
CONVERTER
RESOLVER
AD2S80A/82A/83/90
Figure 9.21
9.33
DSP APPLICATIONS
The functions in the controller block have been integrated into the Analog Devices’
ADMC300, ADMC331, ADMC401, and the ROM-based ADMC326 and ADMC328
DSP motor controller chips. These devices include the peripheral circuitry such as
ADCs, voltage references, PWM controllers, timers, etc., required to perform all the
functions shown in Figure 9.21.
The latest members of the family are the ADMCF326 and ADMCF328 motor
controllers, referred to as DashDSP™, meaning digital plus analog plus flash
memory (see Figure 9.22). The use of flash memory allows the devices to be field
programmable, thereby providing greater flexibility and shortening development
time. These controllers include a 20MIPS 16-bit fixed-point core based on the ADSP217x family architecture. The memory consists of a 512 x 24-bit program memory
RAM, a 512 x 16-bit data memory RAM, 4K x 24-bit program memory ROM, and a
4K x 24-bit programmable flash memory. A completely integrated ADC analog
subsystem allows the three-phase motor currents to be monitored. A 16-bit 3-phase
PWM generates the control signals to the external inverter power stage. The parts
are packaged in a 28-pin SOIC or PDIP package. A block diagram of the ADMCF328
is shown in Figure 9.23.
FULLY INTEGRATED MOTOR CONTROL
WITH DashDSP™
DashDSP™
ADSP-217x
FAMILY
DSP CORE
Other
Peripherals
PWM
POWER
STAGE
(INVERTER)
POR
VREF
PIO
ADCs
FLASH
MEMORY
MOTOR CURRENTS
POSITION, VELOCITY
HOST
COMPUTER
RESOLVER TO
DIGITAL
CONVERTER
AD2S80A/82A/83/90
Figure 9.22
9.34
AC
MOTOR
RESOLVER
DSP APPLICATIONS
ADMCF328 DSP MOTOR CONTROLLER
WITH FLASH MEMORY
MEMORY BLOCK
ADSP-217x
20 MIPS
CORE
DATA
ADDRESS
GENERATORS
PROGRAM
SEQUENCER
DAG 1 DAG 2
PROGRAM
FLASH
4K X 24
PROGRAM
ROM
4K X 24
PROGRAM
RAM
512 X 24
DATA
MEMORY
512 X 16
VREF
2.5V
5
ANALOG
INPUTS
ISENSE
AMP
& TRIP
16-BIT
3-PHASE
PWM
9-BIT
PIO
2 X 8-BIT
AUX
PWM
WATCHDOG
TIMER
PMA
DMA
PMD
DMD
ARITHMETIC UNITS
ALU
MAC
■
SHIFTER
POR
TIMER
SERIAL PORT
SPORT 1
New ADMCF5xx Family Based on ADSP-219x 150 MIPS Core
with 128K Bytes Flash Memory
Figure 9.23
Third-party DSP software, reference designs, and evaluation systems are available
to facilitate motor control system development using these chips.
Analog Devices recently announced a new ADMCF5xx DashDSP family based on
the ADSP-219x, 150MIPS 16-bit fixed point core. At 0.4mA/MIP, this solution
provides engineers a low power and wide ranging speed device. The integration of
up to 128KBytes of embedded flash memory eases development and enables flexible
system software modifications and system upgrades. This also allows users to boot
directly from flash memory into RAM. With 10-, 12-, and 14-bit ADCs, customers
can select a device that matches their system sensing specifications and cost
requirements. The family is supported by ADI’s VisualDSP tools, which incorporate
the industry’s first C++ compiler. The ADMCF5xx series is augmented by a wide
range of high performance peripherals, such as 3 to 6 phase 16-bit PWMs with
single or dual encoder interfaces for multi axis motor control. Motor current
sampling can be achieved through the use of DC link or inverter shunt techniques,
with user programmable converter sampling times. A precision voltage reference, a
power-on reset circuit, and auxiliary PWMs capable of implementing power factor
correction are also included. Additionally, various communication peripherals such
as a full CAN bus, UARTs, SPORTs, and a JTAG interface have been integrated
into the family.
9.35
DSP APPLICATIONS
CODECS AND DSPS IN VOICEBAND AND AUDIO
APPLICATIONS
In voiceband and audio applications such as hands-free car telephone kits and
modems, analog front ends (AFEs), also called codecs, make excellent system
building blocks along with DSPs.
Analog Devices recently announced the ADSP-21ESP202-series specifically
designed for embedded speech processing applications such as hands-free car kits.
This part is based on the AD73322 dual AFE and the ADSP-218x 16-bit fixed point
core. 40KBytes of on-chip RAM is configured as 8K x 24-bit program memory RAM,
and 8K x 24-bit data memory RAM. 24KBytes of on-chip ROM is configured as 8K x
24-bit program memory ROM.
A full-duplex hands-free car kit can be implemented with the ADSP-21ESP202 as
shown in Figure 9.24. This series represents a breakthrough in both the level and
diversity of functional integration.
ADSP-21ESP202 SINGLE-CHIP
HANDS-FREE CAR KIT SOLUTION
ADSP-218x 49MIPs Core
Serial Port 1
BDMA Boot
Serial Port 0
Dual Voice Band Codecs
Microphone
Fully Programmable
Input Gain & Output
Attenuation
Speaker
AMP
Analog
Connection
Digital
Connection
Cellular
Handset
Non-volatile Memory
Speech Prompts
Vocabularies
ADSP-21ESP202
decreases Hands-free IC
count from 4 to 1!!!
Figure 9.24
Specifically, the ADSP-21ESP202-series is the first general purpose ADSP-218xbased product to include analog functionality (see Figure 9.25). All members of the
series contain two 16-bit sigma-delta voiceband codecs which are fully
programmable in sampling rate (up to 64kSPS), input gain, and output gain. There
are two analog comparators which can be used to implement voice detection
algorithms and detect cable insertion/de-insertion events as well as generate
processor interrupts. Dual current sources with switches are included which can be
9.36
DSP APPLICATIONS
configured to be driven by the PWM functionality of the enhanced timer. They have
both trickle and rapid charge settings. The current switches can also be used in
automatic gain control (AGC) algorithms and also for input clock frequency/phase
adjustments.
ADSP-21ESP202 OFFERS
GENERAL PURPOSE ADSP-218x 'FIRSTS'
ADSP-218x
16-Bit DSP
Core
Byte DMA
Controller
Internal DMA
Port
Dual Analog Comparators
8Kx24
Program Memory
SRAM
8Kx16
Data Memory
SRAM
Voice Band Codec #1
Voice Band Codec #2
High Speed
Serial Port
#1
High Speed
Serial Port
#2
8Kx24
Program Memory
ROM
3.3V
Operation!
■ First Integration of analog circuitry
◆ Dual programmable (to 64KSPS)
voice band codecs -- 75dB SNR
◆ Dual analog comparators
◆ Dual current sources
■ First integration of late stage
programmable ROM
◆ Customer defined, or
◆ Standard functions
■ Additional features include
◆ 49MHZ operation @ 3.3 Volts
◆ 8K PM RAM, 8K DM RAM
◆ 8K PM ROM
◆ Extended Interrupts & Flags
◆ Enhanced Timer
■ 128 Pin LQFP package
Figure 9.25
The ADSP-21ESP202 also contains a programmable block of 8K words program
memory ROM. Analog Devices will offer multiple versions of the series with
application specific algorithms embedded into this space. Analog Devices will also
accept customer defined ROM codes.
The system shown in Figure 9.24 provides a DSP engine for noise/echo cancellation
and speech recognition, codecs to interface to the speaker/microphone and cellular
terminal, and non-volatile memory for program storage and speech recognition
vocabularies and prompts. The ADSP-21ESP202 series integrates all these
components in a single package providing a single-chip solution for full-duplex
hands-free car kits. This represents a 75% reduction in IC count over the previous
generation solutions.
PC applications such as voice processing and modems also require high performance
codecs. Figure 9.26 shows a generalized audio/modem application based on the
AD1819B SoundPort® codec.
This codec is fully compliant with the AC’97 specification (Audio Codec ’97,
Component Specification, Revision 1.03, © 1996, Intel Corporation). In addition, the
AD1819 supports multiple codec configurations (up to three per AC Link), a DSP
9.37
DSP APPLICATIONS
serial mode, variable sample rates, modem sample rates and filtering, and built-in
Phat™ Stereo 3D enhancement.
GENERALIZED PC AUDIO, MODEM APPLICATION
SERIAL
DATA
MICROPHONES
AC'97
SoundPort®
CODEC
SPEAKERS
ADSP-218x
16-BIT
FIXED POINT
DSP
EXTERNAL
MEMORY
(OPTIONAL)
PARALLEL
BUS
PC
INTERFACE
LOGIC
PC
AD1885
ANALOG
DATA
AUX LINE-IN
MONITOR
DIRECT ACCESS
ARRANGEMENT
(DAA)
TWISTED PAIR
TELEPHONE LINE
AUX LINE-IN
■ 16-BIT RESOLUTION
■ fs = 7 TO 48 kSPS
■ 90dB THD + N
Figure 9.26
The AD1819B is an analog front end for high performance PC audio, modem, or
DSP applications. The main architectural features of the AD1819B are the high
quality analog mixer section, two channels of 16-bit sigma-delta A/D conversion, two
channels of 16-bit sigma-delta D/A conversion, and a serial port. THD + N is 90dB,
and the sampling frequency is variable from 7kSPS to 48kSPS.
The Analog Devices’ 32-bit floating point SHARC DSP has demonstrated the
highest Dolby Digital AC-3 audio quality. The DSP architecture of the reference
design shown in Figure 9.27 uses the ADSP-21065L SHARC and an integrated
mixed signal IC (AD1836) to provide low cost, high quality, multi-channel audio.
Key applications include A/V receivers for home theater and high-end automotive
audio. The AD1836 provides all the mixed signal functions including 4 ADC input
channels and 6 DAC output channels. The AD1836 codec provides 97dB THD + N
and 105dB SNR for high quality audio. While a fixed-function DSP could be used, a
programmable DSP provides greater flexibility. This reference design can be
programmed for MP3, Dolby Digital AC-3, THX, or DTS decoding. Other audio
processing algorithms can easily be added through additional programs.
9.38
DSP APPLICATIONS
AUTOMOTIVE AND HOME THEATER
REFERENCE DESIGN USING 32-BIT SHARC
RAM
ROM
HOST
ADSP-21065L
SHARC™
■ THD + N = 97dB
= 105dB
■ SNR
■ Up to 96kSPS Sampling Rate
DAC
(AUDIO)
AUDIO IN
ADC
I/F
DAC
ADC
DAC
AD1836
Figure 9.27
In large digital audio systems it is often necessary to distribute the signal
processing tasks between multiple processors. Figure 9.28 shows a 16-channel
mixer that uses two ADSP-21160s. The data comes from sixteen 24-bit ADCs and is
passed to the FPGA. This FPGA converts the serial ADC data into parallel data and
sends it to the two ADSP-21160’s external ports. The external port on each DSP has
hardware support to enable the FPGA to write to both DSPs at once. DMA engines
inside the DSP receive this data and move it to desired locations in internal
memory. The hardware support and DMA engines reduce the complexity of the
FPGA design, because the FPGA is simply required to drive the data onto the bus.
No arbitration logic or address generation is necessary in the FPGA.
The DSPs perform various algorithms, such as mixing, left/right panning,
equalization, and additional effects processing, such as reverberation or
compression/expansion. The audio output samples are then transmitted to a 24-bit
stereo DAC. These algorithms could be mapped in such a way that one of the DSPs
does the mixing and effects, while the other does the equalization. Another
possibility is to give each DSP the task of processing half of the data channels. The
optimal algorithmic mapping depends on what type of processing is required.
For this example, it is estimated that the two ADSP-21160s have enough
computation power to perform various algorithms on 16 channels of data sampled at
48kSPS. In a 20µs period each DSP is capable of executing 2000 instructions in the
core. If each DSP is responsible for half of the channels of data (8 channels), that
means that the DSP can allocate 250 instructions to each channel.
9.39
DSP APPLICATIONS
With the ADSP-21160 SIMD architecture, there are roughly enough instructions for
a 3-band (high, mid, and low) equalizer, mixing, delay effects, and compression for
each channel. The movement of data into the memory does not enter into the
processing loading calculation since it is a zero-overhead task.
16 CHANNEL AUDIO MIXER
USING ADSP-21160 SHARC PROCESSORS
AUDIO OUTPUT
DAC
INPUT DATA
3
ADC #1
.
.
.
.
SERIAL
.
DATA
.
TO
.
PARALLEL
.
DATA
.
(FPGA)
.
.
.
.
ADC #16
TDM
SERIAL PORT
ADSP-21160
ADSP-21160
2 LINK PORTS
64
PARALLEL DATA BUS
■
■
■
■
All 16 channels are broadcast to each ADSP-21160
ADSP-21160s perform mixing, equalization, and audio effects
Link ports pass shared data between ADSP-21160s
Processed data sent to DAC via time division multiplexing
on serial ports
Figure 9.28
A SIGMA-DELTA ADC WITH PROGRAMMABLE DIGITAL
FILTER
Most sigma-delta ADCs have a fixed internal digital filter. The filter’s cutoff
frequency (and the ADC output data rate) scales with the master clock frequency.
The AD7725 is a 16-bit sigma-delta ADC with a programmable internal digital
filter. The block diagram in Figure 9.29 shows that the modulator operates at a
maximum oversampling rate of 19.2MSPS. The modulator is followed by a preset
FIR filter which decimates the modulator output by a factor of 8, yielding an output
data rate of 2.4MSPS. The response of the preset FIR filter is shown in Figure 9.29.
The output of the preset filter drives a programmable digital filter. The diagram
shows typical response for a 300kHz lowpass FIR filter.
9.40
DSP APPLICATIONS
AD7725 16-BIT SIGMA-DELTA ADC
WITH PROGRAMMABLE DIGITAL FILTER
CLKIN = 19.2MSPS
INPUT
16-BIT
SIGMADELTA
MODULATOR
2.5V
REFERENCE
2.4MSPS
1.2MSPS (MAX)
PARALLEL/
SERIAL
INTERFACE
PROGRAMMABLE
FILTER
PRESET
FIR FILTER
0
PRESET
FIR
FILTER
dB
0
TYPICAL
FIR
FILTER
RESPONSE
dB
–70
–90
PROGRAMMABLE FILTER:
■
■
■
■
■
■
■
0
1.2
FREQUENCY (MHz)
INPUT DATA RATE = CLKIN/8
COEFFICIENT PRECISION = 24 BITS
ARITHMETIC PRECISION = 30 BITS
MAXIMUM NUMBER OF TAPS = 108
DECIMATION FACTOR = 2 MIN, 256 MAX
NUMBER OF DECIMATION STAGES = 5 MAX
OUTPUT DATA RATE = CLKIN/16 MAX
2.4
0
0.3
0.6
FREQUENCY (MHz)
SFDR = 90dB
SNR = 84dB
Figure 9.29
The programmable filter is flexible with respect to number of taps and decimation
rate. The filter can have up to 108 taps, up to 5 decimation stages, and a decimation
factor between 2 and 256. Coefficient precision is 24-bits, and arithmetic precision is
30-bits.
The AD7725 contains Systolix’s PulseDSP™ (trademark of Systolix) post processor
which permits the filter characteristics to be programmed through the parallel or
serial microprocessor interface.
The post processor is a fully programmable core which provides processing power of
up to 130 million multiply-accumulates (MAC) per second. To program the post
processor, the user must produce a configuration file which contains the
programming data for the filter function. This file is generated by the FilterWizard
compiler which is available from Analog Devices. The AD7725 compiler accepts
filter coefficient data as an input and automatically generates the required device
programming data.
The coefficient file for the filter response can be generated using a digital filter
design package such as Systolix FilterExpress™ (http://www.systolix.co.uk) or
QEDesign™ from Momentum Data Systems (http://www.mds.com). The response of
the filter can be plotted so the user knows the response before generating the filter
coefficients. The data is available to the processor at a 2.4MSPS rate. When
decimation is employed in a multistage filter, the first filter will be operated at
2.4MSPS, and the user can then decimate between stages. The number of taps
which can be contained in the processor is 108. Therefore, a single filter with 108
9.41
DSP APPLICATIONS
taps can be generated, or a multistage filter can be designed whereby the total
number of taps adds up to 108. The filter characteristic can be lowpass, highpass,
bandstop, or bandpass and can be either FIR or IIR.
The AD7725 operates on a single +5V supply, has an on-chip 2.5V reference, and is
packaged in a 44-pin PQFP. Power dissipation is approximately 350mW when
operating at full power. A half-power mode is available with a master clock
frequency of 10MSPS maximum. Power consumption in the standby mode is 200mW
maximum.
SUMMARY
More examples of DSP applications are summarized in Figure 9.30. There are many
other practical applications of DSP in today’s rapidly expanding industrial,
communication, medical, military, and consumer markets. A discussion of each
would constitute a book in itself. It has been the purpose of this section to highlight
just a few of the more common applications and give the reader a sense of how DSP
affects practically every aspect of modern life.
OTHER DSP APPLICATIONS
■ Hands-Free Car Telephone Kits
■ Digital Answering Machines
■ Voice Recognition Systems
■ Cable Modems
■ Computer Sound Cards
■ Digital Audio: Professional and Consumer
■ Digital Video Signal Processing
■ High Definition Television (HDTV)
■ Computer Generated Graphics
■ Digital Special Effects
■ Direct Broadcast Satellite (DBS)
■ Global Positioning Systems (GPS)
■ Medical: Ultrasound, MRI, CT Scanners
■ Military: Radar, Missile Guidance
Figure 9.30
9.42
DSP APPLICATIONS
REFERENCES
1.
John Bingham, The Theory and Practice of Modem Design,
John-Wiley, 1988.
2.
Digital Signal Processing Applications Using the ADSP-2100
Family, Vol. 1 and Vol. 2, Analog Devices, Free download at:
http://www.analog.com
3.
George Calhoun, Digital Cellular Radio, Artech House, Norwood MA,
1988.
4.
Nick Morley, New Cellular Scheme Muscles In, EDN News, November 15,
1990, p. 1.
5.
P. Vary, K. Hellwig, R. Hofmann, R. J. Sluyter, C. Galand, M. Rosso, Speech
Codec for the European Mobile Radio System, ICASSP 1988 Proceedings,
p. 227.
6.
D. K. Freeman, G. Cosier, C. B. Southcott, I. Boyd, The Voice Activity
Detector for the Pan-European Digital Cellular System, ICASSP 1989
Proceedings, p. 369.
7.
Brad Brannon, Using Wide Dynamic Range Converters for Wide
Band Radios, RF Design, May 1995, pp.50-65.
8.
Joe Mitola, The Software Radio Architecture, IEEE Communications
Magazine, Vol. 33, No.5, May 1995, pp. 26-38.
9.
Jeffery Wepman, Analog-to-Digital Converters and Their Applications
in Radio Receivers, IEEE Communications Magazine,
Vol. 33, No.5, May 1995, pp. 39-45.
10.
Rupert Baines, The DSP Bottleneck, IEEE Communications
Magazine, Vol. 33, No.5, May 1995, pp. 46-54.
11.
Brad Brannon, Overcoming Converter Nonlinearities with Dither,
Application Note AN-410, Analog Devices, 1995.
12.
P.J.M. Coussens, et al, Three Phase Measurements with Vector Rotation
Blocks in Mains and Motion Control, PCIM Conference, Intelligent
Motion, April 1992 Proceedings, Available from Analog Devices.
13.
Dennis Fu, Digital to Synchro and Resolver Conversion with the AC Vector
Processor AD2S100, Available from Analog Devices.
14.
Dennis Fu, Circuit Applications of the AD2S90 Resolver-to-Digital Converter,
AN-230, Analog Devices.
9.43
DSP APPLICATIONS
15.
Aengus Murray and P. Kettle, Towards a Single Chip DSP Based Motor
Control Solution, Proceedings PCIM - Intelligent Motion, May 1996,
Nurnberg Germany, pp. 315-326. Also available at http://www.analog.com.
16.
D. J. Lucey, P. J. Roche, M. B. Harrington, and J. R. Scannell,
Comparison of Various Space Vector Modulation Strategies,
Proceedings Irish DSP and Control Colloquium, July 1994, Dublin,
Ireland, pp. 169-175.
17.
Niall Lyne, ADCs Lend Flexibility to Vector Motor Control Applications,
Electronic Design, May 1, 1998, pp. 93-100.
18.
Motor Control Products, Application Notes, and Tools, CDROM,
Analog Devices, Spring 2000.
19.
Application of Digital Signal Processing in Motion Control Seminar,
Analog Devices, Spring 2000.
20.
Analog Devices’ Motor Control Website:
http://www.analog.com/motorcontrol
9.44
HARDWARE DESIGN TECHNIQUES
SECTION 10
HARDWARE DESIGN TECHNIQUES
■
Low Voltage Interfaces
■
Grounding in Mixed Signal Systems
■
Digital Isolation Techniques
■
Power Supply Noise Reduction and Filtering
■
Dealing with High Speed Logic
10.a
HARDWARE DESIGN TECHNIQUES
10.b
HARDWARE DESIGN TECHNIQUES
SECTION 10
HARDWARE DESIGN TECHNIQUES
Walt Kester
LOW VOLTAGE INTERFACES
Ethan Bordeaux, Johannes Horvath, Walt Kester
For the past 30 years, the standard VDD for digital circuits has been 5V. This voltage
level was used because bipolar transistor technology required 5V to allow headroom
for proper operation. However, in the late 1980s, Complimentary Metal Oxide
Semiconductor (CMOS) became the standard for digital IC design. This process did
not necessarily require the same voltage levels as TTL circuits, but the industry
adopted the 5V TTL standard logic threshold levels to maintain backward
compatibility with older systems (Reference 1).
The current revolution in supply voltage reduction has been driven by demand for
faster and smaller products at lower costs. This push has caused silicon geometries
to drop from 2µm in the early 1980s to 0.25µm that is used in today’s latest
microprocessor and IC designs. As feature sizes have become increasingly smaller,
the voltage for optimum device performance has also dropped below the 5V level.
This is illustrated in the current microprocessors for PCs, where the optimum core
operating voltage is programmed externally using voltage identification (VID) pins,
and can be as low as 1.3V.
The strong interest in lower voltage DSPs is clearly visible in the shifting sales
percentages for 5V and 3.3V parts. Sales growth for 3.3V DSPs has increased at
more than twice the rate of the rest of the DSP market (30% for all DSPs versus
more than 70% for 3.3V devices). This trend will continue as the high volume/high
growth portable markets demand signal processors that contain all of the traits of
the lower voltage DSPs.
On the one hand, the lower voltage ICs operate at lower power, allow smaller chip
areas, and higher speeds. On the other hand, the lower voltage ICs must often
interface to other ICs which operate at larger VDD supply voltages thereby causing
interface compatibility problems. Although lower operating voltages mean smaller
signal swings, and hence less switching noise, noise margins are lower for low
supply voltage ICs.
The popularity of 2.5V devices can be partially explained by their ability to operate
from two AA alkaline cells. Figure 10.2 shows the typical discharge characteristics
for a AA cell under various load conditions (Reference 2). Note that at a load current
of 15mA, the voltage remains above +1.25V (2.5V for two cells in series) for nearly
100 hours. Therefore, an IC that can operate effectively at low currents with a
supply voltage of 2.5V ±10% (2.25V - 2.75V) is very useful in portable designs. Also,
DSPs that have low mA/MIPS ratings and can integrate peripherals onto a single
chip, such as the ADSP-218x L or M-series, are useful in portable applications.
10.1
HARDWARE DESIGN TECHNIQUES
LOW VOLTAGE MIXED-SIGNAL ICs
■ Lower Power for Portable Applications
■ 2.5V ICs Can Operate on Two “AA” Alkaline Cells
■ Faster CMOS Processes, Smaller Geometries, Lower
Breakdown Voltages
■ Multiple Voltages in System: +5V, +3.3V, +2.5V, +1.8V
DSP Core Voltage (VID), Analog Supply Voltage
■ Interfaces Required Between Multiple Logic Types
■ Lower Voltage Swings Produce Less Switching Noise
■ Lower Noise Margins
■ Less Headroom in Analog Circuits Decreases Signal Swings
and Increases Sensitivity to Noise (But that’s the subject of an
entire seminar!)
Figure 10.1
DURACELL MN1500 “AA” ALKALINE BATTERY
DISCHARGE CHARACTERISTICS
VOLTAGE
(V)
1.25
SERVICE HOURS
Courtesy: Duracell, Inc., Berkshire Corporate Park, Bethel, CT 06801
http://www.duracell.com
Figure 10.2
10.2
HARDWARE DESIGN TECHNIQUES
In order to understand the compatibility issues relating to interfacing ICs operated
at different VDD supplies, it is useful to first look at the structure of a typical CMOS
logic stage as shown in Figure 10.3.
TYPICAL CMOS IC OUTPUT DRIVER CONFIGURATION
VDD
VDD
PMOS
High = “1”
VIH MIN
INPUT
VDD
VDD
High = “1”
VOH MIN
PREDRIVER
LOGIC
OUTPUT
NMOS
VIL MAX
Low = “0”
Low = “0”
0V
VOL MAX
0V
VIL MAX = Maximum Allowable Input Low Logic Level
VIH MIN = Minimum Allowable Input High Logic Level
VOL MAX = Maximum Allowable Output Low Logic Level
VOH MIN = Minimum Allowable Output High Logic Level
Figure 10.3
Note that the output driver stage consists of a PMOS and an NMOS transistor.
When the output is high, the PMOS transistor connects the output to the +VDD
supply through its low on-resistance (RON), and the NMOS transistor is off. When
the output is low, the NMOS transistor connects the output to ground through its
on-resistance, and the PMOS transistor is off. The RON of a CMOS output stage can
vary between 5Ω and 50Ω depending on the size of the transistors, which in turn,
determines the output current drive capability.
A typical logic IC has its power supplies and grounds separated between the output
drivers and the rest of the circuitry (including the pre-driver). This is done to
maintain a clean power supply, which reduces the effect of noise and ground bounce
on the I/O levels. This is increasingly important, since added tolerance and
compliance are critical in I/O driver specifications, especially at low voltages.
Figure 10.3 also shows “bars” which define the minimum and maximum required
input and output voltages to produce a valid high or low logic level. Note that for
CMOS logic, the actual output logic levels are determined by the drive current and
the RON of the transistors. For light loads, the output logic levels are very close to 0V
and +VDD. The input logic thresholds, on the other hand, are determined by the
input circuit of the IC.
There are three sections in the “input” bar. The bottom section shows the input
range that is interpreted as a logic low. In the case of 5V TTL, this range would be
10.3
HARDWARE DESIGN TECHNIQUES
between 0V and 0.8V. The middle section shows the input voltage range where it is
interpreted as neither a logic low nor a logic high. The upper section shows where
an input is interpreted as a logic high. In the case of 5V TTL, this would be between
2V and 5V.
Similarly, there are three sections in the “output” bar. The bottom range shows the
allowable voltage for a logic low output. In the case of 5V TTL, the IC must output a
voltage between 0V and 0.4V. The middle section shows the voltage range that is
not a valid high or low - the device should never transmit a voltage level in this
region except when transitioning from one level to the other. The upper section
shows the allowable voltage range for a logic high output signal. For 5V TTL, this
voltage is between 2.4V and 5V. The chart does not reflect a 10%
overshoot/undershoot also allowed on the inputs of the logic standard.
A summary of the existing logic standards using these definitions is shown in
Figure 10.4. Note that the input thresholds of classic CMOS logic (series-4000, for
example) are defined as 0.3VDD and 0.7VDD. However, most CMOS logic circuits in
use today are compatible with TTL and LVTTL levels which are the dominant 5V
and 3.3V operating standards for DSPs. Note that 5V TTL and 3.3V LVTTL input
and output threshold voltages are identical. The difference is the upper range for
the allowable high levels.
LOW VOLTAGE LOGIC LEVEL STANDARDS
VOLTS
5.0
4.5
4.0
5V CMOS
5V TTL
I = INPUT
I
O
I
O
O = OUTPUT
3.3V LVTTL
3.5
3.0
2.5V VCX
2.5V JEDEC
O
I
I
O
I
O
2.5V VCX output
(-12mA)
2.5V JEDEC input
2.5V JEDEC output
(-1mA)
2.0
2.5V VCX input
2.5
1.5
1.0
3.3V LVTTL output
(-2mA)
3.3V LVTTL input
5V TTL output
(-2mA)
5V TTL input
5V CMOS output
(20uA)
0
5V CMOS input
0.5
Figure 10.4
10.4
HARDWARE DESIGN TECHNIQUES
The international standards bureau JEDEC (Joint Electron Device Engineering
Council) has created a 2.5V standard (JEDEC standard 8-5) which will most likely
become the minimum requirement for 2.5V operation (Reference 3). However, there
is no current (2000) dominant 2.5V standard for IC transmission and reception,
because few manufacturers are making products that operate at this voltage. There
is one proposed 2.5V standard created by a consortium of IC manufacturers, titled
the Low Voltage Logic Alliance. Their specification provides a guideline for
semiconductor operation between 1.8V and 3.6V. A standard covering this voltage
range is useful because it ensures present and future compatibility. As an example,
the 74VCX164245, a bus translator/transceiver from Fairchild Semiconductor, is
designed to be operated anywhere between 1.8V - 3.6V and has different input and
output characteristics depending upon the supplied VDD. This standard, named
VCX, was formed by Motorola, Toshiba and Fairchild Semiconductor. It currently
consists primarily of bus transceivers, translators, FIFOs and other building block
logic. There are also a wide range of other low voltage standards, such as GTL
(Gunning Transceiver Logic), BTL (Backplane Transceiver Logic), and PECL
(PseudoECL Logic). However, most of these standards are aimed at application
specific markets and not for general purpose semiconductor systems.
The VCX devices can be operated on a very wide range of voltage levels (1.8V 3.6V). The I/O characteristics of this standard are dependent upon the VDD voltage
and the load on each pin. In Figure 10.4, one voltage (2.5V) was chosen to show the
general I/O behavior of a VCX device. Each of the device’s output voltages is listed
for a specific current. As the current requirements increase, the output high voltage
decreases while the output low voltage increases. Please refer to the appropriate
data sheets for more specific I/O information.
From this chart, it is possible to visualize some of the possible problems in
connecting together two ICs operating on different standards. One example would
be connecting a 5V CMOS device to a 3.3V LVTTL IC. The 5V CMOS high level is
too high for the LVTTL to handle ( > 3.3V). This could cause permanent damage to
the LVTTL chip. Another possible problem would be a system with a 2.5V JEDEC
IC driving a 5V CMOS device. The logic high level from the 2.5V device is not high
enough for it to register as a logic high on the 5V CMOS input (VIH MIN = 3.5V).
These examples illustrate two possible types of logic level incompatibilities - either
a device being driven with too high a voltage or a device not driving a voltage high
enough for it to register a valid high logic level with the receiving IC. These
interfacing problems introduce two important concepts: voltage tolerance and
voltage compliance.
VOLTAGE TOLERANCE AND VOLTAGE COMPLIANCE
A device that is voltage tolerant can withstand a voltage greater than its VDD on its
I/O pins. For example, if a device has a VDD of 2.5V and can accept inputs equal to
3.3V and can withstand 3.3V on its outputs, the 2.5V device is called 3.3V tolerant.
The meaning of input voltage tolerance is fairly obvious, but the meaning of output
voltage tolerance requires some explanation. The output of a 2.5V CMOS driver in
the high state appears like a small resistor (RON of the PMOS FET) connected to
2.5V. Obviously, connecting its output directly to 3.3V is likely to destroy the device
due to excessive current. However, if the 2.5V device has a three-state output which
is connected to a bus which is also driven by a 3.3V IC, then the meaning becomes
10.5
HARDWARE DESIGN TECHNIQUES
clearer. Even though the 2.5V IC is in the off (third-state) condition, the 3.3V IC can
drive the bus voltage higher than 2.5V, potentially causing damage to the 2.5V IC
output.
A device which is voltage compliant can receive signals from and transmit signals to
a device which is operated at a voltage greater than its own VDD. For example, if a
device has a 2.5V VDD and can transmit and receive signals to and from a 3.3V
device, the 2.5V device is said to be 3.3V compliant.
The interface between the 5V CMOS and 3.3V LVTTL parts illustrates a lack of
voltage tolerance; the LVTTL IC input is overdriven by the 5V CMOS device output.
The interface between the 2.5V JEDEC and the 5V CMOS part demonstrates a lack
of voltage compliance; the output high level of the JEDEC IC does not comply to the
input level requirement of a the 5V CMOS device.
LOGIC VOLTAGE TOLERANCE
AND COMPATIBILITY DEFINITIONS
■ Voltage Tolerance:
◆ A device that is Voltage Tolerant can withstand a voltage
greater than its VDD on its input and output pins. If a device
has a VDD of 2.5V and can accept inputs of 3.3V (±10%), the
2.5V device is 3.3V tolerant on its input. Input and output
tolerance should be examined and specified separately.
■ Voltage Compliance:
◆ A device that is Voltage Compliant can transmit and receive
signals to and from logic which is operated at a voltage
greater than its own VDD. If a device has a 2.5V VDD and can
properly transmit signals to and from 3.3V logic, the 2.5V
device is 3.3V compliant. Input and output compliance should
be examined and specified separately.
Figure 10.5
INTERFACING 5V TO 3.3V SYSTEMS USING NMOS FET BUS
SWITCHES
When combining ICs that operate on different voltage standards, one is often forced
to add additional discrete elements to ensure voltage tolerance and compliance. In
order to achieve voltage tolerance between 5V and 3.3V logic, for instance, a bus
switch voltage translator, or QuickSwitch™ can be used (Reference 4, 5). The bus
switch limits the voltage applied to an IC. This is done to avoid applying a larger
input high voltage than the receiving device can tolerate.
As an example, it is possible to place a bus switch between a 5V CMOS and 3.3V
LVTTL IC, and the two devices can then transmit data properly as shown in Figure
10.6. The bus switch is basically an NMOS FET. If 4.3V is placed on the gate of the
FET, the maximum passable signal is 3.3V (approximately 1V less than the gate
10.6
HARDWARE DESIGN TECHNIQUES
voltage). If both input and output are below 3.3V, the NMOS FET acts as a low
resistance (RON ≈ 5Ω). As the input approaches 3.3V, the FET on-resistance
increases, thereby limiting the signal output. The QuickSwitch contains 10 bidirectional FETs with a gate drive enable as shown in Figure 10.6. The VCC of the
QuickSwitch sets the high level for the gate drive.
One way of creating a 4.3V supply on a 5V/3.3V system board is to place a diode
between the 5V supply and VCC on the QuickSwitch. In Figure 10.6, the 4.3V is
generated by a silicon diode in series with a Schottky diode connected to the 3.3V
supply. With 10% tolerances on both 5V and 3.3V supplies, this method produces a
more stable gate bias voltage. Some bus switches are designed to operate on either
3.3V or 5V directly and generate the internal gate bias level internally.
A QuickSwitch removes voltage tolerance concerns in this mixed logic design. One
convenient feature of bus switches is that they are bi-directional; this allows the
designer to place a bus translator between two ICs and not have to create additional
routing logic for input and output signals.
+3.3V / +5V BIDIRECTIONAL INTERFACE USING
NMOS FET ACHIEVES VOLTAGE TOLERANCE
+ 3.3V BUS
≈ 1V
+ 4.3V
470 kΩ
Ω
+ 5V BUS
Si SCHOTTKY
0.1µF
3.3V ± 10%
LOGIC
NMOS
5V ± 10%
LOGIC
RON = 5 Ω
“1” = DISABLE
“0” = ENABLE
0V / VCC
QS3384 QuickSwitch ®
10 SWITCHES/PACKAGE
OPERATE WITH QS3384 VCC = +4.3V
Figure 10.6
A bus switch increases the total power dissipation along with the total area
required to layout a system. Since voltage bus switches are typically CMOS circuits,
they have very low power dissipation ratings. An average value for added
continuous power dissipation is 5mW per package (10 switches), and this is
independent of the frequency of signals which pass through the circuit. Bus
switches typically have 8 - 20 I/O pins per package and take up approximately 25 to
50mm2 of board space.
10.7
HARDWARE DESIGN TECHNIQUES
One concern when adding interface logic into a circuit is a possible increase in
propagation delay. Added propagation delay can create many timing problems in a
design. QuickSwitches have very low propagation delay values (< 0.25ns) as shown
in Figure 10.7.
QS3384 QuickSwitch® TRANSIENT RESPONSE
WITH +4.3V SUPPLY
VIN
VOUT
(VOLTS)
2.0V
0.8V
5V TTL,
3.3V LVTTL
INPUT THRESHOLDS
TIME (ns)
Courtesy: Integrated Device Technology (IDT), Inc., 2975 Stender Way, Santa Clara, CA 95054
http://www.idt.com
Figure 10.7
INTERNALLY CREATED VOLTAGE TOLERANCE / COMPLIANCE
The requirement for low power, high performance ICs has triggered a race among
manufacturers to design devices operating at and below 2.5V that are also
TTL/CMOS compatible. Figure 10.8 is a block diagram of a logic circuit that allows
the logic core to operate at a reduced voltage, while the output driver operates at a
standard supply voltage level of 3.3V, for example.
The technique followed by many IC manufacturers is to provide a secondary I/O
ring, i.e., the I/O drivers in a 2.5V IC are driven by a 3.3V power supply, hence the
device can be TTL compatible and meet the specification for VOH and VOL. The 3.3V
external power supply is required for the part to be 3.3V tolerant. This causes the
added complexity of two power supplies for the chip which have to be maintained in
all future plug-in generations of the IC.
10.8
HARDWARE DESIGN TECHNIQUES
A more flexible technique (used in the ADSP-218xM series DSPs) is to provide a
separate I/O ring with an external voltage with the option of setting that voltage
equal to the core’s operating voltage, if desired. This design can provide tolerance to
3.3V with the external voltage set to 2.5V, or 3.3V tolerance and compliance to 3.3V
with the external voltage set to 3.3V. There are vendors today that use this option
partially, i.e., the VCX devices are 3.3V tolerant at 2.5V internal and external
voltages, but do not have the option of 3.3V compliance. Other existing designs and
patents that address this issue do not support complete tolerance and compliance
and the low standby current specification. This approach is complicated, since the
circuits must meet the noise and power requirements with the external voltages at
3.3V or 2.5V.
CMOS IC WITH SECONDARY I/O RING
VDD INTERNAL(CORE)
VDD EXTERNAL
SECONDARY
I/O RING
PMOS
CORE
LOGIC
INPUT
OUTPUT
NMOS
INTERNAL GND
ADSP-21xx M-SERIES
EXTERNAL GND
VDD EXTERNAL = +3.3V OR +2.5V
= +2.5V
VDD CORE
INPUT +3.3V TOLERANT
Figure 10.8
There are several issues to consider in a dual-supply logic IC design:
•
Power-Up Sequencing: If two power supplies are required to give an IC
additional tolerance / compliance, what is the power-up sequence? Is it a
requirement that the power supplies are switched on simultaneously or can the
device only have a voltage supplied on the core or only on the I/O ring?
10.9
HARDWARE DESIGN TECHNIQUES
•
Process Support and Electro-Static Discharge (ESD) Protection: The transistors
created in the IC’s fabrication process must be able to both withstand and drive
high voltages. The high voltage transistors create additional fabrication costs
since they require more processing steps to build in high voltage tolerance.
Designs with standard transistors require additional circuitry. The I/O drivers
must also provide ESD protection for the device. Most current designs limit the
overvoltage to below one diode drop (0.7V) above the power supply. Protection
for larger overvoltage requires more diodes in series.
•
Internal High Voltage Generation: The PMOS transistors need to be placed in a
substrate well which is tied to the highest on-chip voltage to prevent lateral
diodes from turning on and drawing excessive current. This high voltage can
either be generated on chip using charge pumps or from an external supply. This
requirement can make the design complex, since one cannot efficiently use
charge pumps to generate higher voltages and also achieve low standby current.
•
Chip Area: Die size is a primary factor in reducing costs and increasing yields.
Tolerance and compliance circuitry may require either more or larger I/O devices
to achieve the desired performance levels.
•
Testing: Since the core and the I/O can be at different voltages, testing the
device for all possible combinations of voltages can be complicated, adding to the
total cost of the IC.
3.3V / 2.5V INTERFACES
The Fairchild 74VCX164245-series are low voltage 16-bit dual supply logic
translators/transceivers with three-state outputs. A simplified block diagram is
shown in Figure 10.9. These devices use the VCX low voltage standard previously
discussed. The output driver circuit is supplied from the VDDB power supply bus,
ensuring VDDB compliant and tolerant outputs. The input circuit is supplied from the
VDDA supply, and the input logic threshold adjust circuits optimize the input logic
thresholds for the particular value of VDDA. Figure 10.10 shows the VCX voltage
standards for 3.3V, 2.5V, and 1.8V supply voltages. Note that the input voltage is
3.3V tolerant for all three supply voltages.
These devices dissipate about 2mW per input/output and are packaged in a 48-lead
TSSOP with a 2.5V supply. Propagation delay is about 3.2ns.
Figure 10.11 shows two possibilities for a 3.3V to 2.5V logic interface. The top
diagram (A) shows a direct connection. This will work provided the 2.5V IC is 3.3V
tolerant on its input. If the 2.5V IC is not 3.3V tolerant, the VCX translator can be
used as shown in Figure 10.11B.
10.10
HARDWARE DESIGN TECHNIQUES
LOGIC TRANSLATING TRANSCEIVER
(FAIRCHILD 74VCX164245)
VDDA BUS
VDDB BUS
VDDA
INPUT
LOGIC
INPUT
LOGIC
THRESHOLD
ADJUST
CIRCUITS
VDDB
PREDRIVER
LOGIC
PMOS
OUTPUT
LOGIC
NMOS
74VCX164245 CHARACTERISTICS:
■ Power Dissipation = 2mW / Input or Output
■ 16-bits per 48-lead TSSOP 100mm2 package
■ 3.2ns Propagation Delay at +2.5V
Figure 10.9
VOLTAGE COMPLIANCE FOR VCX STANDARD
(FAIRCHILD 74VCX164245 TRANSLATOR)
3.3V
3.5
3.0
I
1.8V
2.5V
O
I = INPUT
VOLTS
O = OUTPUT
2.5
I
O
2.0
I
1.5
O
1.0
0.5
0
–12mA
–12mA
–6mA
Figure 10.10
10.11
HARDWARE DESIGN TECHNIQUES
+3.3V TO +2.5V INTERFACE
+3.3V BUS
A
+3.3V ± 10%
LOGIC
+2.5V BUS
DIRECT
CONNECTION
+3.3V BUS
+2.5V ± 10%
LOGIC
OK IF 2.5V LOGIC
IS 3.3V TOLERANT
(3.6V MAX.)
+2.5V BUS
TRANSLATOR
B
+3.3V ± 10%
LOGIC
+2.5V ± 10%
LOGIC
Figure 10.11
Figure 10.12A shows a direct connection between 2.5V and 3.3V logic. In order for
this to work, the 2.5V output must be at least 2V minimum. With no loading on the
2.5V output, the 3.3V IC input is connected directly to +2.5V through the onresistance of the PMOS transistor driver. This provides 0.5V noise margin for the
nominal supply voltage of 2.5V. However, the 10% tolerance on the 2.5V bus allows
it to drop to a minimum of 2.25V, and the noise margin is reduced to 0.25V. This
may still work in a relatively quiet environment, but could be marginal if there is
noise on the supply voltages.
Adding a 1.6kΩ pull-up resistor as shown in Figure 10.12B ensures the 2.5V output
will not drop below 2.5V due to the input current of the 3.3V device, but the
degraded noise margin still exists for a 2.25V supply. With a 50% duty cycle, the
resistor adds about 3.4mW power dissipation per output.
A more reliable interface between 2.5V and 3.3V logic is shown in Figure 10.12C,
where a VCX translator is used. This solves all noise margin problems associated
with (A) and (B) and requires about 2mW per output.
10.12
HARDWARE DESIGN TECHNIQUES
+2.5V TO +3.3V INTERFACE
+2.5V BUS
A
+2.5V ± 10%
LOGIC
(+2.25V MIN.)
+2.5V BUS
B
+2.5V ± 10%
LOGIC
(+2.25V MIN.)
+3.3V BUS
DIRECT
CONNECTION
+3.3V ± 10%
LOGIC
NEEDS
+2V MIN.
PULL UP
RESISTOR
1.6kΩ
Ω
3.4mW
50% D.C.
+3.3V BUS
+3.3V ± 10%
LOGIC
TRANSLATOR
+2.5V BUS
C
+2.5V ± 10%
LOGIC
(+2.25V MIN.)
+3.3V BUS
+3.3V ± 10%
LOGIC
2mW
Figure 10.12
10.13
HARDWARE DESIGN TECHNIQUES
REFERENCES ON LOW VOLTAGE INTERFACES
1.
P. Alfke, Low-Voltage FPGAs Allow 3.3V/5V System Design,
Electronic Design, p. 70-76, August 18, 1997.
2.
AA Alkaline Battery Discharge Characteristics, Duracell Inc.,
Berkshire Corporate Park, Bethel, CT 06801, http://www.duracell.com
3.
Joint Electron Device Engineering Council (JEDEC),
Standard 8-5, October 1995.
4.
QS3384 Data Sheet, Integrated Device Technology (IDT), Inc.,
2975 Stender Way, Santa Clara, CA 95054, http://www.idt.com
5.
Pericom Semiconductor Corporation, 2380 Bering Drive,
San Jose, CA 95131, http://www.pericom.com
6.
74VCX164245 Data Sheet, Fairchild Semiconductor, 1997.
http://www.fairchildsemi.com
7.
H. Johnson, M. Graham, High Speed Digital Design, Prentice Hall, 1993.
10.14
HARDWARE DESIGN TECHNIQUES
GROUNDING IN MIXED SIGNAL SYSTEMS
Walt Kester, James Bryant, Mike Byrne
Today's signal processing systems generally require mixed-signal devices such as
analog-to-digital converters (ADCs) and digital-to-analog converters (DACs) as well
as fast digital signal processors (DSPs). Requirements for processing analog signals
having wide dynamic ranges increases the importance of high performance ADCs
and DACs. Maintaining wide dynamic range with low noise in hostile digital
environments is dependent upon using good high-speed circuit design techniques
including proper signal routing, decoupling, and grounding.
In the past, "high precision, low-speed" circuits have generally been viewed
differently than so-called "high-speed" circuits. With respect to ADCs and DACs, the
sampling (or update) frequency has generally been used as the distinguishing speed
criteria. However, the following two examples show that in practice, most of today's
signal processing ICs are really "high-speed," and must therefore be treated as such
in order to maintain high performance. This is certainly true of DSPs, and also true
of ADCs and DACs.
All sampling ADCs (ADCs with an internal sample-and-hold circuit) suitable for
signal processing applications operate with relatively high speed clocks with fast
rise and fall times (generally a few nanoseconds) and must be treated as high speed
devices, even though throughput rates may appear low. For example, the 12-bit
AD7892 successive approximation (SAR) ADC operates on an 8MHz internal clock,
while the sampling rate is only 600kSPS.
Sigma-delta (Σ-∆) ADCs also require high speed clocks because of their high
oversampling ratios. The AD7722 16-bit ADC has an output data rate (effective
sampling rate) of 195kSPS, but actually samples the input signal at 12.5MSPS (64times oversampling). Even high resolution, so-called "low frequency" Σ-∆ industrial
measurement ADCs (having throughputs of 10Hz to 7.5kHz) operate on 5MHz or
higher clocks and offer resolution to 24-bits (for example, the Analog Devices
AD7730 and AD7731).
To further complicate the issue, mixed-signal ICs have both analog and digital
ports, and because of this, much confusion has resulted with respect to proper
grounding techniques. In addition, some mixed-signal ICs have relatively low digital
currents, while others have high digital currents. In many cases, these two types
must be treated differently with respect to optimum grounding.
Digital and analog design engineers tend to view mixed-signal devices from
different perspectives, and the purpose of this section is to develop a general
grounding philosophy that will work for most mixed signal devices, without having
to know the specific details of their internal circuits.
GROUND AND POWER PLANES
The importance of maintaining a low impedance large area ground plane is critical
to all analog circuits today. The ground plane not only acts as a low impedance
return path for decoupling high frequency currents (caused by fast digital logic) but
10.15
HARDWARE DESIGN TECHNIQUES
also minimizes EMI/RFI emissions. Because of the shielding action of the ground
plane, the circuit’s susceptibility to external EMI/RFI is also reduced.
Ground planes also allow the transmission of high speed digital or analog signals
using transmission line techniques (microstrip or stripline) where controlled
impedances are required.
The use of "buss wire" is totally unacceptable as a "ground" because of its
impedance at the equivalent frequency of most logic transitions. For instance, #22
gauge wire has about 20nH/inch inductance. A transient current having a slew rate
of 10mA/ns created by a logic signal would develop an unwanted voltage drop of
200mV at this frequency flowing through 1 inch of this wire:
∆v = L
10mA
∆i
= 20 nH ×
= 200mV.
ns
∆t
For a signal having a 2V peak-to-peak range, this translates into an error of about
200mV, or 10% (approximate 3.5-bit accuracy). Even in all-digital circuits, this error
would result in considerable degradation of logic noise margins.
Figure 10.13 shows an illustration of a situation where the digital return current
modulates the analog return current (top figure). The ground return wire
inductance and resistance is shared between the analog and digital circuits, and
this is what causes the interaction and resulting error. A possible solution is to
make the digital return current path flow directly to the GND REF as shown in the
bottom figure. This is the fundamental concept of a "star," or single-point ground
system. Implementing the true single-point ground in a system which contains
multiple high frequency return paths is difficult because the physical length of the
individual return current wires will introduce parasitic resistance and inductance
which can make obtaining a low impedance high frequency ground difficult. In
practice, the current returns must consist of large area ground planes for low
impedance to high frequency currents. Without a low-impedance ground plane, it is
therefore almost impossible to avoid these shared impedances, especially at high
frequencies.
All integrated circuit ground pins should be soldered directly to the low-impedance
ground plane to minimize series inductance and resistance. The use of traditional
IC sockets is not recommended with high-speed devices. The extra inductance and
capacitance of even "low profile" sockets may corrupt the device performance by
introducing unwanted shared paths. If sockets must be used with DIP packages, as
in prototyping, individual "pin sockets" or "cage jacks" may be acceptable. Both
capped and uncapped versions of these pin sockets are available (AMP part
numbers 5-330808-3, and 5-330808-6). They have spring-loaded gold contacts which
make good electrical and mechanical connection to the IC pins. Multiple insertions,
however, may degrade their performance.
Power supply pins should be decoupled directly to the ground plane using low
inductance ceramic surface mount capacitors. If through-hole mounted ceramic
capacitors must be used, their leads should be less than 1mm. The ceramic
capacitors should be located as close as possible to the IC power pins. Ferrite beads
may be also required for additional decoupling.
10.16
HARDWARE DESIGN TECHNIQUES
DIGITAL CURRENTS FLOWING IN ANALOG
RETURN PATH CREATE ERROR VOLTAGES
ID
IA
+
VD
INCORRECT
+
VA
ANALOG
CIRCUITS
VIN
GND
REF
IA + ID
DIGITAL
CIRCUITS
ID
ID
IA
+
VD
+
VA
GND
REF
VIN
CORRECT
ANALOG
CIRCUITS
DIGITAL
CIRCUITS
IA
ID
Figure 10.13
DOUBLE-SIDED VS. MULTILAYER PRINTED CIRCUIT BOARDS
Each PCB in the system should have at least one complete layer dedicated to the
ground plane. Ideally, a double-sided board should have one side completely
dedicated to ground and the other side for interconnections. In practice, this is not
possible, since some of the ground plane will certainly have to be removed to allow
for signal and power crossovers, vias, and through-holes. Nevertheless, as much
area as possible should be preserved, and at least 75% should remain. After
completing an initial layout, the ground layer should be checked carefully to make
sure there are no isolated ground "islands," because IC ground pins located in a
ground "island" have no current return path to the ground plane. Also, the ground
plane should be checked for "skinny" connections between adjacent large areas
which may significantly reduce the effectiveness of the ground plane. Needless to
say, auto-routing board layout techniques will generally lead to a layout disaster on
a mixed-signal board, so manual intervention is highly recommended.
Systems that are densely packed with surface mount ICs will have a large number
of interconnections; therefore multilayer boards are mandatory. This allows at least
one complete layer to be dedicated to ground. A simple 4-layer board would have
internal ground and power plane layers with the outer two layers used for
10.17
HARDWARE DESIGN TECHNIQUES
interconnections between the surface mount components. Placing the power and
ground planes adjacent to each other provides additional inter-plane capacitance
which helps high frequency decoupling of the power supply. In most systems, 4layers are not enough, and additional layers are required for routing signals as well
as power.
GROUND PLANES ARE MANDATORY!
■ Use Large Area Ground (and Power) Planes for Low Impedance
Current Return Paths (Must Use at Least a Double-Sided Board!)
■ Double-Sided Boards:
◆ Avoid High-Density Interconnection Crossovers and Vias
Which Reduce Ground Plane Area
◆ Keep > 75% Board Area on One Side for Ground Plane
■ Multilayer Boards: Mandatory for Dense Systems
◆ Dedicate at Least One Layer for the Ground Plane
◆ Dedicate at Least One Layer for the Power Plane
■ Use at Least 30% to 40% of PCB Connector Pins for Ground
■ Continue the Ground Plane on the Backplane Motherboard to
Power Supply Return
Figure 10.14
MULTICARD MIXED-SIGNAL SYSTEMS
The best way of minimizing ground impedance in a multicard system is to use a
"motherboard" PCB as a backplane for interconnections between cards, thus
providing a continuous ground plane to the backplane. The PCB connector should
have at least 30-40% of its pins devoted to ground, and these pins should be
connected to the ground plane on the backplane mother card. To complete the
overall system grounding scheme there are two possibilities:
1. The backplane ground plane can be connected to chassis ground at numerous
points, thereby diffusing the various ground current return paths. This is commonly
referred to as a "multipoint" grounding system and is shown in Figure 10.15.
2. The ground plane can be connected to a single system "star ground" point
(generally at the power supply).
10.18
HARDWARE DESIGN TECHNIQUES
MULTIPOINT GROUND CONCEPT
VA
PCB
VD
VA
GROUND PLANE
PCB
VD
GROUND PLANE
BACKPLANE
GROUND PLANE
CHASSIS
GROUND
POWER
SUPPLIES
VA
VD
Figure 10.15
The first approach is most often used in all-digital systems, but can be used in
mixed-signal systems provided the ground currents due to digital circuits are
sufficiently low and diffused over a large area. The low ground impedance is
maintained all the way through the PC boards, the backplane, and ultimately the
chassis. However, it is critical that good electrical contact be made where the
grounds are connected to the sheet metal chassis. This requires self-tapping sheet
metal screws or "biting" washers. Special care must be taken where anodized
aluminum is used for the chassis material, since its surface acts as an insulator.
The second approach ("star ground") is often used in high speed mixed-signal
systems having separate analog and digital ground systems and warrants further
discussion.
SEPARATING ANALOG AND DIGITAL GROUNDS
In mixed-signal systems with large amounts of digital circuitry, it is highly
desirable to physically separate sensitive analog components from noisy digital
components. It may also be beneficial to use separate ground planes for the analog
and the digital circuitry. These planes should not overlap in order to minimize
capacitive coupling between the two. The separate analog and digital ground planes
are continued on the backplane using either motherboard ground planes or "ground
screens" which are made up of a series of wired interconnections between the
10.19
HARDWARE DESIGN TECHNIQUES
connector ground pins. The arrangement shown in Figure 10.16 illustrates that the
two planes are kept separate all the way back to a common system "star" ground,
generally located at the power supplies. The connections between the ground planes,
the power supplies, and the "star" should be made up of multiple bus bars or wide
copper braids for minimum resistance and inductance. The back-to-back Schottky
diodes on each PCB are inserted to prevent accidental DC voltage from developing
between the two ground systems when cards are plugged and unplugged. This
voltage should be kept less than 300mV to prevent damage to ICs which have
connections to both the analog and digital ground planes. Schottky diodes are
preferable because of their low capacitance and low forward voltage drop. The low
capacitance prevents AC coupling between the analog and digital ground planes.
Schottky diodes begin to conduct at about 300mV, and several parallel diodes in
parallel may be required if high currents are expected. In some cases, ferrite beads
can be used instead of Schottky diodes, however they introduce DC ground loops
which can be troublesome in precision systems.
SEPARATING ANALOG AND DIGITAL GROUND PLANES
VA
ANALOG
GROUND
PLANE
PCB
VD
VA
DIGITAL
GROUND
PLANE
ANALOG
GROUND
PLANE
D
A
PCB
VD
DIGITAL
GROUND
PLANE
D
A
DIGITAL GROUND PLANE
BACKPLANE
ANALOG GROUND PLANE
POWER
SUPPLIES
SYSTEM
STAR
GROUND
VA
VD
Figure 10.16
It is mandatory that the impedance of the ground planes be kept as low as possible,
all the way back to the system star ground. DC or AC voltages of more than 300mV
between the two ground planes can not only damage ICs but cause false triggering
of logic gates and possible latchup.
10.20
HARDWARE DESIGN TECHNIQUES
GROUNDING AND DECOUPLING MIXED-SIGNAL ICS WITH LOW DIGITAL
CURRENTS
Sensitive analog components such as amplifiers and voltage references are always
referenced and decoupled to the analog ground plane. The ADCs and DACs (and
other mixed-signal ICs) with low digital currents should generally be treated as
analog components and also grounded and decoupled to the analog ground plane. At
first glance, this may seem somewhat contradictory, since a converter has an analog
and digital interface and usually has pins designated as analog ground (AGND) and
digital ground (DGND). The diagram shown in Figure 10.17 will help to explain this
seeming dilemma.
PROPER GROUNDING OF MIXED-SIGNAL ICs
WITH LOW INTERNAL DIGITAL CURRENTS
VA
VD
FERRITE BEAD
A
LP
LP
CSTRAY
RP
AIN/
OUT
ANALOG
CIRCUITS
LP
R
B
CSTRAY
IA
ID
AGND
A
SEE
TEXT
RP
DIGITAL
CIRCUITS DATA
A
RP
D
VD
VA
SHORT
CONNECTIONS
BUFFER
GATE OR
REGISTER
DATA
BUS
CIN ≈ 10pF
RP
LP
DGND
A
A = ANALOG GROUND PLANE
VNOISE
D
D = DIGITAL GROUND PLANE
Figure 10.17
10.21
HARDWARE DESIGN TECHNIQUES
Inside an IC that has both analog and digital circuits, such as an ADC or a DAC,
the grounds are usually kept separate to avoid coupling digital signals into the
analog circuits. Figure 10.17 shows a simple model of a converter. There is nothing
the IC designer can do about the wirebond inductance and resistance associated
with connecting the bond pads on the chip to the package pins except to realize it's
there. The rapidly changing digital currents produce a voltage at point B which will
inevitably couple into point A of the analog circuits through the stray capacitance,
CSTRAY. In addition, there is approximately 0.2pF unavoidable stray capacitance
between every pin of the IC package! It's the IC designer's job to make the chip
work in spite of this. However, in order to prevent further coupling, the AGND and
DGND pins should be joined together externally to the analog ground plane with
minimum lead lengths. Any extra impedance in the DGND connection will cause
more digital noise to be developed at point B; it will, in turn, couple more digital
noise into the analog circuit through the stray capacitance. Note that connecting
DGND to the digital ground plane applies VNOISE across the AGND and DGND
pins and invites disaster!
The name "DGND" on an IC tells us that this pin connects to the digital ground of
the IC. This does not imply that this pin must be connected to the digital ground of
the system.
It is true that this arrangement may inject a small amount of digital noise onto the
analog ground plane. These currents should be quite small, and can be minimized
by ensuring that the converter output does not drive a large fanout (they normally
can't, by design). Minimizing the fanout on the converter's digital port will also keep
the converter logic transitions relatively free from ringing and minimize digital
switching currents, and thereby reducing any potential coupling into the analog port
of the converter. The logic supply pin (VD) can be further isolated from the analog
supply by the insertion of a small lossy ferrite bead as shown in Figure 10.17. The
internal transient digital currents of the converter will flow in the small loop from
VD through the decoupling capacitor and to DGND (this path is shown with a heavy
line on the diagram). The transient digital currents will therefore not appear on the
external analog ground plane, but are confined to the loop. The VD pin decoupling
capacitor should be mounted as close to the converter as possible to minimize
parasitic inductance. These decoupling capacitors should be low inductance ceramic
types, typically between 0.01µF and 0.1µF.
TREAT THE ADC DIGITAL OUTPUTS WITH CARE
It is always a good idea (as shown in Figure 10.17) to place a buffer register
adjacent to the converter to isolate the converter's digital lines from noise on the
data bus. The register also serves to minimize loading on the digital outputs of the
converter and acts as a Faraday shield between the digital outputs and the data
bus. Even though many converters have three-state outputs/inputs, this isolation
register still represents good design practice. In some cases it may be desirable to
10.22
HARDWARE DESIGN TECHNIQUES
add an additional buffer register on the analog ground plane next to the converter
output to provide greater isolation.
The series resistors (labeled "R" in Figure 10.17) between the ADC output and the
buffer register input help to minimize the digital transient currents which may
affect converter performance. The resistors isolate the digital output drivers from
the capacitance of the buffer register inputs. In addition, the RC network formed by
the series resistor and the buffer register input capacitance acts as a lowpass filter
to slow down the fast edges.
A typical CMOS gate combined with PCB trace and a through-hole will create a load
of approximately 10pF. A logic output slew rate of 1V/ns will produce 10mA of
dynamic current if there is no isolation resistor:
1V
∆v
= 10pF ×
= 10mA .
ns
∆t
A 500Ω series resistors will minimize this output current and result in a rise and
fall time of approximately 11ns when driving the 10pF input capacitance of the
register:
∆I = C
t r = 2.2 × τ = 2.2 × R ⋅ C = 2.2 × 500Ω × 10 pF = 11ns.
TTL registers should be avoided, since they can appreciably add to the dynamic
switching currents because of their higher input capacitance.
The buffer register and other digital circuits should be grounded and decoupled to
the digital ground plane of the PC board. Notice that any noise between the analog
and digital ground plane reduces the noise margin at the converter digital interface.
Since digital noise immunity is of the orders of hundreds or thousands of millivolts,
this is unlikely to matter. The analog ground plane will generally not be very noisy,
but if the noise on the digital ground plane (relative to the analog ground plane)
exceeds a few hundred millivolts, then steps should be taken to reduce the digital
ground plane impedance, thereby maintaining the digital noise margins at an
acceptable level. Under no circumstances should the voltage between the two
ground planes exceed 300mV, or the ICs may be damaged.
Separate power supplies for analog and digital circuits are also highly desirable.
The analog supply should be used to power the converter. If the converter has a pin
designated as a digital supply pin (VD), it should either be powered from a separate
analog supply, or filtered as shown in the diagram. All converter power pins should
be decoupled to the analog ground plane, and all logic circuit power pins should be
decoupled to the digital ground plane as shown in Figure 10.18. If the digital power
supply is relatively quiet, it may be possible to use it to supply analog circuits as
well, but be very cautious.
In some cases it may not be possible to connect VD to the analog supply. Some of the
newer, high speed ICs may have their analog circuits powered by +5V, but the
digital interface powered by +3V to interface to 3V logic. In this case, the +3V pin of
the IC should be decoupled directly to the analog ground plane. It is also advisable
to connect a ferrite bead in series with the power trace that connects the pin to the
+3V digital logic supply.
10.23
HARDWARE DESIGN TECHNIQUES
GROUNDING AND DECOUPLING POINTS
VA FERRITE
VA
VD
BEAD
SEE
TEXT
A
A
VA
VD
D
R
ADC
OR
DAC
AMP
A
VA
AGND
A
R
A
A
SAMPLING
CLOCK
GENERATOR
A
TO OTHER
DIGITAL
CIRCUITS
DGND
A
VOLTAGE
REFERENCE
BUFFER
GATE
OR
REGISTER
D
VA
A
A
ANALOG
GROUND PLANE
D
DIGITAL
GROUND PLANE
Figure 10.18
The sampling clock generation circuitry should be treated like analog circuitry and
also be grounded and heavily-decoupled to the analog ground plane. Phase noise on
the sampling clock produces degradation in system SNR as will be discussed
shortly.
SAMPLING CLOCK CONSIDERATIONS
In a high performance sampled data system a low phase-noise crystal oscillator
should be used to generate the ADC (or DAC) sampling clock because sampling
clock jitter modulates the analog input/output signal and raises the noise and
distortion floor. The sampling clock generator should be isolated from noisy digital
circuits and grounded and decoupled to the analog ground plane, as is true for the
op amp and the ADC.
10.24
HARDWARE DESIGN TECHNIQUES
The effect of sampling clock jitter on ADC Signal-to-Noise Ratio (SNR) is given
approximately by the equation:
 1 
SNR = 20 log10 
,
 2π ft j 
where SNR is the SNR of a perfect ADC of infinite resolution where the only source
of noise is that caused by the RMS sampling clock jitter, tj. Note that f in the above
equation is the analog input frequency. Just working through a simple example, if tj
= 50ps RMS, f = 100kHz, then SNR = 90dB, equivalent to about 15-bits dynamic
range.
It should be noted that tj in the above example is the root-sum-square (RSS) value
of the external clock jitter and the internal ADC clock jitter (called aperture jitter).
However, in most high performance ADCs, the internal aperture jitter is negligible
compared to the jitter on the sampling clock.
Since degradation in SNR is primarily due to external clock jitter, steps must be
taken to ensure the sampling clock is as noise-free as possible and has the lowest
possible phase jitter. This requires that a crystal oscillator be used. There are
several manufacturers of small crystal oscillators with low jitter (less than 5ps
RMS) CMOS compatible outputs. (For example, MF Electronics, 10 Commerce Dr.,
New Rochelle, NY 10801, Tel. 914-576-6570.)
Ideally, the sampling clock crystal oscillator should be referenced to the analog
ground plane in a split-ground system. However, this is not always possible because
of system constraints. In many cases, the sampling clock must be derived from a
higher frequency multi-purpose system clock which is generated on the digital
ground plane. It must then pass from its origin on the digital ground plane to the
ADC on the analog ground plane. Ground noise between the two planes adds
directly to the clock signal and will produce excess jitter. The jitter can cause
degradation in the signal-to-noise ratio and also produce unwanted harmonics. This
can be remedied somewhat by transmitting the sampling clock signal as a
differential signal using either a small RF transformer as shown in Figure 10.19 or
a high speed differential driver and receiver IC. If an active differential driver and
receiver are used, they should be ECL to minimize phase jitter. In a single +5V
supply system, ECL logic can be connected between ground and +5V (PECL), and
the outputs AC coupled into the ADC sampling clock input. In either case, the
original master system clock must be generated from a low phase noise crystal
oscillator.
10.25
HARDWARE DESIGN TECHNIQUES
SAMPLING CLOCK DISTRIBUTION FROM
DIGITAL TO ANALOG GROUND PLANES
DIGITAL GROUND PLANE
VD
VD
LOW PHASE
NOISE
MASTER CLOCK
D
ANALOG GROUND PLANE
SAMPLING
CLOCK
SYSTEM CLOCK
GENERATORS
VD
METHOD 1
D
D
A
VD
DSP OR MICROPROCESSOR
VA
+
SAMPLING
CLOCK
_
METHOD 2
D
SNR = 20 log10
D
1
2π f tj
A
tj = Sampling Clock Jitter
f = Analog Input Frequency
Figure 10.19
THE ORIGINS OF THE CONFUSION ABOUT MIXED-SIGNAL GROUNDING:
APPLYING SINGLE-CARD GROUNDING CONCEPTS TO MULTICARD SYSTEMS
Most ADC, DAC, and other mixed-signal device data sheets discuss grounding
relative to a single PCB, usually the manufacturer's own evaluation board. This has
been a source of confusion when trying to apply these principles to multicard or
multi-ADC/DAC systems. The recommendation is usually to split the PCB ground
plane into an analog plane and a digital plane. It is then further recommended that
the AGND and DGND pins of a converter be tied together and that the analog
ground plane and digital ground planes be connected at that same point as shown in
Figure 10.20. This essentially creates the system "star" ground at the mixed-signal
device.
All noisy digital currents flow through the digital power supply to the digital ground
plane and back to the digital supply; they are isolated from the sensitive analog
portion of the board. The system star ground occurs where the analog and digital
ground planes are joined together at the mixed signal device. While this approach
will generally work in a simple system with a single PCB and single ADC/DAC, it is
not optimum for multicard mixed-signal systems. In systems having several ADCs
or DACs on different PCBs (or on the same PCB, for that matter), the analog and
digital ground planes become connected at several points, creating the possibility of
ground loops and making a single-point "star" ground system impossible. For these
reasons, this grounding approach is not recommended for multicard systems, and
10.26
HARDWARE DESIGN TECHNIQUES
the approach previously discussed should be used for mixed signal ICs with low
digital currents.
GROUNDING MIXED SIGNAL ICs : SINGLE PC BOARD
(TYPICAL EVALUATION/TEST BOARD)
VD
VA
VA
MIXED
SIGNAL
DEVICE
ANALOG
CIRCUITS
AGND
SYSTEM
STAR
GROUND
A
VD
DIGITAL
CIRCUITS
DGND
A
D
ANALOG
GROUND PLANE
D
DIGITAL
GROUND PLANE
D
A
ANALOG
SUPPLY
DIGITAL
SUPPLY
Figure 10.20
SUMMARY: GROUNDING MIXED SIGNAL DEVICES WITH LOW DIGITAL
CURRENTS IN A MULTICARD SYSTEM
Figure 10.21 summarizes the approach previously described for grounding a mixed
signal device which has low digital currents. The analog ground plane is not
corrupted because the small digital transient currents flow in the small loop
between VD, the decoupling capacitor, and DGND (shown as a heavy line). The
mixed signal device is for all intents and purposes treated as an analog component.
The noise VN between the ground planes reduces the noise margin at the digital
interface but is generally not harmful if kept less than 300mV by using a low
impedance digital ground plane all the way back to the system star ground.
However, mixed signal devices such as sigma-delta ADCs, codecs, and DSPs with
on-chip analog functions are becoming more and more digitally intensive. Along
with the additional digital circuitry come larger digital currents and noise. For
example, a sigma-delta ADC or DAC contains a complex digital filter which adds
considerably to the digital current in the device. The method previously discussed
depends on the decoupling capacitor between VD and DGND to keep the digital
transient currents and isolated in a small loop. However, if the digital currents are
10.27
HARDWARE DESIGN TECHNIQUES
significant enough and have components at DC or low frequencies, the decoupling
capacitor may have to be so large that it is impractical. Any digital current which
flows outside the loop between VD and DGND must flow through the analog ground
plane. This may degrade performance, especially in high resolution systems.
GROUNDING MIXED SIGNAL ICs WITH LOW
INTERNAL DIGITAL CURRENTS: MULTIPLE PC BOARDS
VN
VD
VA
FILTER
VA
VD
MIXED
SIGNAL
DEVICE
ANALOG
CIRCUITS
AGND
A
VN = NOISE BETWEEN
GROUND PLANES
R
BUS
BUFFER
LATCH
DGND
A
D
A
ANALOG
GROUND PLANE
A
DIGITAL
CIRCUITS
D
DIGITAL
GROUND PLANE
A
D
TO SYSTEM
ANALOG SUPPLY
D
TO SYSTEM
DIGITAL SUPPLY
TO SYSTEM STAR GROUND
Figure 10.21
It is difficult to predict what level of digital current flowing into the analog ground
plane will become unacceptable in a system. All we can do at this point is to suggest
an alternative grounding method which may yield better performance.
SUMMARY: GROUNDING MIXED SIGNAL DEVICES WITH HIGH DIGITAL
CURRENTS IN A MULTICARD SYSTEM
An alternative grounding method for a mixed signal device with high levels of
digital currents is shown in Figure 10.22. The AGND of the mixed signal device is
connected to the analog ground plane, and the DGND of the device is connected to
the digital ground plane. The digital currents are isolated from the analog ground
plane, but the noise between the two ground planes is applied directly between the
AGND and DGND pins of the device. For this method to be successful, the analog
and digital circuits within the mixed signal device must be well isolated. The noise
between AGND and DGND pins must not be large enough to reduce internal noise
margins or cause corruption of the internal analog circuits.
10.28
HARDWARE DESIGN TECHNIQUES
GROUNDING ALTERNATIVE FOR MIXED SIGNAL ICs WITH
HIGH DIGITAL CURRENTS: MULTIPLE PC BOARDS
VN = NOISE BETWEEN
GROUND PLANES
VN
VD
VA
✱
BACK-TO-BACK
SCHOTTKY
DIODES
OR
FERRITE
BEAD
(SEE TEXT)
VA
VD
MIXED
SIGNAL
DEVICE
ANALOG
CIRCUITS
AGND
A
DIGITAL
CIRCUITS
DGND
A
D
D
✱
ANALOG
GROUND PLANE
A
TO SYSTEM
ANALOG SUPPLY
A
DIGITAL
GROUND PLANE
D
TO SYSTEM STAR GROUND
D
TO SYSTEM
DIGITAL SUPPLY
Figure 10.22
Figure 10.22 shows optional Schottky diodes (back-to-back) or a ferrite bead
connecting the analog and digital ground planes. The Schottky diodes prevent large
DC voltages or low frequency voltage spikes from developing across the two planes.
These voltages can potentially damage the mixed signal IC if they exceed 300mV
because they appear directly between the AGND and DGND pins. As an alternative
to the back-to-back Schottky diodes, a ferrite bead provides a DC connection
between the two planes but isolates them at frequencies above a few MHz where the
ferrite bead becomes resistive. This protects the IC from DC voltages between
AGND and DGND, but the DC connection provided by the ferrite bead can
introduce unwanted DC ground loops and may not be suitable for high resolution
systems.
GROUNDING DSPS WITH INTERNAL PHASE-LOCKED LOOPS
As if dealing with mixed-signal ICs with AGND and DGNDs wasn’t enough, newer
DSPs such as the ADSP-21160 SHARC with internal phase-locked-loops (PLLs)
raise issues with respect to proper grounding. The ADSP-21160 PLL allows the
internal core clock (determines the instruction cycle time) to operate at a userselectable ratio of 2, 3, or 4 times the external clock frequency, CLKIN. The CLKIN
rate is the rate at which the synchronous external ports operates. Although this
allows using a lower frequency external clock, care must be taken with the power
and ground connections to the internal PLL as shown in Figure 10.23.
10.29
HARDWARE DESIGN TECHNIQUES
GROUNDING DSPs WITH INTERNAL
PHASE-LOCKED-LOOPS (PLLs)
+3.3V
10Ω
Ω
+2.5V
SHORT TRACES
40
AVDD
0.1µF
0.01µF
VDD INT
46
VDD EXT
PLL
X1, X2,
X3, X4
DSP
(ADSP-21160)
CLKIN
AGND
GND
83
SHORT TRACES
DIGITAL GROUND PLANE
Figure 10.23
In order to prevent internal coupling between digital currents and the PLL, the
power and ground connections to the PLL are brought out separately on pins labeled
AVDD and AGND, respectively. The AVDD +2.5V supply should be derived from the
VDD INT +2.5V supply using the filter network as shown. This ensures a relatively
noise-free supply for the internal PLL. The AGND pin of the PLL should be
connected to the digital ground plane of the PC board using a short trace. The
decoupling capacitors should be routed between the AVDD pin and AGND pin using
short traces.
GROUNDING SUMMARY
There is no single grounding method which will guarantee optimum performance
100% of the time! This section has presented a number of possible options
depending upon the characteristics of the particular mixed signal devices in
question. It is helpful, however to provide for as many options as possible when
laying out the initial PC board.
It is mandatory that at least one layer of the PC board be dedicated to ground plane!
The initial board layout should provide for non-overlapping analog and digital
ground planes, but pads and vias should be provided at several locations for the
installation of back-to-back Schottky diodes or ferrite beads, if required. Pads and
vias should also be provided so that the analog and digital ground planes can be
connected together with jumpers if required.
10.30
HARDWARE DESIGN TECHNIQUES
The AGND pins of mixed-signal devices should in general always be connected to
the analog ground plane. An exception to this are DSPs which have internal phaselocked-loops (PLLs), such as the ADSP-21160 SHARC. The ground pin for the PLL
is labeled AGND, but should be connected directly to the digital ground plane for
the DSP.
GROUNDING PHILOSOPHY SUMMARY
■ There is no single grounding method which is guaranteed to work
100% of the time!
■ Different methods may or may not give the same levels of
performance.
■ At least one layer on each PC board MUST be dedicated to ground
plane!
■ Do initial layout with split analog and digital ground planes.
■ Provide pads and vias on each PC board for back-to-back
Schottky diodes and optional ferrite beads to connect the two
planes.
■ Provide "jumpers" so that DGND pins of mixed-signal devices can
be connected to AGND pins (analog ground plane) or to digital
ground plane. (AGND of PLLs in DSPs should be connected to
digital ground plane).
■ Provide pads and vias for "jumpers" so that analog and digital
ground planes can be joined together at several points on each
PC board.
■ Follow recommendations on mixed signal device data sheet.
Figure 10.24
SOME GENERAL PC BOARD LAYOUT GUIDELINES FOR MIXED-SIGNAL
SYSTEMS
It is evident that noise can be minimized by paying attention to the system layout
and preventing different signals from interfering with each other. High level analog
signals should be separated from low level analog signals, and both should be kept
away from digital signals. We have seen elsewhere that in waveform sampling and
reconstruction systems the sampling clock (which is a digital signal) is as
vulnerable to noise as any analog signal, but is as liable to cause noise as any digital
signal, and so must be kept isolated from both analog and digital systems. If clock
driver packages are used in clock distribution, only one frequency clock should be
passed through a single package. Sharing drivers between clocks of different
frequencies in the same package will produce excess jitter and crosstalk and
degrade performance.
The ground plane can act as a shield where sensitive signals cross. Figure 10.25
shows a good layout for a data acquisition board where all sensitive areas are
isolated from each other and signal paths are kept as short as possible. While real
life is rarely as tidy as this, the principle remains a valid one.
10.31
HARDWARE DESIGN TECHNIQUES
ANALOG AND DIGITAL CIRCUITS
SHOULD BE PARTITIONED ON PCB LAYOUT
SAMPLING
CLOCK GENERATOR
REFERENCE
ANALOG
ADC
CONTROL
LOGIC
BUFFER
REGISTER
DEMULTIPLEXER
DIGITAL
FILTER
DSP
OR
µP
AMPLIFIER
POWER
TIMING
CIRCUITS
MULTIPLE
ANALOG GROUNDS
INPUT
DATA
BUS
ADDRESS
BUS
BUFFER
MEMORY
MULTIPLE
GROUNDS
Figure 10.25
There are a number of important points to be considered when making signal and
power connections. First of all a connector is one of the few places in the system
where all signal conductors must run in parallel - it is therefore imperative to
separate them with ground pins (creating a faraday shield) to reduce coupling
between them.
Multiple ground pins are important for another reason: they keep down the ground
impedance at the junction between the board and the backplane. The contact
resistance of a single pin of a PCB connector is quite low (of the order of 10mΩ)
when the board is new - as the board gets older the contact resistance is likely to
rise, and the board's performance may be compromised. It is therefore well
worthwhile to allocate extra PCB connector pins so that there are many ground
connections (perhaps 30-40% of all the pins on the PCB connector should be ground
pins). For similar reasons there should be several pins for each power connection,
although there is no need to have as many as there are ground pins.
Manufacturers of high performance mixed-signal ICs like Analog Devices offer
evaluation boards to assist customers in their initial evaluations and layout. ADC
evaluation boards generally contain an on-board low-jitter sampling clock oscillator,
output registers, and appropriate power and signal connectors. They also may have
additional support circuitry such as the ADC input buffer amplifier and external
reference.
10.32
HARDWARE DESIGN TECHNIQUES
The layout of the evaluation board is optimized in terms of grounding, decoupling,
and signal routing and can be used as a model when laying out the ADC PC board
in the system. The actual evaluation board layout is usually available from the ADC
manufacturer in the form of computer CAD files (Gerber files). In many cases, the
layout of the various layers appears on the data sheet for the device.
10.33
HARDWARE DESIGN TECHNIQUES
REFERENCES ON GROUNDING:
1.
William C. Rempfer, Get All the Fast ADC Bits You Pay For,
Electronic Design, Special Analog Issue, June 24, 1996, p.44.
2.
Mark Sauerwald, Keeping Analog Signals Pure in a Hostile Digital
World, Electronic Design, Special Analog Issue, June 24, 1996, p.57.
3.
Jerald Grame and Bonnie Baker, Design Equations Help Optimize
Supply Bypassing for Op Amps, Electronic Design, Special Analog
Issue, June 24, 1996, p.9.
4.
Jerald Grame and Bonnie Baker, Fast Op Amps Demand More Than
a Single-Capacitor Bypass, Electronic Design, Special Analog Issue,
November 18, 1996, p.9.
5.
Walt Kester and James Bryant, Grounding in High Speed Systems,
High Speed Design Techniques, Analog Devices, 1996, Chapter 7, p. 7-27.
6.
Jeffrey S. Pattavina, Bypassing PC Boards: Thumb Your Nose at Rules
of Thumb, EDN, Oct. 22, 1998, p.149.
7.
Henry Ott, Noise Reduction Techniques in Electronic Systems,
Second Edition, New York, John Wiley and Sons, 1988.
8.
Howard W. Johnson and Martin Graham, High-Speed Digital Design,
PTR Prentice Hall, 1993.
9.
Paul Brokaw, An I.C. Amplifier User's Guide to Decoupling, Grounding
and Making Things Go Right for a Change, Application Note,
Analog Devices, Inc., http://www.analog.com.
10.
Walt Kester, A Grounding Philosophy for Mixed-Signal Systems,
Electronic Design Analog Applications Issue, June 23, 1997, p. 29.
11.
Ralph Morrison, Grounding and Shielding Techniques, Fourth Edition,
John Wiley, 1998.
12.
Ralph Morrison, Solving Interference Problems in Electronics,
John Wiley, 1995.
13.
C. D. Motchenbacher and J. A. Connelly, Low Noise Electronic System
Design, John Wiley, 1993.
14.
Crystal Oscillators: MF Electronics, 10 Commerce Drive, New Rochelle,
NY, 10801, 914-576-6570.
15.
Mark Montrose, EMC and the Printed Circuit Board, IEEE Press,
1999 (IEEE Order Number PC5756).
10.34
HARDWARE DESIGN TECHNIQUES
DIGITAL ISOLATION TECHNIQUES
Walt Kester
One way to break ground loops is to use isolation techniques. Analog isolation
amplifiers find many applications where a high degree of isolation is required, such
as in medical instrumentation. Digital isolation techniques offer a reliable method
of transmitting digital signals over interfaces without introducing ground noise.
Optocouplers (also called optoisolators) are useful and available in a wide variety of
styles and packages. A typical optocoupler based on an LED and a phototransitor is
shown in Figure 10.26. A current of approximately 10mA is applied to an LED
transmitter, and the light output is received by a phototransistor. The light
produced by the LED is sufficient to saturate the phototransistor. Isolation of 5000V
RMS to 7000V RMS is common. Although excellent for digital signals, optocouplers
are too nonlinear for most analog applications. One should also realize that since
the phototransistor is operated in a saturated mode, rise and fall-times can range
from 10µs to 20µs in slower devices, thereby limiting applications at high speeds.
DIGITAL ISOLATION USING
LED / PHOTOTRANSISTOR OPTOCOUPLERS
VDD1 (5V) 425Ω
VIN
CMOS
GATE
GND1
10kΩ
IOUT
IIN
HIGH VOLTAGE
ISOLATION BARRIER
VDD2 (5V)
VOUT
GND2
■ Uses Light for Transmission Over a High Voltage Barrier
■ The LED is the Transmitter, and the Phototransistor is the Receiver
■ High Voltage Isolation: 5000V to 7000V RMS
■ Non-Linear -- Best for Digital or Frequency Information
■ Rise and Fall-times can be 10 to 20µs in Slower Devices
■ Example: Siemens ILQ-1 Quad (http://www.siemens.com)
Figure 10.26
A faster optocoupler architecture is shown in Figure 10.27 and is based on an LED
and a photodiode. The LED is again driven with a current of approximately 10mA.
This produces a light output sufficient generate enough current in the receiving
photodiode to develop a valid high logic level at the output of the transimpedance
amplifier. Speed can vary widely between optocouplers, and the fastest ones have
propagation delays of 20ns typical, and 40ns maximum, and can handle data rates
10.35
HARDWARE DESIGN TECHNIQUES
up to 25MBd for NRZ data. This corresponds to a maximum square wave operating
frequency of 12.5MHz, and a minimum allowable passable pulse width of 40ns.
DIGITAL ISOLATION USING
LED / PHOTODIODE OPTOCOUPLERS
VDD1
VDD2
VIN
VOUT
(Data In)
(Data Out)
GND1
GND2
■
■
■
■
■
■
■
+5V Supply Voltage
2500V RMS I/O Withstand Voltage
Logic Signal Frequency: 12.5MHz Maximum
25MBd Maximum Data Rate
40ns Maximum Propagation Delay
9ns Typical Rise/Fall Time
Example: Agilent HCPL-7720
(http://www.semiconductor.agilent.com)
Figure 10.27
The ADuM1100A and ADuM1100B are digital isolators based on Analog Devices’
µmIsolation™ (micromachined isolation) technology. Combining high speed CMOS
and monolithic air core transformer technology, these isolation components provide
outstanding performance characteristics superior to the traditional optocouplers
previously described.
Configured as pin-compatible replacements for existing high speed optocouplers, the
ADuM1100A and ADuM1100B support data rates as high as 25MBd and 100MBb,
respectively. A functional diagram of the devices is shown in Figure 10.28.
Both the ADuM1100A and ADuM1100B operate at either 3.3V or 5V supply
voltages, have propagation delays < 10ns, edge asymmetry of <2ns, and rise and
fall-times < 2ns. They operate at very low power, less than 600µA of quiescent
current (sum of both sides) and a dynamic current of less than 230µA per MBd of
data rate. Unlike common transformer implementations, the parts provide DC
correctness with a patented refresh feature which continuously updates the output
signal.
10.36
HARDWARE DESIGN TECHNIQUES
ADuM1100A/ADuM1100B DIGITAL ISOLATORS
VDD1
VDD2
VIN
ENCODE
DECODE
VOUT
(Data Out)
(Data In)
UPDATE
WATCH
GND1
GND2
■
■
■
■
■
■
■
+5V/+3.3V Supply Voltage
2500V RMS I/O Withstand Voltage
25MBd Maximum Data Rate (ADuM1100A)
100 MBd Maximum Data Rate (ADuM1100B)
10ns Maximum Propagation Delay
2ns Typical Rise/Fall Time
Pin Compatible with Popular Optocouplers
Figure 10.28
The AD260/AD261 family of digital isolators isolates five digital control signals
to/from high speed DSPs, microcontrollers, or microprocessors. The AD260 also has
a 1.5W transformer for a 3.5kV RMS isolated external DC/DC power supply circuit.
Each line of the AD260 can handle digital signals up to 20MHz (40MBd) with a
propagation delay of only 14ns which allows for extremely fast data transmission.
Output waveform symmetry is maintained to within ±1ns of the input so the AD260
can be used to accurately isolate time-based pulse width modulator (PWM) signals.
A simplified schematic of one channel of the AD260/AD261 is shown in Figure
10.29. The data input is passed through a schmitt trigger circuit, through a latch,
and a special transmitter circuit which differentiates the edges of the digital input
signal and drives the primary winding of a proprietary transformer with a "sethigh/set-low" signal. The secondary of the isolation transformer drives a receiver
with the same "set-hi/set-low" data which regenerates the original logic waveform.
An internal circuit operates in the background which interrogates all inputs about
every 5µs and in the absence of logic transitions, sends appropriate "set-hi/set-low"
data across the interface. Recovery time from a fault condition or at power-up is
thus between 5µs and 10µs.
The power transformer (available on the AD260) is designed to operate between
150kHz and 250kHz and will easily deliver more than 1W of isolated power when
driven push-pull (5V) on the transmitter side. Different transformer taps, rectifier
10.37
HARDWARE DESIGN TECHNIQUES
and regulator schemes will provide combinations of ±5V, 15V, 24V, or even 30V or
higher. The output voltage when driven with a low voltage-drop drive will be
37V p-p across the entire secondary with a 5V push-pull drive.
AD260/AD261 DIGITAL ISOLATORS
DATA
IN
SCHMITT
TRIGGER
LATCH
D
TRI STATE
XMTR
RCVR
DATA
OUT
E
ENABLE
ENABLE
CONTINUOUS
UPDATE
CIRCUIT
ISOLATED POWER
XFMR (AD260)
37V p-p, 1.5W
NOTE: SINGLE DATA CHANNEL SHOWN
3500V RMS ISOLATION BARRIER
(AD260B/AD261B)
Figure 10.29
AD260/AD261 DIGITAL ISOLATOR KEY SPECIFICATIONS
■ Isolation Test Voltage to 3500V RMS (AD260B/AD261B)
■ Five Isolated Digital Lines Available in 6 Input/Output Configurations
■ Logic Signal Frequency: 20MHz Max.
■ Data Rate: 40MBd Max.
■ Isolated Power Transformer: 37V p-p, 1.5W (AD260)
■ Waveform Edge Transmission Symmetry: ±1ns
■ Propagation Delay: 14ns
■ Rise and Fall-Times < 5ns
Figure 10.30
10.38
HARDWARE DESIGN TECHNIQUES
POWER SUPPLY NOISE REDUCTION AND
FILTERING
Walt Jung, Walt Kester, Bill Chestnut
Precision analog circuitry has traditionally been powered from well regulated, low
noise linear power supplies. During the last decade however, switching power
supplies have become much more common in electronic systems. As a consequence,
they also are being used for analog supplies. Good reasons for the general popularity
include their high efficiency, low temperature rise, small size, and light weight.
In spite of these benefits, switchers do have drawbacks, most notably high output
noise. This noise generally extends over a broad band of frequencies, resulting in
both conducted and radiated noise, as well as unwanted electric and magnetic
fields. Voltage output noise of switching supplies are short-duration voltage
transients, or spikes. Although the fundamental switching frequency can range
from 20kHz to 1MHz, the spikes can contain frequency components extending to
100MHz or more. While specifying switching supplies in terms of RMS noise is
common vendor practice, as a user you should also specify the peak (or p-p)
amplitudes of the switching spikes, at the output loading of your system.
The following section discusses filter techniques for rendering a switching regulator
output analog ready, that is sufficiently quiet to power precision analog circuitry
with relatively small loss of DC terminal voltage. The filter solutions presented are
generally applicable to all power supply types incorporating switching element(s) in
their energy path. This includes various DC-DC converters as well as popular 5V
(PC type) supplies.
An understanding of the EMI process is necessary to understand the effects of
supply noise on analog circuits and systems. Every interference problem has a
source, a path, and a receptor [Reference 1]. In general, there are three methods for
dealing with interference. First, source emissions can be minimized by proper
layout, pulse-edge rise time control/reduction, filtering, and proper grounding.
Second, radiation and conduction paths should be reduced through shielding and
physical separation. Third, receptor immunity to interference can be improved, via
supply and signal line filtering, impedance level control, impedance balancing, and
utilizing differential techniques to reject undesired common-mode signals. This
section focuses on reducing switching power supply noise with external post filters.
Tools useful for combating high frequency switcher noise are shown by Figure 10.31.
They differ in electrical characteristics as well as their practicality towards noise
reduction, and are listed roughly in an order of priorities. Of these tools, L and C
are the most powerful filter elements, and are the most cost-effective, as well as
small in size.
10.39
HARDWARE DESIGN TECHNIQUES
SWITCHING REGULATOR NOISE REDUCTION TOOLS
■ Capacitors
■ Inductors
■ Ferrites
■ Resistors
■ Linear Post Regulation
■ Proper Layout and Grounding Techniques
■ PHYSICAL SEPARATION FROM SENSITIVE
ANALOG CIRCUITS!!
Figure 10.31
Capacitors are probably the single most important filter component for switchers.
There are many different types of capacitors, and an understanding of their
individual characteristics is absolutely mandatory in the design of effective practical
supply filters. There are generally three classes of capacitors useful in 10kHz100MHz filters, broadly distinguished as the generic dielectric types; electrolytic,
organic, film, and ceramic. These can in turn can be further sub-divided. A
thumbnail sketch of capacitor characteristics is shown in the chart of Figure 10.32.
TYPES OF CAPACITORS
Aluminum
Electrolytic
(General
Purpose)
Aluminum
Electrolytic
(Switching
Type)
Tantalum
Electrolytic
OS-CON
Electrolytic
Polyester
(Stacked
Film)
Ceramic
(Multilayer)
Size
100 µF
120 µF
120 µF
100 µF
1 µF
0.1 µF
Rated
Voltage
25 V
25 V
20 V
20 V
400 V
50 V
0.6 Ω @
100 kHz
0.18 Ω @
100 kHz
0.12 Ω @
100 kHz
0.02 Ω @
100 kHz
0.11 Ω @
1 MHz
0.12 Ω @
1 MHz
≅ 100 kHz
≅ 500 kHz
≅ 1 MHz
≅ 1 MHz
≅ 10 MHz
≅ 1 GHz
ESR
Operating
Frequency
(*)
(*) Upper frequency strongly size and package dependent
Figure 10.32
10.40
HARDWARE DESIGN TECHNIQUES
With any dielectric, a major potential filter loss element is ESR (equivalent series
resistance), the net parasitic resistance of the capacitor. ESR provides an ultimate
limit to filter performance, and requires more than casual consideration, because it
can vary both with frequency and temperature in some types. Another capacitor loss
element is ESL (equivalent series inductance). ESL determines the frequency where
the net impedance characteristic switches from capacitive to inductive. This varies
from as low as 10kHz in some electrolytics to as high as 100MHz or more in chip
ceramic types. Both ESR and ESL are minimized when a leadless package is used.
All capacitor types mentioned are available in surface mount packages, preferable
for high speed uses.
The electrolytic family provides an excellent, cost-effective low-frequency filter
component, because of the wide range of values, a high capacitance-to-volume ratio,
and a broad range of working voltages. It includes general purpose aluminum
electrolytic types, available in working voltages from below 10V up to about 500V,
and in size from 1 to several thousand µF (with proportional case sizes). All
electrolytic capacitors are polarized, and thus cannot withstand more than a volt or
so of reverse bias without damage. They also have relatively high leakage currents
(up to tens of µA, and strongly dependent upon design specifics).
A subset of the general electrolytic family includes tantalum types, generally
limited to voltages of 100V or less, with capacitance of up to 500µF [Reference 3]. In
a given size, tantalums exhibit a higher capacitance-to-volume ratios than do
general purpose electrolytics, and have both a higher frequency range and lower
ESR. They are generally more expensive than standard electrolytics, and must be
carefully applied with respect to surge and ripple currents.
A subset of aluminum electrolytic capacitors is the switching type, designed for
handling high pulse currents at frequencies up to several hundred kHz with low
losses [Reference 4]. This capacitor type competes directly with tantalums in high
frequency filtering applications, with the advantage of a broader range of values.
A more specialized high performance aluminum electrolytic capacitor type uses an
organic semiconductor electrolyte [Reference 5]. The OS-CON capacitors feature
appreciably lower ESR and higher frequency range than do other electrolytic types,
with an additional feature of low low-temperature ESR degradation.
Film capacitors are available in a very broad range of values and an array of
dielectrics, including polyester, polycarbonate, polypropylene, and polystyrene.
Because of the low dielectric constant of these films, their volumetric efficiency is
quite low, and a 10µF/50V polyester capacitor (for example) is actually the size of
your hand. Metalized (as opposed to foil) electrodes do help to reduce size, but even
the highest dielectric constant units among film types (polyester, polycarbonate) are
still larger than any electrolytic, even using the thinnest films with the lowest
10.41
HARDWARE DESIGN TECHNIQUES
voltage ratings (50V). Where film types excel is in their low dielectric losses, a factor
which may not necessarily be a practical advantage for filtering switchers. For
example, ESR in film capacitors can be as low as 10mΩ or less, and the behavior of
films generally is very high in terms of Q. In fact, this can cause problems of
spurious resonance in filters, requiring damping components.
Film capacitors using a wound layer-type construction can be inductive. This can
limit their effectiveness for high frequency filtering. Obviously, only non-inductively
made film caps are useful for switching regulator filters. One specific style which is
non-inductive is the stacked-film type, where the capacitor plates are cut as small
overlapping linear sheet sections from a much larger wound drum of dielectric/plate
material. This technique offers the low inductance attractiveness of a plate sheet
style capacitor with conventional leads [see References 4, 5, 6]. Obviously, minimal
lead length should be used for best high frequency effectiveness. Very high current
polycarbonate film types are also available, specifically designed for switching
power supplies, with a variety of low inductance terminations to minimize ESL
[Reference 7].
Dependent upon their electrical and physical size, film capacitors can be useful at
frequencies to well above 10MHz. At the highest frequencies, only stacked film
types should be considered. Some manufacturers are now supplying film types in
leadless surface mount packages, which eliminates the lead length inductance.
Ceramic is often the capacitor material of choice above a few MHz, due to its
compact size, low loss, and availability up to several µF in the high-K dielectric
formulations (X7R and Z5U), at voltage ratings up to 200V [see ceramic families of
Reference 3]. NP0 (also called COG) types use a lower dielectric constant
formulation, and have nominally zero TC, plus a low voltage coefficient (unlike the
less stable high-K types). NP0 types are limited to values of 0.1µF or less, with
0.01µF representing a more practical upper limit.
Multilayer ceramic “chip caps” are very popular for bypassing/ filtering at 10MHz or
higher, simply because their very low inductance design allows near optimum RF
bypassing. For smaller values, ceramic chip caps have an operating frequency range
to 1GHz. For high frequency applications, a useful selection can be ensured by
selecting a value which has a self-resonant frequency above the highest frequency of
interest.
All capacitors have some finite ESR. In some cases, the ESR may actually be helpful
in reducing resonance peaks in filters, by supplying “free” damping. For example, in
most electrolytic types, a nominally flat broad series resonance region can be noted
by the impedance vs. frequency plot. This occurs where |Z| falls to a minimum
level, nominally equal to the capacitor’s ESR at that frequency. This low Q
resonance can generally cover a relatively wide frequency range of several octaves.
10.42
HARDWARE DESIGN TECHNIQUES
Contrasted to the very high Q sharp resonances of film and ceramic caps, the low Q
behavior of electrolytics can be useful in controlling resonant peaks.
In most electrolytic capacitors, ESR degrades noticeably at low temperature, by as
much as a factor of 4-6 times at –55°C vs. the room temperature value. For circuits
where ESR is critical to performance, this can lead to problems. Some specific
electrolytic types do address this problem, for example within the HFQ switching
types, the –10°C ESR at 100kHz is no more than 2× that at room temperature. The
OSCON electrolytics have a ESR vs. temperature characteristic which is relatively
flat.
As noted, all real capacitors have parasitic elements which limit their performance.
The equivalent electrical network representing a real capacitor models both ESR
and ESL as well as the basic capacitance, plus some shunt resistance (see Figure
10.33). In such a practical capacitor, at low frequencies the net impedance is almost
purely capacitive. At intermediate frequencies, the net impedance is determined by
ESR, for example about 0.12Ω to 0.4Ω at 125kHz, for several types. Above about
1MHz these capacitor types become inductive, with impedance dominated by the
effect of ESL. All electrolytics will display impedance curves similar in general
shape to that of Figure 10.34. The minimum impedance will vary with the ESR, and
the inductive region will vary with ESL (which in turn is strongly affected by
package style).
CAPACITOR EQUIVALENT CIRCUIT
AND PULSE RESPONSE
i
IPEAK = 1A
INPUT
CURRENT
v
Ω
ESR = 0.2Ω
di
1A
=
dt
100ns
Equivalent f = 3.5MHz
0
ESL = 20nH
VPEAK = ESL •
C = 100µF
OUTPUT
VOLTAGE
Ω
XC = 0.0005Ω
@ 3.5MHz
di
+ ESR • IPEAK
dt
= 400mV
ESR • IPEAK = 200mV
0
Figure 10.33
10.43
HARDWARE DESIGN TECHNIQUES
ELECTROLYTIC CAPACITOR
IMPEDANCE VERSUS FREQUENCY
C (100µF)
REGION
LOG
|Z|
ESL (20nH)
REGION
ESR (0.2Ω)
Ω)
REGION
Ω
ESR = 0.2Ω
10kHz
1MHz
LOG FREQUENCY
Figure 10.34
Regarding inductors, Ferrites (non-conductive ceramics manufactured from the
oxides of nickel, zinc, manganese, or other compounds) are extremely useful in
power supply filters [Reference 9]. At low frequencies (<100kHz), ferrites are
inductive; thus they are useful in low-pass LC filters. Above 100kHz, ferrites
become resistive, an important characteristic in high-frequency filter designs.
Ferrite impedance is a function of material, operating frequency range, DC bias
current, number of turns, size, shape, and temperature. Figure 10.35 summarizes a
number of ferrite characteristics, and Figure 10.36 shows the impedance
characteristic of several ferrite beads from Fair-Rite (http://www.fair-rite.com).
Several ferrite manufacturers offer a wide selection of ferrite materials from which
to choose, as well as a variety of packaging styles (see References 10 and 11). A
simple form is the bead of ferrite material, a cylinder of the ferrite which is simply
slipped over the power supply lead to the decoupled stage. Alternately, the leaded
ferrite bead is the same bead, pre-mounted on a length of wire and used as a
component (see Reference 11). More complex beads offer multiple holes through the
cylinder for increased decoupling, plus other variations. Surface mount beads are
also available.
10.44
HARDWARE DESIGN TECHNIQUES
FERRITES SUITABLE FOR HIGH FREQUENCY FILTERS
■ Ferrites Good for Frequencies Above 25kHz
■ Many Sizes and Shapes Available Including Leaded "Resistor
Style"
■ Ferrite Impedance at High Frequencies Primarily Resistive -Ideal for HF Filtering
■ Low DC Loss: Resistance of Wire Passing Through Ferrite is
Very Low
■ High Saturation Current Versions Available
■ Choice Depends Upon:
◆ Source and Frequency of Interference
◆ Impedance Required at Interference Frequency
◆ Environmental: Temperature, AC and DC Field Strength,
Size / Space Available
■
Always Test the Design!
Figure 10.35
IMPEDANCE OF FERRITE BEADS
80
#43
MATERIAL
#64
MATERIAL
60
Z
Ω
#73
MATERIAL
40
20
0
1
10
100
1000
FREQUENCY (MHz)
Courtesy: Fair-Rite Products Corp, Wallkill, NY
(http://www.fair-rite.com)
Figure 10.36
10.45
HARDWARE DESIGN TECHNIQUES
PSpice ferrite models for Fair-Rite materials are available, and allow ferrite
impedance to be estimated [see Reference 12]. These models have been designed to
match measured impedances rather than theoretical impedances.
A ferrite’s impedance is dependent upon a number of inter-dependent variables, and
is difficult to quantify analytically, thus selecting the proper ferrite is not
straightforward. However, knowing the following system characteristics will make
selection easier. First, determine the frequency range of the noise to be filtered.
Second, the expected temperature range of the filter should be known, as ferrite
impedance varies with temperature. Third, the peak DC current flowing through
the ferrite must be known, to ensure that the ferrite does not saturate. Although
models and other analytical tools may prove useful, the general guidelines given
above, coupled with some experimentation with the actual filter connected to the
supply output under system load conditions, should lead to a proper ferrite
selection.
Using proper component selection, low and high frequency band filters can be
designed to smooth a noisy switcher’s DC output to produce an analog ready 5V
supply. It is most practical to do this over two (and sometimes more) stages, each
stage optimized for a range of frequencies. A basic stage can be used to carry all of
the DC load current, and filter noise by 60dB or more up to a 1-10MHz range. This
larger filter is used as a card entry filter providing broadband filtering for all power
entering a PC card. Smaller, more simple local filter stages are also used to provide
higher frequency decoupling right at the power pins of individual stages.
SWITCHING REGULATOR EXPERIMENTS
In order to better understand the challenge of filtering switching regulators, a series
of experiments were conducted with a representative device, the ADP1148
synchronous buck regulator with a 9V input and a 3.3V/1A output.
In addition to observing typical input and output waveforms, the objective of these
experiments was to reduce the output ripple to less than 10mV peak-to-peak, a
value suitable for driving most analog circuits.
Measurements were made using a Tektronix wideband digitizing oscilloscope with
the input bandwidth limited to 20MHz so that the ripple generated by the switching
regulators could be more readily observed. In a system, power supply ripple
frequencies above 20MHz are best filtered locally at each IC power pin with a low
inductance ceramic capacitor and perhaps a series-connected ferrite bead.
Probing techniques are critical for accurate ripple measurements. A standard
passive 10X probe was used with a "bayonet" probe tip adapter for making the
ground connection as short as possible (see Figure 10.37). Use of the "ground clip
lead" is not recommended in making this type of measurement because the lead
length in the ground connection forms an unwanted inductive loop which picks up
high frequency switching noise, thereby corrupting the signal being measured.
10.46
HARDWARE DESIGN TECHNIQUES
PROPER PROBING TECHNIQUES
PROBE
"GROUND CLIP"
CONNECTOR
SLIP-ON
"BAYONET"
GROUND
ADAPTER
"GROUND CLIP"
LEAD
(DO NOT USE!!)
SIGNAL
CONTACT
GROUND PLANE
CONTACT
IC
Figure 10.37
Note: Schematic representation of proper physical grounding is almost impossible. In
all the following circuit schematics, the connections to ground are made to the
ground plane using the shortest possible connecting path, regardless of how they are
indicated in the actual circuit schematic diagram.
The circuit for the ADP1148 9V to 3.3V/1A buck regulator is shown in Figure 10.38.
The output waveform of the ADP1148 buck regulator is shown in Figure 10.39. The
fundamental switching frequency is approximately 150kHz, and the output ripple is
approximately 40mV.
Adding an output filter consisting of a 50µH inductor and a 100µF leaded tantalum
capacitor reduced the ripple to approximately 3mV as shown in Figure 10.40.
Linear regulators are often used following switching regulators for better regulation
and lower noise. Low dropout (LDO) regulators such as the ADP3310 are desirable
in these applications because they require only a small input-to-output series
voltage to maintain regulation. This minimizes power dissipation in the pass device
and may eliminate the need for a heat sink. Figure 10.41 shows the ADP1148 buck
regulator configured for a 9V input and a 3.75V/1A output. The output drives an
ADP3310 linear LDO regulator configured for 3.75V input and 3.3V/1A output. The
input and output of the ADP3310 is shown in Figure 10.42. Notice that the
regulator reduces the ripple from 40mV to approximately 5mV.
10.47
HARDWARE DESIGN TECHNIQUES
ADP1148 BUCK REGULATOR CIRCUIT
+
1µF
VIN, 9V
VIN
10nF
220µF/25V
IRF7204
C1
INT VCC
P-DRIVE
P-CH
ADP1148-3.3
L, 50µH
RSENSE
0.1Ω
Ω
SHUTDOWN
ITH
SENSE (+)
CT
SENSE (–)
RC, 1kΩ
Ω
CC
CT
3300pF
470pF
C2
+
1000pF
N-CH
N-DRIVE
SGND
VOUT
3.3V/1A
10BQ040
100µF
20V
PGND
IRF7403
C1 = 220µF/25V GEN PURPOSE AL ELECTROLYTIC
+ 1µF CERAMIC
L=COILTRONICS CTX-50-4
C2 = 100µF/20V LEADED TANTALUM, KEMET T356-SERIES, ESR = 0.6Ω
Ω
Figure 10.38
ADP1148 BUCK OUTPUT WAVEFORM
VOUT
3.3V
1A
VIN
9V
+
1µF 220µF
25V
C1 = 220 µF
+1 µF
ADP1148
BUCK
REG
CIRCUIT
40mV p-p
+
100µF
20V
C2 = 100µF/20V
VERTICAL SCALE:
HORIZ. SCALE:
10mV / DIV
5µs / DIV
C1 = 1µF CERAMIC + 220µF/25V GENERAL PURPOSE AL ELECTROLYTIC
C2 = 100µF/20V LEADED TANTALUM, KEMET T356-SERIES (ESR = 0.6Ω
Ω)
Figure 10.39
10.48
HARDWARE DESIGN TECHNIQUES
ADP1148 BUCK FILTERED OUTPUT
LF
50µH
VIN
9V
ADP1148
BUCK
REG
CIRCUIT
+
1µF 220µF
25V
C1 = 220 µF
+1 µF
VOUT
3.3V
1A
3mV p-p
+
+
100µF
20V
CF
100µF
20V
C2 = 100µF/20V
VERTICAL SCALE:
HORIZ. SCALE:
10mV / DIV
5µs / DIV
C1 = 1µF CERAMIC + 220µF/25V GENERAL PURPOSE AL ELECTROLYTIC
C2 = 100µF/20V LEADED TANTALUM, KEMET T356-SERIES (ESR = 0.6Ω
Ω)
OUTPUT FILTER
LF=COILTRONICS CTX-50-4
CF = 100µF/20V LEADED TANTALUM, KEMET T356-SERIES
Figure 10.40
ADP1148 BUCK REGULATOR DRIVING
ADP3310 LOW DROPOUT REGULATOR
VIN, 9V
WAVEFORMS
1µF
VIN
10nF
IRF7204
P-DRIVE
IFR7404
C1
INT VCC
P-CH
ADP1148
+
220µF
35V
L, 68µH
RSENSE
0.1Ω
Ω
3.75V
SD
ITH
IN
SENSE (+)
RC, 1kΩ
Ω
CC
2200pF
CT
470pF
C2
SENSE (–)
N-CH
N-DRIVE
SGND
OUT
ADP3310-3.3
1000pF
CT
GATE
10BQ040
FB PGND
+
100µF
20V
GND
3.3V
1A
C3
10µF
35V
R1
20kΩ
Ω
R2
10kΩ
Ω
IRF7403
Figure 10.41
10.49
HARDWARE DESIGN TECHNIQUES
WAVEFORMS FOR ADP1148 BUCK REGULATOR
DRIVING ADP3310 LOW DROPOUT REGULATOR
ADP1148 OUTPUT
(ADP3310 INPUT)
ADP3310 OUTPUT
40mV p-p
VERTICAL SCALE: 10mV/DIV
HORIZ. SCALE:
5µs/DIV
5mV p-p
VERTICAL SCALE: 10mV/DIV
HORIZ. SCALE:
5µs/DIV
Figure 10.42
There are many tradeoffs in designing power supply filters. The success of any
filter circuit is highly dependent upon a compact layout and the use of a large area
ground plane. As has been stated earlier, all connections to the ground plane should
be made as short as possible to minimize parasitic resistance and inductance.
Output ripple can be reduced by the addition of low ESL/ESR capacitors to the
output. However, it may be more efficient to use an LC filter to accomplish the
ripple reduction. In any case, proper component selection is critical. The inductor
should not saturate under the maximum load current, and its DC resistance should
be low enough as not to induce significant voltage drop. The capacitors should have
low ESL and ESR and be rated to handle the required ripple current.
Low dropout linear post regulators provide both ripple reduction as well as better
regulation and can be effective, provided the sacrifice in efficiency is not excessive.
Finally, it is difficult to predict the output ripple current analytically, and there is
no substitute for a prototype using the real-world components. Once the filter is
proven to provide the desired ripple attenuation (with some added safety margin),
care must be taken that parts substitutions or vendor changes are not made in the
final production units without first testing them in the circuit for equivalent
performance.
10.50
HARDWARE DESIGN TECHNIQUES
SWITCHING SUPPLY FILTER SUMMARY
■ Proper Layout and Grounding (using Ground Plane) Mandatory
■ Low ESL/ESR Capacitors Give Best Results
■ Parallel Capacitors Lower ESR/ESL and Increase Capacitance
■ External LC Filters Very Effective in Reducing Ripple
■ Linear Post Regulation Effective for Noise Reduction and Best
Regulation
■ Completely Analytical Approach Difficult, Prototyping is
Required for Optimum Results
■ Once Design is Finalized, Do Not Switch Vendors or Use Parts
Substitutions Without First Verifying Their Performance in
Circuit
■ High Frequency Localized Decoupling at IC Power Pins is Still
Required
Figure 10.43
LOCALIZED HIGH FREQUENCY POWER SUPPLY FILTERING
The LC filters described in the previous section are useful in filtering switching
regulator outputs. However, it may be desirable to place similar filters on the
individual PC boards where the power first enters the board. Of course, if the
switching regulator is placed on the PC board, then the LC filter should be an
integral part of the regulator design.
Localized high frequency filters may also be required at each IC power pin (see
Figure 10.44). Surface mount ceramic capacitors are ideal choices because of their
low ESL. It is important to make the connections to the power pin and the ground
plane as short as possible. In the case of the ground connection, a via directly to the
ground plane is the shortest path. Routing the capacitor ground connection to
another ground pin on the IC is not recommended due to the added inductance of
the trace. In some cases, a ferrite bead in series with the power connection may also
be desirable.
10.51
HARDWARE DESIGN TECHNIQUES
LOCALIZED DECOUPLING TO GROUND PLANE
USING SHORTEST PATH
CORRECT
POWER
SUPPLY
TRACE
OPTIONAL
FERRITE BEADS
DECOUPLING
CAPACITOR
INCORRECT
POWER
SUPPLY
TRACE
DECOUPLING
CAPACITOR
V+
IC
V+
VIAS TO
GROUND
PLANE
GND
IC
PCB
TRACE
GND
VIA TO
GROUND
PLANE
Figure 10.44
The following list summarizes the switching power supply filter layout/construction
guidelines which will help ensure that the filter does the best possible job:
(1) Pick the highest electrical value and voltage rating for filter capacitors which is
consistent with budget and space limits. This minimizes ESR, and maximizes filter
performance. Pick chokes for low ∆L at the rated DC current, as well as low DCR.
(2) Use short and wide PCB tracks to decrease voltage drops and minimize
inductance. Make track widths at least 200 mils for every inch of track length for
lowest DCR, and use 1 oz or 2 oz copper PCB traces to further reduce IR drops and
inductance.
(3) Use short leads or better yet, leadless components, to minimize lead inductance.
This minimizes the tendency to add excessive ESL and/or ESR. Surface mount
packages are preferred. Make all connections to the ground plane as short as
possible.
(4) Use a large-area ground plane for minimum impedance.
(5) Know what your components do over frequency, current and temperature
variations! Make use of vendor component models for the simulation of prototype
designs, and make sure that lab measurements correspond reasonably with the
10.52
HARDWARE DESIGN TECHNIQUES
simulation. While simulation is not absolutely necessary, it does instill confidence in
a design when correlation is achieved (see Reference 15).
HIGH DENSITY DSP LOCALIZED DECOUPLING CONSIDERATIONS
High pin-count DSP packages require special consideration with respect to localized
decoupling due to their high digital transient currents. Typical decoupling
arrangements are shown in Figure 10.45. The surface mount capacitors are placed
on the top side of the PC board in Figure 10.45A. For the SHARC-family, eight
0.02µF ceramic capacitors are recommended. They should be placed as close to the
package as possible. The connections to the VDD pins should be as short as possible
using wide traces. The connections to ground should be made directly to the ground
plane with vias. A less desirable method is shown in Figure 10.45B, where the
capacitors are mounted on the back side of the PC board underneath the footprint of
the package. If the ground plane underneath the package footprint is perforated
with many signal vias, the capacitor return transient current must flow to the
outside ground plane, which may be poorly connected to the inside ground plane due
to the vias.
DECOUPLING HIGH PIN-COUNT DSPs
IN PQFP PACKAGES
A
B
TOP VIEW
OF PACKAGE
TOP VIEW
OF PACKAGE
CAPACITORS ON BACK SIDE OF PC BOARD
CAPACITORS ON TOP SIDE OF PC BOARD
Figure 10.45
10.53
HARDWARE DESIGN TECHNIQUES
The PC board for a ball grid array (BGA) package is shown in Figure 10.46. Note
that all connections to the balls must be made using vias to other layers of the
board. The “dogbone” pattern shown is often used for the BGA packages. The
shaded area indicates the location of the solder mask. As in the case of PQFP
packages, the localized decoupling capacitors should be placed as close as possible to
the package with short connections to the VDD pins and direct connections to vias to
the ground plane layer.
DECOUPLING HIGH PIN-COUNT DSPs
IN BALL GRID ARRAY (BGA) PACKAGES
TOP VIEW
OF PC BOARD
WITHOUT
PACKAGE
"DOGBONE" PAD
VIA
SOLDER MASK
CROSS SECTION
VIEW
BGA PACKAGE
SOLDER BALLS
MULTILAYER
PC BOARD
VIAS
Figure 10.46
The ADSP-21160 400-ball BGA 27mm by 27mm package approximate power and
ground assignments are shown in Figure 10.47. The ball pitch is 1.27mm.
Approximately 84 balls are used in the center of the pattern for ground connections.
The connections to the core voltage (40 balls) and the external voltage (46 balls)
surround the ground balls. The remaining outer balls are used for the various
signals.
The centrally located ground balls serve a dual function. Their primary function is
make a low impedance connection directly to the ground plane layer. Their
secondary function is to conduct the package heat to the ground plane layer which
also acts as a heat sink, since the device must dissipate about 2.5W under average
operating conditions. The addition of an external heatsink as shown lowers the
junction-to-ambient thermal resistance even further.
10.54
HARDWARE DESIGN TECHNIQUES
ADSP-21160 DSP 400-PIN
PBGA PACKAGE BALL LOCATIONS
BALL LOCATIONS
ARE APPROXIMATE,
ALL BALLS NOT SHOWN
GROUND BALLS
(84)
VOLTAGE BALLS
(86)
PACKAGE SIZE:
27mm X 27mm
BALL PITCH: 1.27mm
SIGNAL BALLS
MULTILAYER
PC BOARD
GROUND AND HEATSINK
Figure 10.47
10.55
HARDWARE DESIGN TECHNIQUES
REFERENCES: NOISE REDUCTION AND FILTERING
1.
EMC Design Workshop Notes, Kimmel-Gerke Associates, Ltd.,
St. Paul, MN. 55108, (612) 330-3728.
2.
Walt Jung, Dick Marsh, Picking Capacitors, Parts 1 & 2, Audio,
February, March, 1980.
3.
Tantalum Electrolytic and Ceramic Capacitor Families, Kemet
Electronics, Box 5928, Greenville, SC, 29606, (803) 963-6300.
4.
Type HFQ Aluminum Electrolytic Capacitor and type V Stacked
Polyester Film Capacitor, Panasonic, 2 Panasonic Way, Secaucus,
NJ, 07094, (201) 348-7000.
5.
OS-CON Aluminum Electrolytic Capacitor 93/94 Technical Book,
Sanyo, 3333 Sanyo Road, Forrest City, AK, 72335, (501) 633-6634.
6.
Ian Clelland, Metalized Polyester Film Capacitor Fills High Frequency
Switcher Needs, PCIM, June 1992.
7.
Type 5MC Metallized Polycarbonate Capacitor, Electronic Concepts, Inc.,
Box 1278, Eatontown, NJ, 07724, (908) 542-7880.
8.
Walt Jung, Regulators for High-Performance Audio, Parts 1 and 2,
The Audio Amateur, issues 1 and 2, 1995.
9.
Henry Ott, Noise Reduction Techniques in Electronic Systems,
2d Ed., 1988, Wiley.
10.
Fair-Rite Linear Ferrites Catalog, Fair-Rite Products, Box J, Wallkill,
NY, 12886, (914) 895-2055, http://www.fair-rite.com.
11.
Type EXCEL leaded ferrite bead EMI filter, and type EXC L leadless
ferrite bead, Panasonic, 2 Panasonic Way, Secaucus, NJ, 07094,
(201) 348-7000.
12.
Steve Hageman, Use Ferrite Bead Models to Analyze EMI Suppression,
The Design Center Source, MicroSim Newsletter, January, 1995.
13.
Type 5250 and 6000-101K chokes, J. W. Miller, 306 E. Alondra Blvd.,
Gardena, CA, 90247, (310) 515-1720.
14.
DIGI-KEY, PO Box 677, Thief River Falls, MN, 56701-0677,
(800) 344-4539.
10.56
HARDWARE DESIGN TECHNIQUES
15.
Tantalum Electrolytic Capacitor SPICE Models, Kemet Electronics,
Box 5928, Greenville, SC, 29606, (803) 963-6300.
16.
Eichhoff Electronics, Inc., 205 Hallene Road, Warwick, RI., 02886,
(401) 738-1440, http://www.eichhoff.com.
17.
Practical Design Techniques for Power and Thermal Management,
Analog Devices, 1998, Chapter 8.
10.57
HARDWARE DESIGN TECHNIQUES
DEALING WITH HIGH SPEED LOGIC
Much has been written about terminating printed circuit board traces in their
characteristic impedance to avoid reflections. A good rule-of-thumb to determine
when this is necessary is as follows: Terminate the line in its characteristic
impedance when the one-way propagation delay of the PCB track is equal to or
greater than one-half the applied signal rise/fall time (whichever edge is faster). A
conservative approach is to use a 2 inch (PCB track length)/nanosecond (rise-, falltime) criterion. For example, PCB tracks for high-speed logic with rise/fall time of
1ns should be terminated in their characteristic impedance if the track length is
equal to or greater than 2 inches (including any meanders). Figure 10.48 shows the
typical rise/fall times of several logic families including the SHARC DSPs operating
on +3.3V supplies. As would be expected, the rise/fall times are a function of load
capacitance.
TYPICAL DSP OUTPUT RISE TIMES AND FALL TIMES
■ GaAs: 0.1ns
■ ECL: 0.75ns
■ ADI SHARC DSPs: 0.5 ns to 1 ns (Operating on +3.3V Supply)
ADSP-21060L
SHARC:
Figure 10.48
This same 2 inch/nanosecond rule of thumb should be used with analog circuits in
determining the need for transmission line techniques. For instance, if an amplifier
must output a maximum frequency of fmax, then the equivalent risetime, tr, can be
calculated using the equation tr = 0.35/fmax. The maximum PCB track length is
then calculated by multiplying the risetime by 2 inch/nanosecond. For example, a
maximum output frequency of 100MHz corresponds to a risetime of 3.5ns, and a
track carrying this signal greater than 7 inches should be treated as a transmission
line.
10.58
HARDWARE DESIGN TECHNIQUES
Equation 10.1 can be used to determine the characteristic impedance of a PCB track
separated from a power/ground plane by the board’s dielectric (microstrip
transmission line):
Z o (Ω) =
 5.98d 
ln 

 0.89w + t 
ε r + 1.41
87
Eq. 10.1
where εr = dielectric constant of printed circuit board material;
d = thickness of the board between metal layers, in mils;
w = width of metal trace, in mils; and
t = thickness of metal trace, in mils.
The one-way transit time for a single metal trace over a power/ground plane can be
determined from Eq.10.2:
t pd (ns / ft) = 1.017 0.475ε r + 0.67
Eq. 10.2
For example, a standard 4-layer PCB board might use 8-mil wide, 1 ounce (1.4 mils)
copper traces separated by 0.021" FR-4 (εr=4.7) dielectric material. The
characteristic impedance and one-way transit time of such a signal trace would be
88Ω and 1.7ns/ft (7"/ns), respectively.
The best ways to keep sensitive analog circuits from being affected by fast logic are
to physically separate the two and to use no faster logic family than is dictated by
system requirements. In some cases, this may require the use of several logic
families in a system. An alternative is to use series resistance or ferrite beads to
slow down the logic transitions where the speed is not required. Figure 10.49 shows
two methods. In the first, the series resistance and the input capacitance of the gate
form a lowpass filter. Typical CMOS input capacitance is 5pF to10pF. Locate the
series resistor close to the driving gate. The resistor minimizes transient currents
and may eliminate the necessity of using transmission line techniques. The value of
the resistor should be chosen such that the rise and fall times at the receiving gate
are fast enough to meet system requirement, but no faster. Also, make sure that the
resistor is not so large that the logic levels at the receiver are out of specification
because of the voltage drop caused by the source and sink current which flow
through the resistor. The second method is suitable for longer distances (>2 inches),
where additional capacitance is added to slow down the edge speed. Notice that
either one of these techniques increases delay and increases the rise/fall time of the
original signal. This must be considered with respect to the overall timing budget,
and the additional delay may not be acceptable.
Figure 10.50 shows a situation where several DSPs must connect to a single point,
as would be the case when using read or write strobes bidirectionally connected
from several DSPs. Small damping resistors shown in Figure 10.50A can minimize
ringing provided the length of separation is less than about 2 inches. This method
will also increase rise/fall times and propagation delay. If two groups of processors
must be connected, a single resistor between the pairs of processors as shown in
Figure 10.50B can serve to damp out ringing.
10.59
HARDWARE DESIGN TECHNIQUES
DAMPING RESISTORS SLOW DOWN FAST
LOGIC EDGES TO MINIMIZE EMI/RFI PROBLEMS
R
LOGIC
GATE
< 2 inches
LOGIC
GATE
CIN
Risetime = 2.2 R·CIN
> 2 inches
R
LOGIC
GATE
LOGIC
GATE
C
CIN
Risetime = 2.2 R·(C + CIN)
Figure 10.49
SERIES DAMPING RESISTORS
FOR SHARC DSP INTERCONNECTIONS
A
STAR CONNECTION
DAMPING RESISTORS
SHARC
DSP
USE FOR RD, WR
STROBES
SHARC
DSP
<2"
10Ω
Ω
EACH
SHARC
DSP
SHARC
DSP
NOTE: THESE TECHNIQUES
INCREASE RISE/FALL TIMES
AND PROPAGATION DELAY
B
SINGLE DAMPLING
RESISTOR BETWEEN
PROCESSOR GROUPS
SHARC
DSP
SHARC
DSP
20Ω
Ω
SHARC
DSP
Figure 10.50
10.60
<2"
SHARC
DSP
HARDWARE DESIGN TECHNIQUES
The only way to preserve 1ns or less rise/fall times over distances greater than
about 2 inches without ringing is to use transmission line techniques. Figure 10.51
shows two popular methods of termination: end termination, and source
termination. The end termination method (Figure 10.51A) terminates the cable at
its terminating point in the characteristic impedance of the microstrip transmission
line. Although higher impedances can be used, 50Ω is popular because it minimizes
the effects of the termination impedance mismatch due to the input capacitance of
the terminating gate (usually 5pF to 10pF). In Figure 10.51A, the cable is
terminated in a Thevenin impedance of 50Ω terminated to +1.4V (the midpoint of
the input logic threshold of 0.8V and 2.0V). This requires two resistors (91Ω and
120Ω), which add about 50mW to the total quiescent power dissipation to the
circuit. Figure 10.51A also shows the resistor values for terminating with a +5V
supply (68Ω and 180Ω). Note that 3.3V logic is much more desirable in line driver
applications because of its symmetrical voltage swing, faster speed, and lower
power. Drivers are available with less than 0.5ns time skew, source and sink
current capability greater than 25mA, and rise/fall times of about 1ns. Switching
noise generated by 3.3V logic is generally less than 5V logic because of the reduced
signal swings and lower transient currents.
TERMINATION TECHNIQUES FOR CONTROLLED
IMPEDANCE MICROSTRIP TRANSMISSION LINES
TYPICAL DRIVERS:
■ 74FCT3807/A (IDT)
■ 74ACTQ240 (Fairchild)
A
ZO = 50Ω
Ω
+1.4V
+3.3V
+5.0V
120Ω
Ω
30mW
180Ω
Ω
72mW
+1.4V
91Ω
Ω
22mW
GROUND PLANE
B
68Ω
Ω
29mW
END TERMINATION
ZO ≈ 10Ω
Ω
SOURCE TERMINATION
39Ω
Ω
RULE OF THUMB:
ZO = 50Ω
Ω
USE TRANSMISSION LINE IF DISTANCE IS
MORE THAN 2"/ns OF LOGIC RISE/FALL TIME
50Ω
Ω PC BOARD TRANSMISSION LINE DELAY ≈ 1ns / 7"
Figure 10.51
The source termination method, shown in Figure 10.51B, absorbs the reflected
waveform with an impedance equal to that of the transmission line. This requires
about 39Ω in series with the internal output impedance of the driver, which is
generally about 10Ω. This technique requires that the end of the transmission line
be terminated in an open circuit, therefore no additional fanout is allowed. The
10.61
HARDWARE DESIGN TECHNIQUES
source termination method adds no additional quiescent power dissipation to the
circuit.
Figure 10.52 shows a method for distributing a high speed clock to several devices.
The problem with this approach is that there is a small amount of time skew
between the clocks because of the propagation delay of the microstrip line
(approximately 1ns /7"). This time skew may be critical in some applications. It is
important to keep the stub length to each device less than 0.5" in order to prevent
mismatchs along the transmission line.
CLOCK DISTRIBUTION USING
END-OF-LINE TERMINATION
+3.3V
TRANSMISSION LINE ZO = 50Ω
Ω
CLOCK
+1.4V
120Ω
Ω
30mW
91Ω
Ω
22mW
SHARC
DSP
SHARC
DSP
SHARC
DSP
50Ω
Ω PC BOARD TRANSMISSION LINE DELAY ≈ 1ns / 7"
NOTE: KEEP STUB LENGTH < 0.5"
NOT RECOMMENDED FOR SYNCHRONIZED SHARC OPERATION!
Figure 10.52
The clock distribution method shown in Figure 10.53 minimizes the clock skew to
the receiving devices by using source terminations and making certain the length of
each microstrip line is equal. There is no extra quiescent power dissipation as would
be the case using end termination resistors.
Figure 10.54 shows how source terminations can be used in bi-directional link port
transmissions between SHARC DSPs. The output impedance of the SHARC driver
is approximately 17Ω, and therefore a 33Ω series is required on each end of the
transmission line for proper source termination.
The method shown in Figure 10.55 can be used for bi-directional transmission of
signals from several sources over a relatively long transmission line. In this case,
the line is terminated at both ends, resulting in a DC load impedance of 25Ω.
SHARC drivers are capable of driving this load to valid logic levels.
10.62
HARDWARE DESIGN TECHNIQUES
PREFERRED METHOD OF CLOCK DISTRIBUTION
USING SOURCE TERMINATED TRANSMISSION LINES
ZO ≈ 10Ω
Ω
*
39Ω
Ω
> 4"
ZO = 50Ω
Ω
SHARC
DSP
ZO = 50Ω
Ω
SHARC
DSP
ZO = 50Ω
Ω
SHARC
DSP
ZO ≈ 10Ω
Ω
CLOCK
*
39Ω
Ω
ZO ≈ 10Ω
Ω
* Same
Package
*
39Ω
Ω
Figure 10.53
SOURCE TERMINATION FOR BI-DIRECTIONAL
TRANSMISSION BETWEEN SHARC DSPs
ADSP-2106x
ADSP-2106x
OFF
LENGTH > 6"
33Ω
Ω
ZO = 50Ω
Ω
33Ω
Ω
ON
Ω
ZO ≈ 17Ω
LINK PORT
TRANSMITTER
LINK PORT
RECEIVER
Figure 10.54
10.63
HARDWARE DESIGN TECHNIQUES
SINGLE TRANSMISSION LINE
TERMINATED AT BOTH ENDS
SHARC
DSP
SHARC
DSP
SHARC
DSP
+3.3V
+3.3V
120Ω
Ω
30mW
Ω
ZO = 50Ω
LENGTH > 10"
+1.4V
91Ω
Ω
22mW
120Ω
Ω
30mW
91Ω
Ω
22mW
SHARC
DSP
SHARC
DSP
SHARC
DSP
NOTE: KEEP STUB LENGTH < 0.5"
NOT RECOMMENDED FOR CLOCKS IN SYNCHRONIZED SHARC OPERATION!
Figure 10.55
10.64
HARDWARE DESIGN TECHNIQUES
REFERENCES ON DEALING WITH HIGH SPEED
LOGIC:
1.
Howard W. Johnson and Martin Graham, High-Speed Digital Design,
PTR Prentice Hall, 1993.
2.
EDN’s Designer’s Guide to Electromagnetic Compatibility, EDN,
January, 20, 1994, material reprinted by permission of Cahners Publishing
Company, 1995.
3.
Designing for EMC (Workshop Notes), Kimmel Gerke Associates, Ltd., 1994.
4.
Mark Montrose, EMC and the Printed Circuit Board, IEEE Press,
1999 (IEEE Order Number PC5756).
10.65
HARDWARE DESIGN TECHNIQUES
10.66
INDEX
INDEX
■
SUBJECT INDEX
■
ANALOG DEVICES PARTS INDEX
Index-a
INDEX
Index-b
INDEX
Subject Index
A
AA Alkaline Battery Discharge
Characteristics, 10.14
Absolute value amplifier, 3.29
Active and Passive Electrical Wave Catalog,
2.42
AD185x, multibit sigma-delta DAC, 4.11
AD260/AD261:
digital isolators
key specifications, 10.38
schematic, 10.38
AD820, op amp, 3.8-9
AD1819B SoundPort, codec, for
audio/modem, 9.37-38
AD1836, mixed-signal IC, codec, 9.38-39
AD1852, sigma-delta audio DAC, 3.17
AD1853:
dual 24-bit DAC, 4.12
sigma-delta audio DAC, 3.17
AD1854, sigma-delta audio DAC, 3.17
AD1877:
16-bit sigma-delta ADC
characteristics, 3.16
FIR filter characteristics, 3.17
AD189x, sample rate converters, 6.35
AD77xx, 24-bit sigma-delta ADC, low
frequency/high resolution, filters,
3.17
AD77XX-Series Data Sheets, 3.33
AD773x, 24-bit ADC, with PGA, 1.5
AD977x, 14-bit TxDAC, 4.6
AD983x, 10-bit DDS system, 4.16
AD985x, 14-bit TxDAC, 4.6
AD5322:
12-bit dual DAC
block diagram, 8.18
power-down feature, 8.17
serial interface, 8.18-19
AD5340:
12-bit parallel input DAC
block diagram, 8.10
interface diagram, 8.12
wait states, 8.11
AD6521, voiceband/baseband mixed-signal
CODEC, 9.21-22
AD6522, DSP-based baseband processor,
9.21-22
AD6523, Zero-IF Transceiver, 9.23
AD6524, Multi-Band Synthesizer, 9.23
AD6600, diversity receiver ADC, 9.30
AD6622, quad digital TSP, 9.30-31
AD6624, quad digital RSP, 9.30-31
AD6640:
12-bit ADC
with MagAmps, 3.24-25
multitone testing, diagram, 2.30
AD6644, 14-bit ADC, 9.30-31
AD7722, 16-bit ADC, 10.15
AD7725:
16-bit sigma-delta ADC
with on-chip PulseDSP filter, 6.2
programmable digital filter, 9.40-42
block diagram, 9.41
Systolix PulseDSP, 9.41
sigma-delta ADC, with Systolix PulseDSP
processor, 3.19-20
AD7730, sigma-delta ADC, 10.15
AD7731, sigma-delta ADC, 10.15
AD7853/7853L:
12-bit ADC, block diagram, 8.16
serial ADC interface, 8.17
AD7853L:
serial clock, 8.15-17
output timing, 8.16
AD7854/AD7854L:
12-bit ADC
block diagram, 8.5
interface diagram, 8.6-7
key interface timing comparisons, 8.5-6
AD7858/59:
12-bit ADC, 3.8-9
SAR ADC, circuit, 3.8-9
AD7892, 12-bit SAR ADC, 10.15
AD9042:
12-bit ADC, NPR, diagram, 2.32
12-bit wideband ADC, high SFDR,
2.27-28
AD9201, 10-bit, dual-channel ADC, 8.22-23
AD9220:
12-bit ADC, SINAD and ENOB, diagram,
2.25
12-bit CMOS ADC, latency or pipeline
delay, 3.27
AD9221, 12-bit CMOS ADC, latency or
pipeline delay, 3.27
AD9223, 12-bit CMOS ADC, latency or
pipeline delay, 3.27
AD9288-100, 8-bit dual ADC, functional
diagram, 3.32
AD9410, 10-bit flash ADC, interpolation,
diagram, 3.23
AD9761, 10-bit, dual channel DAC, 8.22-23
AD9772:
14-bit interpolating DAC, 9.31
14-bit oversampling interpolating
TxDAC, 4.9-10
block diagram, 4.10
14-bit TxDAC, 4.6
Index-1
INDEX
diagram, 4.6
SFDR, 2.39-40
TxDAC 14-bit DAC, 9.30-31
AD9850, DDS/DAC synthesizer system,
diagram, 4.16
AD73322:
block diagram, 8.19-20
simplified interface timing, 8.21
AD73422, dspConverter, specifications, 8.22
AD20msp430:
baseband processing chipset, 9.21-23
components, 9.22-23
construction, 9.21
Adams, R.W., 3.33, 3.34
Adaptive filter, 6.35-39
basic concept, 6.35-36
speech compression and synthesis, 6.36
ADC, 1.2
1-bit, comparator, 3.11
3-bit unipolar, transfer characteristics,
2.8-9
10-bit, theoretical NPR, 2.32
11-bit, theoretical NPR, 2.32
12-bit
pipelined architecture, diagram, 3.7
theoretical NPR, 2.32
24-bit, with PGAs, 1.5
analog bandwidth, 2.25-26
analog input, 2.10
bit-per-stage, 3.1
boot memory select, 8.2
communications application, SFDR, 2.26
conversion complete output, 8.1-2
conversion process, 2.2
convert start, 8.1-2
data memory address, 8.4
data memory select, 8.2
digital output, grounding, 10.22-24
DNL errors, 2.22
DSP applications, 3.1-35
high speed architectures, 3.2
sigma-delta, 3.2
successive approximation, 3.1-9
types, 3.1
dynamic performance, quantification,
2.21
ENOB versus frequency, diagram, 2.26
equivalent input referred noise, 2.20-21
excess DNL, and missing codes, 2.12-13
flash, 3.1
high speed, pipelined, latency, 3.6, 3.8
ideal 12-bit
FFT, noise floor, 2.19
SFDR sampling clock to input
frequency ratio, 2.18
ideal, distortion and noise, 2.15-19
ideal N-bit
Index-2
dynamic performance analysis,
diagram, 2.17
errors, 2.15
quantization noise, diagram, 2.16
input/output memory select, 8.2
low power/low voltage, design issues, 3.1
memory address bus, 8.2
memory read, 8.2
memory select line, 8.2
non-ideal 3-bit, transfer function,
diagram, 2.13
non-monotonic, 2.12
normalized signal to reference, 2.10
output, quantization, 2.7
output enable/read, 8.2
oversampling ratio, 3.11
pipelined, 3.1
latency or pipeline delay, 3.25
timing, diagram, 3.7
practical
distortion and noise, 2.19-20
noise and distortion sources, 2.20
processor interrupt request line, 8.1-2
program memory select, 8.2
quantization error, 2.9-10
quantization uncertainty, 2.9-10
quantized output, 2.9, 2.10
ripple, 3.1
sampling clock jitter, 2.33
sampling frequency, versus antialiasing
characteristics, 2.5
SAR, 3.3
external high frequency clock, 3.6
fundamental timing, 3.5
resolutions, table, 3.5
typical timing, 3.5
sigma-delta, 3.1-2, 3.9-21
circuitry, 3.9-10
as oversampling converter, 2.6
programmable digital filter, 9.40-42
VLSI technology, 3.9
see also Sigma-delta ADC
sign-magnitude, use, 2.10
single-tone sinewave FFT testing, 2.18
SNR decrease with input frequency, 2.33
static transfer functions, DC errors,
2.7-14
subranging, 3.1
pipelined converter, 3.6
successive approximation, 3.1-9
basic, 3.3
resolutions, table, 3.5
see also SAR ADC
thermal noise, 2.20-21
Address bus, DAC, 8.9
ADI DSP collaborative, 7.54
INDEX
ADI modified Harvard architecture, in
microprocessor, 7.6-7
ADI SHARC floating point DSPs, 7.26-30
ADMC300, motor controller, 9.34
ADMC331, motor controller, 9.34
ADMC401, motor controller, 9.34
ADMCF326, motor controller, 9.34
ADMCF328:
motor controller, 9.34
block diagram, 9.35
ADP1148:
synchronous buck regulator, 10.46-50
circuit, 10.47-48
driving ADP3310
circuit diagram, 10.49
waveforms, 10.50
filtered output, 10.49
output waveform, 10.47-48
ADP3310, linear LDO regulator, 10.47,
10.49-50
ADPCM, adaptive pulse code modulation,
1.2
ADSL, 9.11-16
advantages, 9.12-13
block diagram, 9.13
data transmission capability, 9.15
definition, 9.12
installation advantages, 9.14
modem, block diagram, 9.15
modems, three-channel approach, 9.13-14
ADSP-21ESP202, codec, embedded speech
processing, 9.36-37
ADSP-21mod870:
digital modem processor, 9.10
expanding central office capability, 9.9
in voice-based RAS modem, 9.8-9
ADSP-21xx:
16-bit fixed point DSP core, 7.6-23
arithmetic, signed fractional format,
7.23-24
assembly code for FIR filter, 1.8
buses, 7.10-11, 7.11
computational units, 7.11-13
core architecture
diagram, 7.9
summary, 7.10
data address generators and program
sequencer, 7.13-14
digital filter example, 7.7-10
DSP optimized, 7.9
FIR filter assembly code, 7.8
fixed-point DSP, 7.7-8
FIR filter, 6.11-12
assembly code, 6.13
internal buses, 7.10-11
internal peripherals, powerdown, 7.18
memory-write cycle timing diagram, 8.7-8
on-chip peripherals, 7.14-23
byte DMA port, 7.17
internal DMA, 7.17
SPORTs, 7.16
read-cycle, timing diagram, 8.3
serial ports
block diagram, 8.13
features, 8.13
operation, 8.12-14
SPORTs, 7.16
ADSP-218x:
architecture, 7.15
byte memory interface, 7.17
DSP, EZ-ICE, 7.49
implementation, multi-channel VOIP
server, diagram, 9.11
interface with CODEC, 8.20-22
internal direct-memory-access port, 7.16
memory-mapped peripherals, 7.16
modified Harvard architecture, 7.16
multiple core devices, 7.22
on-chip peripherals, memory interface,
7.15
roadmap, 7.22
VisualDSP software, 7.52
ADSP-219x:
architecture, 7.19
code compatibility with ADSP-218x, 7.19
key specifications, 7.20
roadmap, 7.23
VisualDSP software, 7.52
ADSP-2100, core in ADSP-218x
architecture, 7.15
ADSP-2100 EZ-KIT Lite Reference Manual,
7.55, 8.25
ADSP-2100 Family EZ Tools Manual, 7.55,
8.25
ADSP-2100 Family Users Manual, 3rd
Edition, 7.55, 8.25
ADSP-2106x:
characteristics, 7.28
external ports, 7.29
host interface, 7.29
I/O processor, 7.29
IEEE Standard P1149.1 Joint Test Action
Group standard, 7.30
instruction set, 7.28
internal memory, 7.30
multiprocessing systems, 7.29
on-chip DMA, 7.30
SHARC processors, 7.26-27
Super Harvard Architecture, 7.26-28
ADSP-2106x SHARC EZ-KIT Lite Manual,
7.55, 8.25
ADSP-2106x SHARC User's Manual, 2nd
Edition, July 1996, 7.55, 8.25
ADSP-2116x:
Index-3
INDEX
32-bit DSP, second-generation, 7.31
SIMD core architecture, 7.31-36
diagram, 7.31
SIMD DSP, 7.33
SIMD features, 7.32
ADSP-2189M:
69-tap FIR filter, design example, 6.22
75 MIPS DSP
filter subroutine, 6.12
throughput time, 6.28
75 MIPS processor, 6.20
EZ-KIT Lite, 7.47
fixed-point DSP, 6.1
key timing specifications, 8.9-10
parallel read timing, 8.4, 8.7
parallel write interface, timing
specifications, 8.11
serial port, receive timing diagram,
8.14-15
system interface, full memory mode,
diagram, 8.24
ADSP-21000 Family Application Handbook,
5.25, 6.40
ADSP-21060L SHARC, DSP output rise and
fall times, 10.58
ADSP-21065L:
connected to ADC and DAC, 8.22-23
EZ-KIT Lite, 7.48
SHARC, DSP benchmarks, 7.34
ADSP-21065L SHARC EZ-LAB User's
Manual, 7.55, 8.25
ADSP-21065L SHARC User's Manual, Sept.
1, 1998, 7.55, 8.25
ADSP-21160:
16-channel audio mixer, SIMD
architecture, 9.40
32-bit SHARC, key features, 7.32
DSP, BGA package locations, 10.54-55
integrated peripherals, 7.32
SIMD/multiple channels, DSP
benchmarks, 7.34
SISD, DSP benchmarks, 7.34
ADSP-21160 SHARC DSP Hardware
Reference, 7.55, 8.25
ADSP-21160M, EZ-KIT Lite, 7.47
ADSP-TS001:
features, 7.40
TigerSHARC
16-bit fixed-point DSP, 5.18-19
architecture, 7.36-44
benchmarks, 7.44
diagram, 7.38
ADuM1100A, digital isolator, 10.36-37
Advanced mobile phone service see AMPS
Agilent HCPL-7720, 10.36
Aiken, Howard, 7.7
Alfke, P., 10.14
Index-4
Aliasing, 2.2-3
frequency domain, representation, 2.3
All-pole lattice filter, parameters, from
speech samples, 6.38-39
Alternate framing mode, 8.14
Aluminum electrolytic capacitor, 10.40-41
Amplifier Applications Guide (1992), 2.43,
3.35
Amplitude shift keying, POTS, 9.2
Analog bandwidth, ADC, 2.25-26
Analog cellular basestation, 9.27-28
Analog Devices Inc., 16-bit DSP, roadmap,
7.22
Analog Devices' Motor Control Website, 9.44
Analog filter:
frequency response, 6.4
popular types, 6.24, 6.27-28
requirements, 6.3
for oversampling, 4.10
versus digital, 6.3
Analog front end, 8.19-22
Analog receiver design, 9.26-27
Analog return current, 10.16-17
Analog signal:
characteristics, 1.1-2
discrete time sampling, 2.2-7
diagram, 2.1
normalized ratio, 2.9
quantization, 2.2
diagram, 2.1
sampling, aliasing, 2.3
Analog superheterodyne receiver, 9.27
Andreas, D., 3.33
ANSI/IEEE Standard 754-1985, floating
point arithmetic, 7.24-25
Anti-imaging filter, 6.2
Antialiasing filter, 6.2
baseband sampling, oversampling, 2.5
requirements, relaxing, 2.6
specifications, 2.4-5
Aperture delay, 2.33-35
Aperture jitter, 2.33-35, 10.25
Apex-ICE, 7.48-49
USB simulator, 7.49
Application of Digital Signal Processing in
Motion Control Seminar (2000),
9.44
Architecture:
computer
ADI modified Harvard, 7.6
Harvard, 7.6-7
Von Neumann, 7.6
Arithmetic logic unit:
in DSP, 7.4
features, 7.11
Armstrong, Edwin H., Major, 9.27
INDEX
The ARRL Handbook for Radio Amateurs,
4.17
Asymmetric digital subscriber line see
ADSL
Audio system, synthesized, 1.3
Auto-correlation, 1.3
Automotive/home theater, using 32-bit
SHARC, 9.39
B
Backplane ground plane, 10.18
Baines, Rupert, 9.43
Baker, Bonnie, 10.34
Ball grid array, 10.54
Band filter, 10.46
Bandpass filter:
design, 6.23-24
from lowpass and highpass filters, 6.24
Bandpass sampling, 2.6-7
Bandpass sigma-delta ADC, undersampling,
3.18
Bandstop filter:
design, 6.23-24
equivalent impulse response, 6.23
from lowpass and highpass filters, 6.24
Baseband data signal, POTS, 9.2
Baseband sampling:
antialiasing filter, 2.4-6
oversampling, 2.5
Nyquist zone, 2.6
Basestation, block diagram, 9.30
Basis function, 5.3
correlation, DFT, 5.4
BDC binary coding, data converters, 2.10
Bennett, W.R., 2.42
Bessel filter, 6.24, 6.27-28
Best straight line, 2.11-12
Best, R.E., 4.17
Bilinear transformation, 6.28
Binary ADC:
3-bit
diagram, 3.29
input and residue waveforms, 3.29
single-stage, diagram, 3.28
Binary coding, data converters, 2.10
Binary DAC, 5-bit, architectures, diagram,
4.4
Bingham, John, 9.43
Bipolar converter, types, 2.10
Biquad, in IIR filter, 6.25
Bit reversal:
for 8-point DFT, 5.14
algorithm, 5.11-14
Bit-per-stage ADC, 3.27-32
diagram, 3.27
Blackman window function, 5.22-23
Blackman, R.B., 2.43
Block floating point, in FFT, 5.17
Boot memory select, ADC, 8.2
Bordeaux, Ethan, 7.55, 10.1
Boser, B., 3.33
Boyd, I., 9.43
Branch target buffer, 7.39, 7.41
Brannon, Brad, 9.43
Brokaw, Paul, 10.34
Bryant, James, 2.1, 3.1, 3.9, 4.1, 10.15,
10.34
Buck regulator, 10.46-50
Buffer register, 10.22-23
Bus:
data memory address, 7.10
data memory data, 7.10
data transfer, 7.10-11
internal result, 7.10-13
program memory address, 7.10
program memory data, 7.10
Buss wire, 10.16
Butterfly, 5.11-12
DIF FFT, 5.15
DIT FFT, 5.12
Butterworth filter, 4.9, 6.24, 6.27-28
characteristics, 2.5
Byrne, Mike, 10.15
C
Cage jack, 10.16
Calhoun, George, 9.43
Capacitive coupling, doublet glitch, 2.37
Capacitor:
equivalent circuit, pulse response, 10.43
finite ESR, 10.42-43
parasitic elements, 10.43
types, 10.40-42
Card entry filter, 10.46
Carrier, 2.23
Cascaded biquads, 6.25
Cauer filter, 6.27-28
CCD image processing, 1.5
CDMA, digital telephone system, 9.17
Cellular phone:
basic system, diagram, 9.16
frequency reuse, diagram, 9.16
Ceramic, capacitor, 10.40-42
Charpentier, A., 3.33
Chebyshev filter, 1.6, 6.2-3, 6.24, 6.27-28
Chestnut, Bill, 10.39
Chip select, DAC, 8.9
Circular buffering:
DSP application, 7.5
DSP requirement, 7.5
in FIR filter, 6.10
FIR filter, 7.5
in FIR filter, output calculation, 6.11
FIR filter pseudocode, in DSP, 7.8
Index-5
INDEX
Clelland, Ian, 10.56
Clock distribution:
end-of-line termination, diagram, 10.62
source terminated transmission lines,
diagram, 10.63
Cluster multiprocessing, 7.34
SHARC family, 7.36
CMOS IC, secondary I/O ring, diagram, 10.9
CMOS IC output driver, configuration,
10.3-4
Code division multiple access see CDMA
Code transition noise, and DNL, effects,
2.14
Codec, 1.5, 8.1
interfacing, 8.19-22
sampling rate, 8.20
voiceband/audio applications, 9.36-40
CODEC and DSP, in voiceband and audio,
9.36-40
COder/DECoder, 8.1
Coding, types, 2.10
Coleman, Brendan, 2.43
Colotti, James J., 2.43
Comfort noise insertion, 9.20
Communications, external port, versus link
port, 7.35
Computational unit:
arithmetic logic, 7.11
multiplier-accumulator, 7.11-12
shifter, 7.11-12
Computer, general purpose, Von Neumann
architecture, 7.6
Computing applications:
CISC, 7.2
data manipulation, 7.1
mathematical calculation, 7.1
RISC, 7.2-3
tabular summary, 7.1
Concentrator, 9.7
Connelly, J.A., 10.34
Convert start, ADC, 8.1-2
Convolution, 1.3
loop, 6.12
Convolving, filter responses, 6.23
Cooley, J.W., 5.10, 5.25
Cosier, G., 9.43
Coussens, P.J.M., 9.43
Crystal Oscillators, 10.34
D
DAC, 1.2
3-bit switched capacitor, diagram, 3.4
3-bit unipolar, transfer characteristics,
2.8-9
address bus, 8.9
binary weighted, 4.1
in CD player, use of interpolation, 6.33
Index-6
chip select, 8.9
conversion process, 2.2
current output, diagram, 4.2
in direct digital synthesis, 4.9
distortion, specification, 2.38
DNL errors, 2.22
double-buffered, latches, 4.7
DSP applications, 4.1-17
dynamic performance, 2.35-41
high-speed, 4.9
interpolating, 4.9-10
interrupt request, 8.9
ladder networks, 4.1
logic, 4.7-8
low distortion, architectures, 4.3-7
memory select, 8.9
midscale glitch, diagram, 2.37
monotonicity, 2.12
non-ideal 3-bit, transfer function,
diagram, 2.13
non-monotonic, and DNL, 2.12
output, quantization, 2.7
reconstruction, output, 2.41
segmented voltage, diagram, 4.3
settling time, applications, 2.38
sigma-delta, 4.11-12
1-bit, in CD players, 4.9
sin(x)/x frequency rolloff, 2.41
static transfer functions, DC errors,
2.7-14
structures, 4.1-3
transitions, with glitch, 2.36
voltage output, Kelvin divider, 4.1
write, 8.9
Damping resistor:
EMI/RFI minimization, 10.60
series, for SHARC DSP interconnections,
10.60
DashDSP, motor controllers, 9.34-35
Data address generator:
features, 7.13
mode status register, 7.13
Data converter:
analog bandwidth specification, 2.26
applications, 2.8
DC errors, 2.10, 2.10-11
DC performance, 2.15
gain error, diagram, 2.11
integral and differential non-linearity
distortion effects, 2.22-23
offset error, diagram, 2.11
sampling and reconstruction systems, 2.8
Data manipulation, by computer, 7.1
Data memory address, ADC, 8.4
Data memory address bus, 7.10
Data memory buffer, circular, 7.5
Data memory data bus, 7.10
INDEX
Data memory select, ADC, 8.2
Data reduction algorithms, in signal
processing, 1.2
Data sampling, 2.1-44
block diagram, 2.1
distortion and noise, 2.15-19
FFT, 2.1-2
quantization error, 3.10
quantization noise, 3.10
real-time system, 2.1
Data scrambling, in sigma-delta DAC, 4.11
Data-flow multiprocessing, 7.34
Dattorro, J., 3.33
DC error, types, 2.10
DDS, 4.12-16
basic architecture, 4.13
code-dependent glitches, 2.38
DAC, 4.3
distortion, contributors, 2.39
problems, 4.13
system, flexible, diagram, 4.14
tuning equation, 4.15
Dead time, DAC settling time, 2.35
Decimation, 3.10, 3.11
multirate filter, diagram, 6.31
Decimation-in-frequency, DIF, 5.13-14
Decimation-in-time, DIT, 5.11
Decoupling, localized, diagram, 10.52
Decoupling points, diagram, 10.24
Deglitching, 3.25
Del Signore, B.P., 3.33
Delta phase register, 4.14
Designing for EMC (Workshop Notes), 10.65
DFT, 5.1-8
8-point, 5.9, 5.12
using DIT, 5.12-13
applications, 5.2
basis functions, 5.3
butterfly operation, 5.11-12
characteristics, 5.4
complex, 5.5
equations, 5.6
from real, 5.7
input/output, 5.6
real/imaginary values, 5.7
equations, real versus imaginary, 5.8
expansion, 5.9-10
FIR filter, 6.15
fundamental analysis equation, 5.3
inverse, 5.5
output spectrum, 5.3
real, 5.5
input/output, 5.6
real versus imaginary
equations, 5.7-8
output conversion, 5.8
relationship, 5.7-8
sampled time domain signal, 5.3
versus FFT, 5.11
DIF:
butterfly, 5.13-14
decimation-in-frequency, 5.13-14
Differential non-linearity:
DNL, 2.12
from encoding process, 2.22
DIGI-KEY, 10.56
Digital cellular basestation, 9.28-31
Digital cellular basestations, advantages,
9.28
Digital cellular telephone, 9.16-20
GSM system, 9.18-20
Digital communications services see DCS
Digital correction, for subranging ADC, 3.24
Digital filter, 6.1-39
coefficient values, 6.1
design procedure, 6.1
diagram, 6.2
filter coefficient modification, 6.35
frequency response, 6.4
general equation, 6.25-26
lattice, 6.5
moving average, 6.5-10
programming ease in ADSP-21xx, 7.7-8
real-time, 6.1, 6.3
processing requirements, 6.4
requirements, 6.3
types, 6.5
versus analog, 6.3
Digital filtering, 3.10
Digital FIR filter, 1.6
Digital isolation:
by LED/photodiode optocouplers, 10.36
by LED/phototransistor optocouplers,
10.35
Digital isolation technique, 10.35-38
Digital isolator, 10.36-38
Digital mobile radio, standards, 9.18
Digital receiver, characteristics, 9.27-28
Digital return current, 10.16-17
Digital signal, characteristics, 1.1-2
Digital Signal Processing Applications
Using the ADSP-2100 Family, 5.25,
6.40, 9.43
Digital signal processor see DSP
Digital system, with ADC, quantization
noise, 2.31
Digital telephone system, CDMA and
TDMA, 9.17
Digital transmission, using adaptive
equalization, 6.37
Direct data scrambling, in sigma-delta
DAC, 4.11-12
Direct digital synthesis see DDS
Direct form 1 biquad filter, diagram, 6.26
Index-7
INDEX
Direct IF to digital conversion, 2.6-7
Direct memory access, DSP controller, 8.22
Direct-memory-access see DMA
Discontinuous transmission, 9.18-19, 9.20
Discrete Fourier transform see DFT
Discrete time Fourier series, 5.1
Discrete time sampling, 2.2
analog signal, 2.2-7
Distortion, DAC performance, 2.35-41
Distortion and noise, practical ADCs,
2.19-20
DIT:
decimation-in-time, 5.11
FFT, algorithm, 5.11, 5.13
Dither signal, 2.18
DMA, internal, in ADSP-21xx, 7.17
DNL:
ADC and DAc errors, 2.22
converter error, definition, 2.12
differential non-linearity, 2.12
Doernberg, Joey, 2.43
Double precision, 64-bit, floating point
arithmetic, 7.25
Double-buffered DAC:
advantages, 4.7-8
diagram, 4.8
latches, 4.7
Doublet glitch, 2.37
DSP, 1.2, 7.1-3
16-bit, fixed-point family, history, 7.20-22
32-bit, second-generation, 7.31
ADSP-21xx, characteristics, 1.6-8
analog versus digital, 1.4-5
options, diagram, 1.5
applications, 9.1-44
summary, 9.42
arithmetic
fixed-point versus floating point,
7.23-26
fixed-point versus floating-point,
7.23-26
characteristics, 7.2
code, compilation, 7.45
and computer applications, 7.1
core voltage, 10.2
development tools, 7.46
dot-product, 7.2
efficiency, 7.18
evaluation and development tools,
7.45-54
VisualDSP and VisualDSP++, 7.51-54
fundamental mathematical operation, 7.3
grounding, 10.29-30
hardware, 7.1-55
high density, localized decoupling,
10.53-55
interfacing, 8.1-25
Index-8
interfacing I/O ports, analog front ends,
and CODECS, 19-22
internal phase-locked loops, grounding,
10.29-30
kernel, 7.1-2
optimization, 7.6
output rise and fall times, diagram, 10.58
parallel interfacing, 8.1-12
block diagram, 8.8
reading data from memory-mapped
peripheral ADCs, 8.1-7
to external ADC, block diagram, 8.2
writing data to memory-mapped DACs,
8.7-12
practical example, 1.5-8
requirements, 7.3-6
circular buffering, 7.5
dual operand fetch, 7.4-5
extended precision, 7.4
fast arithmetic, 7.4
summary, 7.6
zero overhead looping, 7.5-6
sampled data, block diagram, 2.1
serial interfacing, 8.12-19
serial ADC, 8.14-19
serial DAC, 8.17-19
system interface, 8.23-24
using VLSI, 1.4
voiceband/audio applications, 9.36-40
DSP Designer's Reference (DSP Solutions),
7.55, 8.25
DSP Navigators: Interactive Tutorials about
Analog Devices' DSP
Architectures, 7.55
DSP Navigators: Interactive Tutorials about
Analog Device's DSP
Architectures, 8.25
Dual operand fetch:
DSP application, 7.4-5
DSP requirement, 7.4-5
Duracell MN1500 "AA" alkaline battery,
discharge characteristics, 10.2
Dynamic performance analysis, ideal N-bit
ADC, 2.17
Dynamic range compression, 1.4
E
Echo, voiceband telephone connection, 9.2
Eckbauer, F., 3.33
EDN's Designer's Guide to Electromagnetic
Compatibility, 10.65
Edson, J.O., 2.42, 3.34
Effective aperture delay time, 2.34
diagram, 2.35
Effective number of bits see ENOB
Eichhoff Electronics, Inc., 10.57
Electrolytic capacitor:
INDEX
characteristics, 10.40-41
impedance versus frequency, diagram,
10.44
Elliptic filter, 6.24, 6.27-28
characteristics, 2.5
Embedded speech processing, 9.36-37
EMC Design Workshop Notes, 10.56
End point, 2.11-12
Engelhardt, E., 3.33
ENOB, 2.21, 2.24-25, 3.10
ADC, effective resolution, 3.10
definition, 3.10
Equiripple FIR filter design, program
inputs, 6.19-20
Equivalent input referred noise, thermal
noise, 2.20-21
Error voltage, 10.17
Euler's equation, 5.3
Extended precision:
DSP application, 7.4
DSP requirement, 7.4
floating point arithmetic, 7.25
External clock jitter, 10.25
External port, versus link port, in
communications, 7.35
EZ-ICE, in-circuit simulator, 7.48
EZ-KIT Lite, 7.45-46, 7.53
EZ-LAB, evaluation board, 7.45
F
Fair-Rite, 10.44-46
Fair10.Rite Linear Ferrites Catalog, 10.56
Fairchild 74VCX164245, 16-bit low voltage
dual logic
translators/transceivers, 10.10-13
Faraday shield, 10.22
Fast arithmetic:
DSP application, arithmetic logic unit, 7.4
DSP requirement, 7.4
Fast Fourier transform see FFT
Fast logic, and analog circuits, 10.59
FDMA:
frequency division multiple access, 1.2
telephone system, 1.2
FDMA communications link, NPR testing,
2.30
Ferguson, P. Jr., 3.34
Ferguson, P.F. Jr., 3.33
Ferrite, 10.44
beads, 10.16, 10.20, 10.29, 10.30
impedance, 10.45-46
suitable for high-frequency filters, 10.45
FFT, 2.1, 5.1-25, 5.9-17
8-point DIF, 5.15
8-point DIT algorithm, 5.13
and DFT, 5.10
DIF, 5.13-14
FIR filter, 6.15
first Nyquist zone aliases, 2.4
floating point DSP, 5.18
frame-based systems, 5.18
hardware implementation and
benchmarks, 5.17-18
noise floor determination, 5.19
processing gain, 2.18-19, 5.20
processor, 5.17
radix-2, 5.16
radix-2 complex, hardware benchmark
comparisons, 5.18
radix-4, DIT butterfly, 5.16-17
real-time
considerations, 5.20
DSP requirements, 5.18-20
processing, 5.19
sinewave
integral number of cycles, 5.21
nonintegral number of cycles, 5.22
spectral leakage and windowing, 5.21-24
twiddle factors, 5.9-10
versus DFT, 5.11
Film capacitor, 10.40-42
Filter:
analog, for oversampling, 4.10
analog versus digital, 6.2-4
frequency response comparison, 1.7
antialiasing, specifications, 2.4-5
band, 10.46
bandpass, design, 6.23-24
bandstop, design, 6.23-24
baseband antialiasing, 2.4-6
Butterworth, characteristics, 2.5
capacitor, types, 10.40
Chebyshev, 1.6
corner frequency, 2.4
design, and FFT, 5.1
digital, diagram, 1.6
digital FIR, 1.6
elliptic, characteristics, 2.5
FIR versus IIR, comparisons, 6.30
high frequency, 10.51
high-frequency, ferrites, 10.45
highpass, design, 6.23-24
impulse convolving, 6.23
lowpass, analog versus digital, in
sampled data system, 1.5-6
power supply, 10.39-57
rolloff, sharpness, 6.9
switching supply, summary, 10.51
transfer function, 3.13
Filtering, 1.3, 1.4
FilterWizard, compiler, 9.41
Finite amplitude resolution due to
quantization, 2.2
Finite impulse response see FIR
Index-9
INDEX
FIR filter, 6.5-30
4-tap, output calculation, via circular
buffer, 6.11
arbitrary frequency response, design,
6.17-18
CAD design programs, 6.18-19
characteristics, 6.14
circular buffering, 6.10-13
in circular buffering, 7.5
circular buffering, fixed boundary RAM,
6.10
coefficients, 6.15
compared with IIR filter, 6.30
computational efficiency increase, 6.32
design
Fourier series method, with
windowing, 6.17
frequency sampling method, 6.17-18
fundamental concepts, 6.13
Parks-McClellan program, 6.18-22
programming ease in ADSP-21xx, 7.7-8
windowed-sinc method, 6.16
designing, 6.13-24
implementation, circular buffering,
6.10-13
impulse response, and coefficients, 6.14
Momentum Data Systems design
frequency response, 6.21
impulse response, 6.22
step response, 6.21
N-tap, general form, 6.9-10
output decimation, 6.32
program outputs, 6.20
pseudocode, 6.12, 7.7-8
simplified diagram, 6.10
transfer function
frequency domain, 6.15
time domain, 6.15
versus IIR filter, 6.30
First-order sigma-delta ADC, 3.12
Fisher, J., 3.33
Fixed-point:
16-bit, fractional format, 7.24
DSP arithmetic, 7.23-26
versus floating point, DSP arithmetic,
7.23-26
Flash ADC:
10-bit, diagram, 3.23
limited to 8-bits, 3.22
parallel ADCs, 3.21
diagram, 3.22
problems, 3.21
Flash converter, 3.21-23
Floating-point:
DSP arithmetic, 7.23-26
versus fixed point arithmetic, 7.26
Folding converter, 3.29
Index-10
Fourier series, 5.1
FIR filter design, windowing, 6.17
Fourier transform, 5.1, 6.15
and time domain signal, 5.2
Fourier, Jean Baptiste Joseph, 5.1
FPBW, full power bandwidth, 2.26
Framing mode, 8.14
Freeman, D.K., 9.43
Frequency division multiple access, FDMA,
1.2
Frequency domain, versus time domain,
6.13, 6.15
Frequency sampling method, FIR filter
design, 6.17-18
Frequency synthesis, using PLLs and
oscillators, diagram, 4.12
Fu, Dennis, 9.43
Full memory mode, 8.23
Full power bandwidth, FPBW, 2.26
Full-duplex hands-free car kit, diagram,
9.36
Fully decoded DAC:
5-bit, diagram, 4.4
diagram, 4.2
G
Gain error, 2.10
Galand, C., 9.43
Ganesan, A., 3.33, 3.34
Gardner, F.M., 4.17
Gaussian noise, 2.20-21, 2.30
Gaussian-filtered minimum-shift keying,
9.26
Geerling, Greg, 7.1
General DSP Training and Workshops,
7.55, 8.25
General purpose aluminum electrolytic
capacitor, 10.40-41
Gerber files, 10.33
Ghausi, M.S., 2.44
Gibbs effect, 5.1
Glitch:
code-dependent, effects, 2.38
DAC performance, 2.35-41
definition, 2.36
energy, 2.37
and harmonic distortion and SFDR,
prediction, 2.39
impulse area, 2.37
Global System for Mobile Communication
see GSM
Glue logic, 8.1
Gold, B., 5.25, 6.40
Gold, Bernard, 2.44
Gosser, Roy, 2.42, 3.34
Graham, M., 10.14
Graham, Martin, 10.34, 10.65
INDEX
Grame, Jerald, 10.34
Gray bit, 3.30
Gray code binary coding, data converters,
2.10
Gray, A.H., Jr., 6.40
Gray, G.A., 2.42
Ground loop, 10.20, 10.35
Ground plane, 10.30
decoupling high frequency current, 10.15
impedance, 10.20
low-impedance return path, 10.15-16
printed circuit board, 10.17-18
Ground screen, 10.19
Grounded-input histogram, 2.20-21
Grounding:
DSP, internal phase-locked loops,
diagram, 10.30
mixed signal devices, in multicard
system, 10.27-29
mixed signal systems, 10.15-34
mixed-signal ICs, 10.15
multiple ground pins, 10.32
philosophy, summary, 10.31
points, diagram, 10.24
separate analog and digital grounds,
10.19-20
single-card versus multicard, concepts,
10.26-27
Groupe Speciale Mobile see GSM
GSM, block diagram, 9.18-19
GSM handset:
components, 9.21-26
using SoftFone baseband processor, and
Othello radio, 9.21-26
H
Hageman, Steve, 10.56
Half-flash ADC, 3.23
Hamming window function, 5.22-23
Hanning window function, 5.22-23
Hard limiter, 2.29
Hardware design, techniques, 10.1-65
Harmonic distortion, 2.23-24
DAC, 2.38
definition, 2.23
products, location, diagram, 2.23
Harmonic sampling, 2.6-7
Harmonic undersampling, 2.7
Harrington, M.B., 9.44
Harris, Fredrick J., 2.43, 5.25, 6.40
Harris, Steven, 3.34
Harvard architecture:
in DSP, 7.5
in microprocessor, 7.6-7
Hauser, Max W., 3.34
Haykin, S., 6.40
HDTV, high definition television, 1.2
Heise, B., 3.33
Hellwig, K., 9.43
Henning, H.H., 2.42, 3.34
Higgins, Richard J., 1.9, 2.44, 5.25, 6.40,
7.55, 8.25
High definition television, HDTV, 1.2
High density DSP:
localized decoupling
diagram, 10.53
using BGA packages, 10.54
High Speed Design Techniques (1996), 3.35
High speed logic, 10.58-65
High-level language support, architectural
features, 7.42
High-speed interfacing, 8.22-23
Highpass filter:
design, 6.23-24
using lowpass FIR, 6.23
Hilton, Howard E., 2.44
Hodges, David A., 2.43
Hofmann, R., 9.43
Honig, Michael L., 6.40
Horvath, Johannes, 10.1
Host memory mode, 8.23
HP Journal, 2.43-44
HP Product Note, 2.43
I
IC, low-voltage mixed signal, 10.2
IEEE Trial-Use Standard for Digitizing
Waveform Recorders, 2.44
IF sampling, 2.6-7
IIR biquad filter:
basic, 6.25
simplified notations, 6.27
IIR elliptic filter, 6.30
IIR filter, 6.24-27
analog counterparts, 6.24
CAD design, using Fletcher-Powell
algorithm, 6.28
characteristics, 6.25
compared with FIR filter, 6.30
design techniques, 6.27-29, 6.29
direct form implementation, 6.26
feedback, 6.5
IIR filter, 6.24-27
implementation, 6.25
throughput considerations, 6.29
versus FIR filter, 6.30
IMD:
measurement, 2.28
second- and third-order, diagram, 2.29
two tone intermodulation distortion,
2.28-30
Impulse invariant transformation, 6.28
IMT-2000 protocol, 7.36
In-circuit simulator, 7.48
Index-11
INDEX
Indirect field-oriented control, 9.32
Induction motor:
control, 9.32-35
block diagram, 9.33
Infinite impulse response see IIR
Information, in signal, 1.2
Input filtering, 1.5
Input noise rejection, 1.5
Input/output memory select, ADC, 8.2
Instruction register, data address
generator, 7.14
Integral linearity error, measurement,
2.11-12
Integral sample-and-hold, 2.19
Interface:
2.5V/3.3V, diagram, 10.13
3.3V/2.5V, 10.10-13
diagram, 10.11-12
Interfacing, high-speed, 8.22-23
Interference, components, 10.39
Intermodulation distortion, DAC, 2.38
Internal aperture jitter, 10.25
Internal result bus, 7.10-13
Interpolation:
frequency domain effects, 6.33-34
implementation example, 6.33-34
multirate filter, 6.32-33
Interrupt request, DAC, 8.9
Intranet, 9.7
I.Q. convention, 7.24
IRQ, 8.1
J
Jantzi, S.A., 3.33
JEDEC, 10.14
specification, for packaging, 7.20
standards bureau, 10.5
Jitter:
aperture, 10.25
effects, 2.33
external clock, 10.25
internal apeture, 10.25
sampling clock, 10.25
Johnson noise, 2.8
Johnson, H., 10.14
Johnson, Howard W., 10.34, 10.65
Joint Electron Device Engineering Council
see JEDEC
Jung, Walt, 10.39, 10.56
K
Kelvin divider, 4.1
disadvantages, 4.1-2
ladder network, 4.3
Kelvin-Varley divider, string DAC, 4.3
Kerr, Richard J., 4.17
Index-12
Kester, Walt, 1.1, 2.1, 2.43, 3.1, 3.35, 4.1,
5.1, 6.1, 7.1, 8.1, 9.1,
10.1, 10.15, 10.34-35, 10.39
Kettle, P., 9.44
King, Dan, 7.1, 8.1
Koch, R., 3.33
L
Lagrange, Joseph Louis, 5.1
Laker, K.R., 2.44
Lane, Chuck, 2.42, 3.34
Laplace transform, 6.15
conversion to z-transform, 6.28
Laplace, Pierre Simon de, 5.1
Latency, 3.25
Lattice filter, 6.5
Leaded ferrite bead, 10.44
Least significant bit:
definition, 2.7
quantization, chart, 2.8
size, 2.8
Least-mean-square algorithm, for filter
coefficients, 6.36
Lee, Hae-Seung, 2.43
Lee, Wai Laing, 3.33
Lee, W.L., 3.33
Levine, Noam, 7.1
Linear Design Seminar (1995), 2.42, 3.33,
3.34
Linear integrity error, 2.11
Linear predictive coding see LPC
Linear settling time, DAC settling time,
2.35
Linearity errors, 2.10
Link descriptor file see LDF
Link port, versus external port, in
communications, 7.35
Link port multiprocessing, 7.34
LMS, least-mean-square, 6.36
Logarithmic pulse code modulation, 9.19
Logic:
CMOS IC output driver, configuration,
10.3-4
high speed, 10.58-65
Logic translating transceiver:
diagram, 10.11
voltage compliance, 10.11
Long term prediction see LTP
Low voltage interface, 10.1-14
Low Voltage Logic Alliance, standards, 10.5
Low voltage logic level, standards, 10.4-5
Low voltage mixed-signal IC, 10.2
Low-pass filter, LPF, 3.11
Lowpass filter, design of other filters, 6.23
LPC, 6.36-37
in all-pole lattice filter, 6.38-39
model of speech production, 6.37
INDEX
speech companding system, 6.38
speech processing system, 9.19
speech system, digital filters, 6.38
LPF:
analog, in sigma-delta ADC, 3.13
low-pass filter, 3.11
LQFP, new package designation, 7.20
Lucey, D.J., 9.44
LVTTL, logic level, 10.4
Lyne, Niall, 9.44
M
McClellan, J.H., 6.40
MagAmp, 3.29
3-bit folding ADC
block diagram, 3.31
functional equivalent circuit, 3.30
input and residue waveforms, 3.31
Magnitude amplifier, 3.29
Magnitude-amplifier architecture,
MagAmp, 3.24
Mahoney, Matthew, 2.44
Main lobe spreading, 5.22
Manolakis, Dimitris G., 5.25, 6.40
Markel, J.D., 6.40
Marsh, Dick, 10.56
Matched z-transform, 6.28
MathCad 4.0 software package, 2.44
Mathematical calculation, by computer, 7.1
Matsuya, Y., 3.33, 3.34
Maximally flat filter, 6.27
Mayo, J.S., 2.42, 3.34
Meehan, Pat, 2.43
Memory address bus, ADC, 8.2
Memory read, ADC, 8.2
Memory select:
ADC, 8.2
DAC, 8.9
Messerschmitt, David G., 6.40
Microcontroller, 7.1-3
characteristics, 7.2
MicroConverter, precision analog circuitry,
1.3
Microprocessor, 7.1-3
architectures, comparison, 7.6
characteristics, 7.2
Million operations per second see MOPS
Millions of instructions per second see MIPS
miniBGA package, 7.20-21
Minimum 4-term Blackman-Harris window
function, 5.22-23
MIPS, 7.2, 7.18
Missing codes:
in ADC, 2.12-13
defining, 2.14
Mitola, Joe, 9.43
Mixed Signal Design Seminar (1991), 3.33
Mixed signal processing, 1.4
Mixed-signal device:
grounding, 10.26-27
diagram, 10.27
Mixed-signal IC:
grounding, multiple printed circuit
boards, 10.28-29
high digital currents, grounding, 10.28-29
low digital current, grounding and
decoupling, 10.21-22
low digital currents, grounding, 10.27-28
Mobile telephone service, superheterodyne
receiver, diagram, 9.27
Mode status register, data address
generator, 7.13
Modem:
full-duplex, 9.1
half-duplex, 9.1
high performance, telephone service,
9.1-7
high speed, 1.2
RAS, 9.7-10
standards, chart, 9.3
V.90 analog
block diagram, 9.4
details, 9.6
V.90 analog versus V.34, diagram, 9.7
Momentum Data Systems, QED1000
program, for FIR filter design, 6.18
Momentum Data Systems, Inc., 6.40
Montrose, Mark, 10.34, 10.65
MOPS, 7.18
Moreland, Carl, 2.42, 3.34
Morley, Nick, 9.43
Morris, Jesse, 7.1
Morrison, Ralph, 10.34
Motchenbacher, C.D., 10.34
Motherboard, grounding, 10.18-19
Motor control, 9.32-35
fully integrated, DashDSP, 9.35
Motor Control Products, Application Notes,
and Tools, 9.44
Moving average filter, 6.5-10
calculating output, 6.7
diagram, 6.6
frequency response, 6.8
noise, 6.6-8
step function response, 6.6-7
MSP, 1.4
Multi-channel high frequency
communication system, NPR, 2.31
Multi-channel VOIP server, 9.10-11
Multi-tone SFDR, measurement, 2.29
Multicard mixed-signal system, grounding,
10.18-19
Multilayer ceramic chip caps, 10.42
Multiplier-accumulator unit, features, 7.12
Index-13
INDEX
Multiplying DACs, 4.3
Multipoint ground, diagram, 10.19
Multiprocessing, using SHARCs, 7.34-36
Multirate filter, 6.31-35
decimation, 6.31
interpolation, 6.31
Murden, Frank, 2.42, 3.34
Murray, Aengus, 9.44
N
Narrowband, 9.27
Network:
LAN, 1.2
local area, 1.2
Nicholas, Henry T. III, 4.17
NMOS FET, bus switches, interfacing,
10.6-8
Noise:
and grounding, 10.22
power supply, 10.39-57
Noise power ratio see NPR
Noise shaping, using analog filter, 3.13
Non-ideal ADC, 3-bit, transfer function,
diagram, 2.13
Non-ideal DAC, 3-bit, transfer function,
diagram, 2.13
Normal framing mode, 8.14
NPR, 2.30-32
measurement, diagram, 2.31
NPR, 2.30-32
theoretical, for various ADCs, 2.32
Numerically controlled oscillator, 4.13
block diagram, 4.14
Nyquist band, 3.10
Nyquist bandwidth, 2.15, 2.17
aliasing, 2.4
definition, 2.4
Nyquist criteria, 2.2
Nyquist frequency, 2.41, 9.5
Nyquist zone, 2.4, 2.26-27
baseband sampling, 2.6
spurious frequency component, 2.4
O
Offset binary coding, data converters, 2.10
Offset error, 2.10
1's complement binary coding, data
converters, 2.10
Oppenheim, A.V., 5.25, 6.40
Optocoupler, 10.35
Optoisolator, 10.35
OS-CON Aluminum Electrolytic Capacitor
93/94 Technical Book, 10.56
OS-CON electrolytic capacitor, 10.40-41,
10.43
Othello:
radio, chipset, 9.21
Index-14
radio receiver
advantages, 9.24
block diagram, 9.24-25
compactness, advantages, 9.26
superhomodyne architecture, 9.23-24
Ott, Henry, 10.34, 10.56
Output enable/read, ADC, 8.2
Output ripple, reduction, 10.50
Oversampling, 3.10
and baseband antialiasing filter, 2.5
definition, 2.16-17
ratio, 3.11
P
Parallel ADC, diagram, 3.22
Parallel peripheral device:
read interface, key requirements, 8.3
write interface, key requirements, 8.9
Parasitic capacitance, 10.43
Parasitic inductance, 10.22
Parks-McClellan program:
equiripple FIR filter design, program
inputs, 6.19-20
FIR filter, 6.28
design, 6.18-22
Parks, T.W., 6.40
Parzefall, F., 3.33
Passband ripple, 6.2
in filter, 1.6
Pattavina, Jeffrey S., 10.34
Peak glitch area, 2.37
Peak spurious spectral content, ADC, 2.26
Pentium-Series, 7.2
Pericom Semiconductor Corporation, 10.14
Permanent magnet synchronous machine,
9.33
PGA, programmable gain amplifier, 1.5
Phase accumulator, 4.14
Phase jitter, 2.33
Phase-locked loop, PLL, 4.12
Phase-Locked Loop Design Fundamentals,
4.17
PIC, 7.2
Picocell, 9.30
Pin socket, 10.16
Pipeline delay, 3.25
Pipelined ADC, 3.23-27
12-bit CMOS ADC, block diagram, 3.26
error correction, 3.25-26
Pipelined subranging ADC, 12-bit, digital
error correction, block diagram,
3.25
Plain Old Telephone Service see POTS
PMOS transistor current switch, diagram,
4.7
Point of presence see POP
Polyester, capacitor, 10.40-41
INDEX
POP, 9.8
POTS:
block diagram, 9.2
hybrid curcuit, 9.1
Power inverter, 9.32
Power plane, 10.15-17
Power supply:
filtering, 10.39-57
localized high frequency
filter, 10.51-53
construction guidelines, 10.52-53
noise reduction, 10.39-57
pins, decoupling, 10.16
separate for analog and digital circuits,
10.23
PowerPC, 7.2
Practical Analog Design Techniques (1995),
2.42, 3.34
Practical Design Techniques for Power and
Thermal Management, 10.57
Practical Design Techniques for Sensor
Signal Conditioning, 1.9
Printed circuit board:
double-sided versus multilayer, 10.17-18
ground plane, 10.17
impedance, calculation, 10.59
mixed-signal system, layout guidelines,
10.31-33
"motherboard", grounding, 10.18-19
trace termination, 10.58
Proakis, John G., 5.25, 6.40
Processing gain:
definition, 2.16-17
FFT, 2.18-19
Processor interrupt request line, ADC, 8.1-2
Program memory address bus, 7.10
Program memory data bus, 7.10
Program memory select, ADC, 8.2
Program sequencer:
data address generator, 7.14
features, 7.14
Programmable gain amplifier, PGA, 1.5
Prom splitter, 7.45
Pseudocode, FIR filter, using DSP with
circular buffering, 7.8
Pulse code modulation, 9.3
Q
QEDesign, 9.41
filter package, 3.20
QS3384 Data Sheet, 10.14
QS3384 QuickSwitch, 10.7-8
transient response, diagram, 10.8
Quadrature amplitude modulation, 9.26
diagram, 9.5
POTS, 9.3
Quantization:
effects on ideal ADC, 2.15
size of LSB, chart, 2.8
Quantization error, 2.9-10
signal, 2.15
Quantization noise:
in digital system with ADC, 2.31
for ideal N-bit ADC, diagram, 2.16
shaping, 3.10
in sigma-delta ADC, 3.14
spectrum, diagram, 2.16
Quantization uncertainty, 2.9-10
QuickSwitch, in bidirectional interface,
10.6-8
R
Rabiner, Lawrence, 2.44
Rabiner, L.R., 5.25, 6.40, 6.41
RAMDAC, 1.3
Ramierez, Robert W., 2.43
Ramirez, R.W., 5.25
RAS:
equipment, 9.10
modem, 9.7-10
as Internet gateway, diagram, 9.8
on-switch based, 9.9
RAS/VOIP servers, 9.10
Receive data register, 8.14
Receive frame sync, 8.14
Receive shift register, 8.14
Reconstruction, definition, 2.14
Reconstruction filter, 6.2
Recovery time, DAC settling time, 2.35
Recursive filter, 6.24
Recursive-least-squares algorithm, for filter
coefficients, 6.36
Reference frame theory, 9.33
Regular pulse excitation see RPE
Reidy, John, 2.43
Remez exchange algorithm, 6.18-19
Remote access server see RAS
Remote network access, 9.7
Rempfer, William C., 10.34
Ripple, passband, in filter, 1.6
Ripple ADC, 3.27-32
diagram, 3.27
RISC, and DSP, 7.3
RLS, recursive-least-squares, 6.36
Roche, P.J., 9.44
Rolloff:
sharpness, 6.9
sidelobe, 5.22
Root-sum-square, RSS, 2.24
Rorabaugh, C. Britton, 5.25, 6.40, 7.55, 8.25
Rosso, M., 9.43
Ruscak, Steve, 2.42
Index-15
INDEX
S
Sample rate converter, using interpolator
and decimator, 6.35
Sample-and-hold see SHA
Sampled data system, definition, 2.14
Sampling:
above first Nyquist zone, 2.7
bandpass, 2.6-7
harmonic, 2.6-7
IF, 2.6-7
rate versus bandwidth, 2.7
Sampling clock, 10.24-26
ground planes, diagram, 10.26
jitter, effect on SNR, 10.25
Sampling rate, increased, 1.5
Samueli, Henry, 4.17
SAR, 3.2
SAR ADC, 3.3
external high frequency clock, 3.6
fundamental timing, 3.5
resolutions, table, 3.5
switched capacitor, 3.8
with switched capacitor DAC, 3.3
typical timing, 3.6
Sauerwald, Mark, 10.34
Scannell, J.R., 9.44
Schafer, R.W., 5.25, 6.40, 6.41
Schmid, Hermann, 2.42, 3.34
Schmitt trigger, 10.38
Schottky diode, 10.7, 10.20, 10.29, 10.30
Segmentation, in thermometer DAC, 4.5
Segmented DAC, 4.2-3
10-bit, diagram, 4.5
Segmented voltage DAC, segmented
voltage, 4.3
Sensor:
as analog device, 1.4
in industrial data acquisition and control
systems, 1.3
uses, 1.1-2
Serial ADC, 3.27-32
diagram, 3.27
to DSP interface, 8.14-17
Serial clock, 8.14
Serial DAC, to DSP interface, 8.17-19
Serial ports, in ADSP-21xx, 7.16
Serial-Gray, ADC architecture, 3.29
Settling time:
DAC performance, 2.35-41
diagram, 2.36
periods, 2.35
SFDR, 2.17, 2.21, 2.26-28, 2.35-41
ADC, definition, 2.26
DAC
distortion, 2.38
performance, 2.35-41
test setup, 2.40
Index-16
diagram, 2.27
sampling clock to input frequency ratio,
for ideal 12-bit ADC, 2.18
SHA, 2.19-20
and aperture jitter, 2.33
SHARC:
32-bit, key features, 7.32
architecture, 7.27
coding, 7.33
decoupling, 10.53
DSP
evaluation device, 7.48-49
floating and fixed point arithmetic,
7.25
DSP benchmarks, 7.34
family roadmap, 7.33
FFT butterfly processing, 7.26
floating point DSP, 7.26-30
floating-point DSP, 7.26-30
key features, 7.27
multi-function instruction, 7.33
multiprocessing, 7.34-36
multiprocessor communication, examples,
7.35
program sequencer, 7.26
VisualDSP++ software, 7.52
SHARC DSP:
32-bit floating point, 9.38-39
bi-directional transmission, source
termination, diagram, 10.63
Sheingold, Dan, 2.44
Sheingold, Daniel H., 1.9, 3.34
Shifter unit, features, 7.12
Shunt resistance, 10.43
Sidelobe, 5.22
Sigma-delta ADC:
bandpass filters, 3.18-19
characteristics, table, 3.10
first-order, diagram, 3.12
fixed internal digital filter, 3.19
and missing codes, 2.14
multi-bit
diagram, 3.18
flash ADC with DAC, 3.17-18
noise shaping, 3.12-13
order versus oversampling, 3.14-15
oversampling, 3.20
converter, 2.6
programmable digital filter, 9.40-42
quantization noise shaping, 3.14
second-order, block diagram, 3.15
settling time, and digital filter, 3.16
SNR, 3.20
summary, 3.21
Sigma-delta audio DAC, 3.17
Sigma-delta DAC, 4.11-12
architecture, diagrams, 4.11
INDEX
high resolution, oversampling, 4.11
Sign-magnitude bipolar converter, 2.10
Signal:
amplitude, 1.2
analog, 1.1
bandwidth
aliasing, 2.2
Nyquist criteria, 2.2
characteristics, table, 1.1
conditioning, 1.1
analog signal processor, 1.4
sensors, 1.1-2
successive approximation ADCs, 3.2
continuous aperiodic, 5.1
continuous periodic, 5.1
sinusoidal waves, 5.1
definition, 1.1, 1.2
digital, 1.1-2
frequency, 1.2
processing
frequency compression, 1.2
information extraction, 1.2
methods, 1.4-5
mixed, 1.4
real-time, comparison chart, 1.8
reasons, 1.3
real-world
generation, 1.3
origins, 1.1-2
processing, 1.2-3
units of measurement, 1.1-2
recovery, 1.3
sampled aperiodic, 5.1
sampled periodic, 5.1
spectral content, 1.2
timing, 1.2
units of measurement, table, 1.1
Signal-to-noise ratio see SNR
Signal-to-noise-and-distortion ratio, SINAD,
2.21, 2.24-25
Signed fractional format, DSP arithmetic,
7.23
Signed integer format, DSP arithmetic, 7.23
Silence descriptor, 9.20
SINAD, 2.21, 2.24-25
conversion to ENOB, equation, 2.25
Singer, Larry, 2.42
Single precision, floating point arithmetic,
standard, 7.25
single-instruction, multiple data see SIMD
Single-instruction, single-data see SISD
Sinusoidal wave, continuous periodic signal,
5.1
SISD, older DSP architecture, 7.31
Sluyter, R.J., 9.43
Small signal bandwidth, 2.25
Smith, Steven W., 5.25, 6.40, 7.55, 8.25
Snelgrove, M., 3.33
SNR, 2.15, 2.21, 2.24-25
decrease with input frequency in ADC,
2.33
definition, 2.15
degradation, from external clock jitter,
10.25
due to aperture and sampling clock jitter,
diagram, 2.34
SNR-without-harmonics, calculation, 2.25
Sodini, C.G., 3.33
SoftCell:
chipset
block diagram, 9.30
for digital phone receivers, 9.29
SoftFone, baseband processor, chipset, 9.21
Software development environment, 7.51
Software simulator, for debugging, 7.45
Southcott, C.B., 9.43
Spectral inversion, in filter design, 6.23
Spectral leakage:
FFT, 5.21
windowing, 5.23
Spectral reversal, in filter design, 6.23
Speech:
compression, adaptive filter use, 6.36
encoder/decoder, 9.18-19
processing, standards, 9.18-19
synthesis, adaptive filter use, 6.36
synthesized, 1.3
SPORT:
in ADSP-21xx, 7.16
transfer rate, 8.20
Spurious free dynamic range see SFDR
SSBW, small signal bandwidth, 2.25
Stacked-film capacitor, 10.40, 10.42
Star ground, 10.16, 10.19, 10.26
Static branch prediction logic, 7.39
Static superscalar, 7.39
Stearns, S.D., 6.40
Stopband ripple, 6.2
String DAC:
diagram, 4.2
disadvantages, 4.1-2
Kelvin divider, 4.1
Subranging ADC, 3.23-27
8-bit, diagram, 3.24
12-bit, digitally corrected, block diagram,
3.25
digital correction, 3.24
Successive approximation ADC, 3.1-9
basic, 3.3
Successive approximation register see SAR
SUMMIT-ICE, PCI emulator, 7.50
Super Harvard Architecture, for 32-bit DSP,
7.27-28
Superheterodyne architecture, 9.23-24
Index-17
INDEX
Superhomodyne architecture:
diagram, 9.24-25
operation, 9.25-26
Supply voltage reduction, 10.1
Surface mount ferrite bead, 10.44
Swanson, E.J., 3.33
Switching capacitor, characteristics,
10.40-41
Switching regulator, 10.46-51
filtering, experiment, 10.46-50
noise, reduction tools, 10.40
Switching supply, filter, summary, 10.51
Switching time, DAC settling time, 2.35
System Applications Guide (1993), 2.43,
3.33, 3.35
Systolix FilterExpress, 9.41
Systolix PulseDSP, 9.41
filter core, 6.2
processor, 3.19-20
T
T-Carrier system, 1.2
Tantalum electrolytic capacitor, 10.40-41
Tantalum Electrolytic Capacitor SPICE
Models, 10.57
Tantalum Electrolytic and Ceramic
Capacitor Families, 10.56
Tant, M.J., 2.42
TDMA, 1.2
Telephone, FDMA, 1.2
Tesla, Nikola, 9.32
THD+N, 2.23-24
definition, 2.24
THD, 2.21, 2.23-24
definition, 2.24
Thermal noise, equivalent input referred
noise, 2.20-21
Thermometer code, 3.21
Thermometer DAC:
5-bit, diagram, 4.4
diagram, 4.2
disadvantage, 4.5
segmentation, 4.5
Thevenin impedance, 10.61
Thompson filter, 6.27-28
TigerSHARC:
ADSP-TS001 static superscalar DSP,
7.36-44
architecture, 7.38
branch target buffer, 7.39, 7.41
features, 7.39
flexibility, 7.36
key features, 7.37
multiprocessor key elements, 7.36-37
static branch prediction logic, 7.39
static superscalar, 7.39
development tools, 7.53
Index-18
DSP, 9.30-31
key features, 7.40
multiprocessing implementation, sample
configuration, 7.42-43
peak computation rates, 7.41
roadmap, 7.43
Time division multiple access see TDMA
Time domain, versus frequency domain,
6.13, 6.15
Time sampling, analog signal, 2.2-7
Total harmonic distortion see THD
Total harmonic distortion plus noise see
THD+N
TQFP package designation, new
designations, 7.20
Transmission line:
controlled impedance microstrip,
termination, 10.61
termination at both ends, 10.64
Transmit data register, 8.14
Transmit shift register, 8.14
TREK-ICE, Ethernet emulator, 7.50
TTL, logic level, 10.4
Tukey, J.W., 2.43, 5.10, 5.25
Twiddle factors, 5.9-10
Two tone intermodulation distortion, IMD,
2.28-30
2's complement binary coding, data
converters, 2.10
Type 1 Chebyshev filter, 6.27-28
Type 2 Chebyshev filter, 6.27
Type 5MC Metallized Polycarbonate
Capacitor, 10.56
Type 5250 and 6000-101K chokes, 10.56
Type EXCEL leaded ferrite bead EMI filter,
and type EXC L leadless
ferrite ead, 10.56
Type HFQ Aluminum Electrolytic Capacitor
and type V Stacked Polyester
Film Capacitor, 10.56
U
Undersampling, 2.6-7
diagram, 2.6
harmonic sampling, 2.7
Nyquist zone, 2.6
Unipolar converter, 2.9-10
Unsigned fractional format, DSP
arithmetic, 7.23
Unsigned integer format, DSP arithmetic,
7.23
Using the ADSP-2100 Family, 7.55, 8.25
V
Vary, P., 9.43
VCX device, 10.9
74VCX164245 Data Sheet, 10.14
INDEX
Vector control, 9.32-33
Very Large Scale Integration, VLSI, 1.4
Very long instruction word, in SHARC
architecture, 7.37
Video raster scan display system, 1.3
Virtual-IF transmitter, in AD6523, 9.23
VisualDSP++, DSP development software,
7.51
VisualDSP:
DSP development software, 7.51
test drive, 7.53
Viterbi algorithm, 9.6
Viterbi decoder, 9.5
Voice activity detector, 9.20
Voice over the Internet protocol, 7.42
Voice-over-Internet-provider see VOIP
Voltage:
identification pin, 10.1
low, interfaces, 10.1-14
supply, reduction, 10.1
Voltage compliance, 10.5-6
definition, 10.6
internally created, 10.8-10
considerations, 10.9-10
Voltage tolerance, 10.5-6
definition, 10.5-6
internally created, 10.8-10
Voltage-controlled-oscillator, 4.12
Von Neumann architecture, 7.6
W
Waldhauer, F.D., 2.42, 3.29, 3.34
Waurin, Ken, 7.1
Weaver, Lindsay A., 4.17
Weeks, Pat, 2.43
Welland, D.R., 3.33
Wepman, Jeffery, 9.43
Wideband, 9.27
Widrow, B., 6.40
Window function, 5.22-23, 5.23, 6.16
characteristics, 5.24
frequency responses, 5.24
Windowed-sinc method:
FIR filter design, 6.16
responses, diagrams, 6.16
Witte, Robert A., 2.43
Wooley, Bruce, 3.33
Worst harmonic, 2.23-24
Write, DAC, 8.9
X
xDSL protocol, 7.36
Analog Devices Parts Index
AD260/AD261, 10.36-38
AD820, 3.8-9
AD974, 3.5
AD2S80A, 9.33-34
AD2S82A, 9.33-34
AD2S83A, 9.33-34
AD2S90A, 9.33-34
AD185x, 4.11
AD189x, 6.35
AD1819B, 1.5, 9.37-38
AD1836, 9.38-39
AD1852, 3.17
AD1853, 3.17, 4.12
AD1854, 3.17
AD1877, 3.16-17
AD1890, 6.35
AD1891, 6.35
AD1892, 6.35
AD1893, 6.35
AD1896, 6.35
AD5322, 8.17-19
AD5340, 8.10-12
AD6521, 9.21-22
AD6522, 9.21-22
AD6523, 9.22-23, 9.25
AD6524, 9.22-23, 9.25
AD6600, 9.30
AD6622, 9.30-31
AD6624, 9.30-31
AD6640, 2.29-30, 3.24-25
AD6644, 9.30-31
AD7472, 3.5
AD77xx, 3.17
AD773x, 1.5
AD977x, 4.6
AD983x, 4.16
AD984x, 1.5
AD985x, 4.6
AD7660, 3.5
AD7664, 3.5
AD7722, 10.15
AD7725, 3.19, 6.2, 9.40-42
AD7730, 10.15
AD7731, 10.15
AD7853/7853L, 8.15-18
AD7853L, 8.15-17
AD7854/AD7854L, 8.4-7
AD7856/67, 3.5
AD7858/59, 3.5, 3.8-9
AD7861, 9.33
AD7862, 9.33
AD7863, 9.33
Index-19
INDEX
AD7864, 9.33
AD7865, 9.33
AD7887/88, 3.5
AD7891, 3.5
AD7892, 10.15
AD9042, 2.27, 2.31-32
AD9201, 8.22-23
AD9220, 2.24-25, 3.26
AD9221, 3.26
AD9223, 3.26
AD9288-100, 3.32
AD9410, 3.23
AD9761, 8.22-23
AD9772, 2.39, 4.6, 4.9-10, 9.30-31
AD9814, 1.5
AD9815, 1.5
AD9850, 4.15-16
AD73311, 8.1
AD73322, 1.5, 8.1, 8.19-22, 9.36
AD73422, 8.21-22
AD74222-80, 8.21-22
AD20msp430, 9.21-23
AD20msp910, 9.12, 9.15
AD20msp918, 9.15
ADMC200/ADMC201, 9.33
ADMC300, 9.34
ADMC326, 9.34
ADMC328, 9.34
ADMC331, 9.34
ADMC401, 9.34
ADMCF5xx, 9.35
ADMCF326, 9.34
ADMCF328, 9.34
ADP1148, 10.46
ADP3310, 10.47, 10.49-50
ADP33xx, 9.22
ADP34xx, 9.22
ADSP-21ESP202, 9.36-37
ADSP-21mod870, 9.8-10
ADSP-21mod970, 9.10
ADSP-21mod980, 9.10
ADSP-21modxxx, 9.9
ADSP-21msp5x, 7.21
ADSP-21xx, 1.6-8, 6.11-13, 6.28, 7.2, 7.4,
7.6-24, 8.1-3, 8.7-8, 8.12-
Index-20
14, 9.5, 9.8, 9.33, 10.9
ADSP-210x, 7.21
ADSP-216x, 7.21
ADSP-217x, 7.18, 9.34-35
ADSP-218x, 7.7, 7.11, 7.15, 7.18, 7.22, 7.49,
7.51-53, 8.20-22, 9.5,
9.8, 9.10-11, 9.22, 9.36-38
ADSP-218xL/M, 10.1
ADSP-219x, 7.7, 7.11, 7.19-21, 7.23, 7.51-53,
9.35
ADSP-2100, 7.9, 7.15
ADSP-2105, 7.16
ADSP-2106x, 7.26-30
ADSP-2111, 7.21
ADSP-2116x, 7.31-36
ADSP-2181/3, 7.21
ADSP-2183, 9.15
ADSP-2184/L, 7.21
ADSP-2185/L/M, 7.21
ADSP-2185L/86L, 8.21-22
ADSP-2186/L, 7.21
ADSP-2187L/M, 7.21
ADSP-2188M, 7.20, 7.21, 9.10-11
ADSP-2189M, 5.18, 6.1, 6.12, 6.20, 6.22,
6.28, 7.21, 7.47, 8.2-7, 8.912, 8.14-15, 8.23-24
ADSP-21000, 7.26-27
ADSP-21060, 7.29-30, 7.33
ADSP-21060L, 10.58
ADSP-21061, 7.29-30, 7.33
ADSP-21062, 7.29-30, 7.33
ADSP-21065, 7.29-30, 7.33
ADSP-21065L, 7.27, 7.34, 7.48, 8.22-23,
9.38-39
ADSP-21160, 5.18, 7.27, 7.31-32, 7.34, 7.36,
9.39-40, 10.29-31, 10.54-55
ADSP-21160M, 7.33, 7.47
ADSP-21161N, 7.33
ADSP-TS001 TigerSHARC, 5.18-19,
7.36-44, 9.30-31
ADuM1100A, 10.36-37
ADuM1100B, 10.36-37
EZ-ICE, 7.48
EZ-KIT Lite, 7.45-48
EZ-LAB, 7.45
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertisement