User manual | vd

'I
V...
»lv
< • >
1-.
•
J ^
•^-
'y.''^
II
*
i;>S;F^
lEs I S M I 3 E R
F^: e~* v ;i. sis- SE? c i
D :3.. s"" IE? C : t
vd
i . C" r's
. il
J..
E K e c; LI t i v e
II
I n t r o d u c t i o n t o S i g n a l F'-'roceBB i, n q t-. DSP
A . I n t r o d u c t i o n - v-^Krat i ' s i t , w h y f i e e d e d ?
Ei„ S i g n a l P r o c e s s i n g Pipp 1 i c a t i o n s
^
C. S i g n a l F'rocesBinq
TechyiiqueB
j^ *'
-- A n a l o g
—Digital
>•
f *
D- I m p a c t o f DSP
I
D . S i g n a l F ' r o c e B B ;i riq I C B >
^ -- Analog
' •
Digi t a 1
— M i •;•; e d A n a 1 o Q / D i Q i t a 1.
11
EJigvial P r o c e s i s i n e ! M a r k e t s
A„ EJy E n d p r o d u c t
1» Consumer' Audio
S . Corisufner V i d e o
'*
3 . Corrst-imer- O t t i e r '
4« A u t o m o t i v e
5.
Telecom
6» C o m p u t e r
7., H i l i t a r y
S .. I n s t r u m e r i t a t i o n
9 . 1 nc;!uB t r" i a I / P \ • c ^ c e s s Co ri t r"o 1
B. Regional! N a t u r e o f A c t i v i t ies/i'1ai
C , DE)P p e r i e t r a t i o r i &: F o r e c a s t
*. 1
IV.
(-*-:
if.t^
S u in in a s" y
DSP ICs
t1 i c i o p r o c e s B o r " B
J•
E< u i 1 d i n g E( 1 o c k: s
A. M a r k e t
'•*
B., F o r e c a ^ i t
C » t ' 1 a r " k e t B h a T - e £•: Compejt i t i - / e A n a l
D „ E" ille r g i n g I B B U e B
A r i a 1 (:;i g D S P 5i !.,t p f:) o i • t IC" s
SensorE
Amp 1 i f i er B
Swi tches/MuxeB ?>•. Sample/Holds
D a t a C; o rr v E: r t e r" B
A, Market
Ei» Forecast
C. F'roduct AnaiyBis ••- Data Converters
D„ Marketshare £•: Competitive Analysis
E. Emerging Issuers wi tli respect to DSP
*
VIII
DSP A S IC
( D i q i t a 1 l-. M x ;; ed )
kets
'^<\
ii
^ ; \ .
f
ri- •
*•*'
T^'
Ar
'f;
*•
A„
B»
C.
•^D»
'
-
Market
-fForecast
M a r k e t s ^ h a r e &: C o m p e t i t i v e
E m e v g i viq I s s n e s
V
Ai
^'
\
:
DSP T' e c: tT n c 1 o q y T ren d s
i, •
DeVe 1 o p men t Sup p c v t
i'lajor
arid
Einerginq
DSP
Supplii
-^'.
^
t-l
^
.
/
•
7
•-•ju\;-
;<^
r
- ^
*-.
t
L.
[ Vr
,
At
^^wrf
\ '
^1
•!
.
J
Semiconductor Industry Service
Digital Signal Processing
Dataquest
a company of
Tlie Dun & Kad^icct Coqwration
1290 Ridder Park Drive
San Jose, California 95131-2398
(408) 437-8000
Telex: 171973
Fax: (408) 437-0292
Sales/Service Offices:
UNITED KINGDOM
Dataquest UK Limited
13th Floor, Centrepoint
103 New Oxford Street
London WCIA IDD
England
01-379-6257
Telex: 266195
Fax: 01-240-3653
FRANCE
Dataquest SARL
Tour Gallieni 2
36, avenue Gallieni
93175 Bagnolet Cedex
France
(1)48 97 31 00
Telex: 233 263
Fax: (1)48 97 34 00
EASTTERN U.S.
Dataquest Boston
1740 Massachusetts Ave.
Boxborough, MA 01719
(617) 264-4373
Telex: 171973
Fax: (617) 263-0696
GERMANY
Dataquest GmbH
Rosenkavalierplatz 17
D-8000 Munich 81
West Germany
(089)91 10 64
Telex: 5218070
Fax: (089)91 21 89
JAPAN
Dataquest Japan, Ltd.
Taiyo Ginza Building/2nd Floor
7-14-16 Ginza, Chuo-ku
Tokyo 104 Japan
(03)546-3191
Telex: 32768
Fax: (03)546-3198
KOREA
Dataquest Korea
63-1 Chungjung-ro, 3Ka
Seodaemun-ku
Seoul, Korea
(02)392-7273-5
Telex: 27926
Fax: (02)745-3199
The content of this report represents our interpretation and analysis of information generally available to the public or released by responsible individuals in the subject companies, but is not guaranteed
as to accuracy or completeness. It does not contain material provided to us in confidence by our clients.
This information is not furnished in connection with a sale or offer to sell securities, or in connection with the solicitation of an offer to buy securities. This firm and its parent and/or their officers,
stockholders, or members of their families may, from time to time, have a long or short position in
the securities mentioned and may sell or buy such securities.
Printed in the United States of America. All rights reserved. No part of this publication may be
reproduced, stored in retrieval systems, or transmitted, in any form or by any means—mechanical,
electronic, photocopying, duplicating, microfilming, videotape, or otherwise—without the prior written
permission of the publisher.
© 1988 Dataquest Incorporated
Digital Signal Processing
Table of Contents
TABLE OF CONTENTS
How To Use the Table of Contents
Digital Signal Processing Table of Contents
INTRODUCTION
Introduction to Digital Signal Processing
DSP OVERVIEW
Executive Summary
DSP Executive Summary
Forecast Summary
DSP Forecast Summary
Market Dynamics
DSP Markets and Applications
Family Tree and Definitions
DSP Family Tree
DSP Definitions
DSP MICROPROCESSORS
Executive Summary
DSMPU—Executive Summary
Forecast
DSMPU—Forecast
Product Analysis
DSMPU—Product Comparison
Emerging Technology and Trends
DSMPU—Emerging Technology and Trends
SIS DSP
0002976
© 1989 Dataquest Incorporated February
Digital Signal Processing
Table of Contents
DSP BUILDING BLOCKS
Executive Summary
DSP Building Blocks—Executive Summary
Forecast
DSP Building Blocks—Forecast
Product Analysis
DSP Building Blocks—Product Comparison
Emerging Technology and Trends
DSP Building Blocks—Emerging Technology and Trends
ASICs AND SPECIAL FUNCTION DSPs
Executive Summary
ASICs/SFICs—Executive Summary
Forecast
ASICs/SFICs—Forecast
Competitive Analysis
ASICs/SFICs—Suppliers
DSP TECHNICAL OVERVIEW
DSP—Technical Overview
DISTRIBUTION
Potential Users
Directory of Potential Users
PACKAGING
Worldwide IC Packaging Update
Surface-Mount Technology Overview
© 1989 Dataquest Incorporated February
SIS DSP
0002976
How to Use the Table of Contents
The Digital Signal Processing (DSP) notebook is arranged in a series of
major sections defined by blue primary tabs with labels in all capital
letters. These major topic areas may be further divided by white subtabs
whose labels begin with initial capital letters.
In most cases, the
documents following either of these tabs will repeat in their running
headlines the wording of the primary subject tab and/or that particular
subtab.
To assist in locating a particular document, a table of contents has been
provided. This table lists the primary tabs, the subtabs, and the running
headline of each document contained in the notebook. The following is an
annotated example of the format of the DSP table of contents:
SAMPLE TABLE OF CONTENTS
DSP OVERVIEW*
Executive Summary**
DSP Executive Summary^
Forecast Summary
DSP Forecast Summary
Market Dynamics
DSP Market Strategies
DSP Market Analysis
Family Tree and Definitions
DSP Family Tree
DSP Definitions
PACKAGING
Worldwide IC Packaging Update*'^
•Entries shown in this
**Entries shown in this
"Entries shown in this
##A primary subject tab
followed directly by
SIS DSP
manner are primary subject tabs (blue).
manner are subtabs (white).
manner are document titles.
may, in some cases, have no subtabs and will be
documents.
© 1987 Dataquest Incorporated June
3)
o
o
c
Introduction to Digital Signal Processing
SEMICONDUCTOR INDUSTRY SERVICE
Dataquest's Semiconductor Industry Service (SIS) is a comprehensive information
service covering the worldwide semiconductor industry. It provides a product-oriented,
executive-level perspective intended to assist key executives and product managers with
their strategic decisions. In recognition of the fact that some semiconductor
organizations focus on specific product areas within the industry, Dataquest offers
various service options. These product-focused options provide detailed analysis in
specific product areas while omitting information about irrelevant product areas.
This research notebook focuses on the product, market, and technology issues
influencing the digital signal processing (DSP) market.
INTRODUCTION
The subject of digital signal processing (DSP) encompasses a broad range of exciting
technology, market, and semiconductor product categories. As a base technology,
advances in digital signal processing algorithms and architectures are important to all
industry participants because of the promise for new ways to solve problems that in the
past were either technologically impractical or prohibitively expensive. Simultaneously
with the maturing of the base technology, important market opportunities have emerged
that require digital signal processing techniques. This situation, in turn, presents new
opportunities for semiconductor manufacturers, systems companies, and software houses
to provide new products and services that will ultimately benefit all industry
participants. Illustrative of this fact, DSP technology today provides the basis for a wide
range of applications, such as advanced military radar systems, medical ultrasonic.
imaging, personal computer modems, and interactive talking toys.
DIGITAL SIGNAL PROCESSING NOTEBOOK
Undertaking the task of providing a comprehensive research notebook on the digital
signal processing marketplace presents some interesting challenges. First, digital signal
processing is really a technology, not simply a product category like the more mature
areas of memories or microprocessors. Hence, a significant number of different types of
products, each with its own unique characteristics, market niches, and development
tools, combine to form the complete semiconductor DSP product categories. Dataquest
partitions these products into the following four categories:
•
DSP microprocessors (DSMPUs)
•
Microprogrammable DSP (MPDSP)
•
Special-function DSP integrated circuits (SFDSP)
•
Application-specific DSP integrated circuits (ASDSP)
SIS DSP
0001092
© 1988 Dataquest Incorporated September
t
Introduction to Digital Signal Processing
A second challenge is that the market for DSP products is still relatively young. It
is difficult to adequately project future growth rates for products in this market without
fully understanding where digital signal processing can (and cannot) be effectively used
in the end applications. It is also important to understand the growth rates of the end
markets themselves. For these reasons, Dataquest has included a section on markets and
applications in this research notebook that is independent of the previously listed DSP
product categories. This section identifies where DSP can contribute to the end
applications and also presents forecast market growth rates for the next five years.
DSP Overview Section
The notebook opens with an overview of the DSP market. All of the remaining
sections of the notebook are summarized in this section. This section quickly provides
the reader with an understanding of the DSP market and products by summarizing the
information presented in each of the sections, as follows:
•
Executive Summary—Identifies the pertinent issues influencing this market in
easy-to-read bullet form
•
Forecast Summary—Combines the individual product forecasts contained in
the product sections into a total market forecast by product category and
end-use market
•
Market Dynamics—Discusses the global issues affecting the DSP market,
applications, and the players involved
•
Family Tree—Provides a definition of the product categories used in this
analysis and how one relates to another
Markets and Applications
This section analyzes the end DSP markets and applications. These markets and
applications are analyzed using a number of criteria, including technical issues,
performance requirements, DSP content within a system, and forecasts. Dataquest
broadly partitions semiconductor applications markets into the following six categories:
2
•
Military
•
Industrial
•
Communications
•
Data processing (computer)
•
Consumer
•
Transportation
© 1988 Dataquest Incorporated September
SIS DSP
0001092
Introduction to Digital Signal Processing
Product Sections
Each product section is designed to provide the reader with in-depth, detailed
information on the structure and makeup of the market segments that these products
represent. Topics discussed may include:
•
Executive summary—Provides the reader with a bulleted overview of the
factors affecting this product area
•
Forecast—Summarizes Dataquest's five-year forecast for this product segment
•
Product analysis—Compares market share, product features, pricing, life
cycles, and design wins on a product-type or family basis
•
Competitive analysis—Discusses the configuration of the market based on
market share, product positioning, applications support issues, and number of
suppliers
•
Emerging technology and trends—Identifies emerging
marketplace, including product or technology developments
•
Applications and user issues—Addresses the appropriate applications areas for
each product category as a function of technical, performance, and user
requirements
•
Historical shipment data—Records the shipment history of the various
products in this category by manufacturer
trends
in
the
DSP Technical Overview
This section provides an overview of the technical aspects of
processing. It begins with a discussion of the domain encompassed by
processing, discusses the concepts of sampled systems, and introduces
common DSP functions, such as filtering and Fast Fourier Transforms.
provided at the end of this section to define DSP terminology.
digital signal
digital signal
some of the
A glossary is
Company Profiles
The major semiconductor companies manufacturing DSP products are profiled, with
an emphasis on their strengths, weaknesses, and market presence.
SIS DSP
0001092
© 1988 Dataquest Incorporated September
Introduction to Digital Signal Processing
NEWSLETTERS
In addition to general executive-issue newsletters, the subscriber will receive
newsletters focused specifically on topics in digital signal processing. These newsletters
provide information such as:
•
•
•
•
•
•
Analyses of emerging markets and applications
New product and technology trends
Analyses of DSP-related products
Summaries of key industry events
Changes or updates to the reference material in this notebook
Other dynamic business or product issues that are of interest to industry
participants
INQUIRY PRIVILEGE
To support the unique information needs of each of our subscribers, Dataquest
provides the registered subscriber and one designated alternate the privilege of direct
access to our semiconductor staff. Two forms of this service are available:
•
Access to the semiconductor inquiry center—The inquiry center provides
assistance in finding or interpreting material in the data base notebooks or
other Dataquest published material. It is manned full-time by a staff that is
dedicated to providing immediate response to the needs of our subscribers.
•
Access to the semiconductor research staff—Clients may seek additional
commentary on or clarification of the published material from the
semiconductor research staff. Using this feature, clients may interact with
industry experts on a one-on-one basis to disciKS attitudes and opinions about
topics covered in the service.
RESEARCH METHODOLOGY
The methodology used by Dataquest in support of this research has a number of
unique features. The primary research staff has extensive industry experience in the
market area covered by this research. Adding to this base knowledge, Dataquest has
significant numbers of contacts within both the semiconductor manufacturer and
end-user communities to contribute a balanced perspective to this work. Additional
insight is provided from other research groups at Dataquest studying technology areas
related to digital signal processing, such as the Telecommunications Industry Service and
Technical Computer Systems Industry Service. A financial outlook provided through
Dataquest's extensive business contacts complements this research. Figure 1
summarizes the perspectives that contribute to this work.
4
© 1988 Dataquest Incorporated September
SIS DSP
0001092
Introduction to Digital Signal Processing
Figure 1
Worldwide Input to Dataquest Semiconductor Forecasts
End-Use Perspective
Semiconductor iVIanufacturers
Equipment Forecast
Procurement Survey
I/O Ratio Analysis
Inventory Levels
Operation Analysis
Buyer Attitudes
End-Market Growth
End-Market Contacts
Product Shipment Data Base
Product Introductions and Withdrawals
Technology Trends
Product Application Analysis
Competitive Analysis
Semiconductor Start-Ups
Japanese, European, and
Asian Market Trends
Capacity Analysis
Capital Spending
Silicon Purchases
Revenue/Square-Inch Analysis
Japanese, European, and
Asian Manufacturers
Executive Contacts
Financial Perspective
Other Input
Dun & Bradstreet Economic Analysis
Dataquest Financial Services Program
Wail Street Viewpoint
Prudential-Bache Contacts
Annual Reports
Other Dataquest Services
- Computers
- Telecommunications
- CAD/CAM
- Computer Storage
- Office Automation
- Printers
- Semiconductor Applications
Government Contacts
Trade Publications
R&D Facilities
Industry Associations
Source: Dataquest
September 1988
SIS DSP
0001092
© 1988 Dataquest Incorporated September
Introduction to Digital Signal Processing
Manufacturing Perspective
Dataquest's Semiconductor Information Group comprises seven different semiconductor information services. The combined worldwide research network of these
services provides input to the forecasts based on all relevant aspects of the industry.
Dataquest uses its extensive product shipment data base and information about new and
obsolete products when compiling the annual consumption forecasts.
Executive level input from almost all semiconductor manufacturers in the free
world provides insight into the industry environment, current business levels, and major
competitors. Information regarding technology trends, capital spending, capacity
utilization, and product acceptance is analyzed by Dataquest's industry experts in
formulating the forecasts.
Other considerations are analyses of wafer starts, wafer consumption, and revenue
per wafer. Our worldwide research offices and contacts in Western Europe, Japan, and
Asia provide primary input as well as checks and balances for each forecast cycle.
End-Use Perspective
Research about trends in the application and procurement of semiconductors is
another cornerstone of our forecasts. The Semiconductor Application Markets service
contributes this end-use analysis through equipment forecasts, procurement surveys, and
I/O ratio analysis. The Semiconductor User Information Service studies buyer attitudes,
changes in operation methods, inventory analysis, and short-term pricing trends.
Financial Perspective
Dataquest's general economic outlook and financial viewpoint is based on business
surveys and financial analyses from Dataquest's parent company. Dun & Bradstreet. In
addition, Dataquest's Financial Services Program provides close contact with venture
capitalists, fund managers, and financial institutions. Through our exclusive relationship
with Prudential-Bache, Dataquest has access to Wall Street's attitudes about all areas of
technology. These sources, combined with Dataquest's own research on financial and
economic trends, provide one of the cornerstones of our forecasts.
Other Input
Close association with other Dataquest information services provides additional
input to the forecast accuracy and quality. This includes analysis of key industries that
consume semiconductors such as technical, business, and personal computers. In
addition, input is solicited from services that analyze CAD/CAM, telecommunications,
computer storage, printers, office automation, and graphics terminal industries.
Dataquest analysts also attend industry trade shows and technical symposiums to monitor
the latest product introductions and technology trends. Other input comes from
government contracts and publications, industry associations, and trade publications.
6
© 1988 Dataquest Incorporated September
SIS DSP
0001092
DSP Executive Summary
EXECUTIVE SUMMARY
Broadly, the digital signal processing (DSP) marketplace consists of a tremendous
variety of different semiconductor product offerings targeted at a tremendous variety of
end applications. The need for different DSP semiconductor products is driven by the
performance requirements of DSP systems, where sample rates range from hundreds of
hertz at the low-performance end to hundreds of megahertz at the high-performance
end. Historically, this has allowed plenty of room for manufacturers to search for unique
market niches in which to target their business.
Currently, the products serving the end DSP marketplace are undergoing some
fundamental changes. It is important for semiconductor manufacturers to understand
and react effectively to these changes, or risk disastrous business consequences.
Highlights of these changes are presented in this section.
The semiconductor DSP marketplace is characterized by four different product
categories, each addressing a different portion of the user market. Currently, each of
these product categories is roughly the same size in revenue. However, the growth rates
of these markets are not the same. There are nearly 50 semiconductor companies
supplying products to different portions of this market. Very few of these companies
have product offerings in all of the product categories; many offer products only in
niches within one of the categories.
At this time, the largest market for DSP semiconductors is communications,
followed closely by the military. In 1987, these two market areas together accounted for
an astounding 77 percent of total DSP revenue. However, market applications for DSP
semiconductors are all-pervasive. DSP semiconductors are being designed into
applications ranging from simple toys to high-performance military radar systems.
Driven by both price declines and technology advances, Dataquest expects that by 1992
the largest application market for DSP semiconductors will be consumer products.
DSP Technical Introduction
Digital signal processing is a technique for manipulating (processing) signals
digitally. A simple block diagram of a DSP system is shown in Figure 1. Because the
world we live in is analog (continuous time, continuous sight, continuous smell, etc.),
Figure 1 shows the requisite analog-to-digital and digital-to-analog converters. These
converters change the analog world into a digital representation upon which a DSP
system can operate. Usually at least one (sometimes both) of these converters is present
in a DSP system.
SIS DSP
0001215
© 1988 Dataquest Incorporated September
DSP Executive Summary
Figure 1
Block Diagram of a Generic DSP System
I
Analog-to-Digital
Converter
%
I
Digital Signal
Processing
Digital-to-Analog
Converter
•
Output
Input
Source: Dataquest
September 1988
DIGITAL SIGNAL PROCESSING MARKET FORECAST
The Dataquest market forecast for DSP semiconductors is shown graphically in
Figure 2. It includes breakouts of the four product subcategories that combine to form
the cumulative forecast. Highlights of this forecast include the following:
•
We forecast the worldwide DSP revenue for 1988 to be $586 million, an
increase of 30.8 percent over 1987.
•
Worldwide DSP revenue will grow to more than $1 billion in 1990, representing
a compound annual growth rate (CAGR) of nearly 34 percent over 1987
revenue.
•
All product categories except MPDSP are growing at impressive CAGRs of
greater than 30 percent; MPDSP through 1992 will likely experience a CAGR
of less than 2 percent.
© 1988 Dataquest Incorporated September
SIS DSP
0001215
DSP Executive Summary
Figure 2
Worldwide DSP Revenue Forecast
Millions of Dollars
2100
$1.S66
ASOSP
1800
SFDSP
MPDSP
1500
^'7^
CAGR
1985-19S2
33.6%
DSMPU
1200
900-1
600
$448
300
0
1987
1992
Source: Dataquejt
September 1988
Product Categories/Revenue Estimates
Dataquest partitions DSP semiconductor products into four different categories.
These categories, including 1987 and 1988 revenue estimates, are shown in Table 1.
Table 1
DSP Product Category Revenue Estimates
(Millions of Dollars)
Product Category
Revenue Estimates
1987
1988
Abbreviation
DSP Microprocessors
(DSMPU)
$ 98
$147
Microprograiranable DSP
(MPDSP)
$139
$150
S p e c i a l - F u n c t i o n DSP
(SFDSP)
$113
$158
A p p l i c a t i o n - S p e c i f i c DSP
(ASDSP)
$ 98
$131
Source;
SIS DSP
0001215
© 1988 Dataquest Incorporated September
Dataquest
September 1988
DSP Executive Summary
Factors Influencing DSP Revenue Growth
A number of factors are contributing to the continued growth of revenue for DSP
products:
•
DSP design expertise, which had existed primarily in the military sector, has
recently spread into the commercial sector.
•
Entirely new applications, which were not practical prior to maturing DSP
techniques, have developed. An example is the talking "Julie" doll introduced
by Worlds of Wonder.
•
Tremendous price declines of DSP products (some close to 40 percent per year)
have brought previously sophisticated, high-priced technology to low-cost,
mass-market items like toys and personal computer modems.
•
Conversion of older analog products to newer, more reliable designs with more
features using digital signal processing is taking place. An example is the
migration from analog oscilloscopes to newer digital scopes.
•
Availability of powerful hardware and software development tools has aided
the system designer in incorporating digital signal processing into end products.
DIGITAL SIGNAL PROCESSING PRODUCTS
This section highlights some of the key issues affecting each of the specific product
categories. Included are a category description, reasons for future revenue growth,
major suppliers, trends, and competitive issues.
DSP Microprocessors (DSMPU)
•
DSMPU products are general-purpose, programmable digital signal processors.
They are similar to microprocessor architectures containing hardware
multipliers and other architectural optimizations that address specifically the
needs of the DSP marketplace.
•
The DSMPU market can expect significant revenue growth through 1992,
achieving annual revenue of nearly $700 million at a CAGR of nearly
48 percent.
•
The three leading DSMPU manufacturers continue to be Texas Instruments,
NEC, and Fujitsu, in that order. Combined, they contribute more than
70 percent of DSMPU revenue.
© 1988 Dataquest Incorporated September
SIS DSP
0001215
DSP Executive Summary
•
Analog Devices, AT&T, and Motorola form the core group of "second-tier"
DSMPU manufacturers. While the revenue for each of them is less than
$10 million, they have each adopted strategies designed to secure their
long-term commitment to this market.
•
Dataquest expects the new generation of floating-point DSMPUs to gain a
nearly 40 percent share of the market by 1992.
•
For applications where DSMPUs are sufficiently fast, they will dominate over
MPDSPs and ASDSPs on system cost alone.
•
There are currently 13 manufacturers of different DSMPU architectures
available to designers of DSP systems.
•
Dataquest expects a shakeout in the number of suppliers to this market
segment over the next three years, similar to the shakeout that occurred in
the microprocessor market in the late 1970s and early 1980s. Ultimately we
expect no more than three major suppliers and two minor suppliers to the
general-purpose DSMPU market.
•
U.S. manufacturers are currently better positioned than Japanese manufacturers to win the bulk of the DSMPU design slots available for the foreseeable
future.
Microprogrammable DSP (MPDSP)
•
In the past, MPDSP components have been labeled by the industry as
"bit-slice" processors. However, the term "bit slice" is somewhat archaic and
does not adequately describe the newer 32- and 64-bit processors available
today. The primary components which form this category are:
-
Microprogrammable Arithmetic Units (MAUs)
-
Multipliers and Multiplier-Accumulators (MACs)
•
. The MPDSP category is heavily populated with IC manufacturers. Sixteen
manufacturers supply products to this segment of the market. The leading
suppliers are AMD, Analog Devices, IDT, Texas Instruments, TRW, and Weitek.
•
Annual revenue in this product category is expected to peak at $177 million in
1990, then to decline gradually.
SIS DSP
0001215
© 1988 Dataquest Incorporated September
DSP Executive Summary
•
While overall growth in this market is expected to be relatively flat, a
subportion of this market, made up of floating-point multipliers and MAUs,
should continue to experience growth of approximately 20 percent through
1990 before beginning to flatten out.
•
Competitive pressure from products in both the DSMPU and SFDSP categories
are the major contributors to the revenue slowdown in the MPDSP market.
Special Functions DSP (SFDSP)
•
Products in this category have dedicated (usually not general-purpose) DSP
features. They include modems, codecs, speech processors, digital
television/circuits, filters, and Fast Fourier Transform (FFT) functions. Some
of these functions have been traditionally implemented using analog
techniques. This category includes only those devices which implement the
functions using DSP architectures.
•
Revenue in this product category is expected to continue growing at a CAGR
of about 36 percent through 1992, achieving annual revenue of about
$550 million.
•
Historically, DSP modem chip sets have fueled the growth of this market,
achieving 1987 revenue of about $100 million.
•
Digital chip television sets are expected to provide impressive growth to this
segment over the next few years, achieving annual revenue of $270 million
by 1992.
Application Specific DSP (ASPSP)
•
Products in this category are generally not standard products sold on the open
market. Instead, they are usually custom architectures designed using
standard cell or gate array techniques for a specific application and user.
•
Revenue growth in this category is expected to be strong through 1992 with a
CAGR of 37 percent, achieving annual revenue of $461 million.
•
There are over 30 manufacturers able to supply ASDSP solutions to their
customers.
•
Many DSMPU manufacturers are in the process of migrating their early
DSMPU architectures into standard cells.
•
Many of the applications now using MPDSP will shift to ASICs constructed
from ALUs, MACs, and registers currently in cell libraries.
© 1988 Dataquest Incorporated September
SIS DSP
0001215
DSP Executive Summary
APPLICATION TRENDS
Dataquest segments the semiconductor application markets into six primary
segments. A snapshot of the percentage of total DSP revenue distributed across these
markets both now and in 1992 is shown in Figure 3.
•
Currently military and communication applications account for about
77 percent of all DSP revenue.
•
As observed in Figure 3, Dataquest expects consumer application areas, led by
digital television, to become the largest DSP application segment by 1992.
Figure 3
Percentage of DSP Revenue
Across the Six Applications Markets
k\'"'i
Communications
^M
Military
Wi:f:4 Consumer
I
1^1
0.2%
I Computer
Industrial
2.8%
t^S;:;S::;si Automotive
1992
1987
Source: Dataquest
September 1988
GENERAL ISSUES AND TRENDS
•
DSP devices are "design win" products, much like microprocessors.
Manufacturers must work with and support end users in order to secure design
slots.
•
As device architectures become more sophisticated, product support and
application assistance become important issues to end users; in some cases,
more important than the details of architectural differences between suppliers.
SIS DSP
0001215
© 1988 Dataquest Incorporated September
DSP Executive Summary
S
•
In all four product categories, prices have been dropping dramatically, at rates
exceeding 30 percent per year. The significance of this is underscored by
observing revenue growth projections: greater than 33 percent per year
through 1992.
•
Across all four DSP product categories, CMOS is the dominant process
technology. Alternate technologies such as ECL or CaAs may achieve
penetration in small performance niches, but none are expected to displace
CMOS in the foreseeable future.
© 1988 Dataquest Incorporated September
SIS DSP
0001215
DSP Forecast Summary
HISTORICAL ECONOMIC FACTORS
The worldwide semiconductor industry experienced a much-needed recovery during
1987. For the North American semiconductor market, 1987 turned out to be a better
year than even Dataquest had predicted. Although our original 1987 forecast for North
American consumption projected 12 percent growth over the preceding year, the actual
growth of the U.S. market was 19 percent.
The fundamental optimism that this recovery would continue in 1988 was shaken
somewhat by events on Wall Street in the final quarter of 1987. Nevertheless, Dataquest
believes that the industry will maintain its momentum in 1988. Dataquest forecasts
more than 30 percent growth in worldwide semiconductor shipments in 1988 over 1987.
This translates into a worldwide semiconductor market that will reach nearly $50 billion
by the end of this year, with the highest overall growth rate in the area of MOS (metal
oxide semiconductor) products. Through 1992, Dataquest expects a worldwide
semiconductor compound annual growth rate (CAGR) of 15.5 percent, achieving revenue
of $75 billion.
Figure 1 establishes that although some analysts speak of the semiconductor
business as a maturing industry, its most meteoric growth has really been during the last
five years. In fact, in the last five years, annual consumption has grown by $18 billion,
surpassing an annual level that took more than 25 years to attain.
Figure 1
Historical Worldwide Semiconductor Consumption
1956-1987
Billions of Dollars
19S6
1961
1966
1971
1976
1981
1986
Source: I>ataqucst
October ISBS
SIS DSP
0001470
© 1988 Dataquest Incorporated October
DSP Forecast Summary
Digital Signal Processing Revenue Forecast
Digital signal processing (DSP) is still a relatively new technology that should
outgrow the worldwide semiconductor industry over the next few years. While Dataquest
estimates that revenue for worldwide semiconductor shipments will grow at a CAGR of
15.5 percent through 1992, DSP product revenue during this same period is forecast to
grow at a significantly higher CAGR of 33.6 percent. Table 1 and Figure 2 illustrate
both historical and estimated worldwide DSP revenue growth through 1992.
Table 1
Worldwide DSP Revenue Forecast
(Millions of Dollars)
Total DSP
Revenue
Actual
1985 1986 1987
1988 1989
Forecast
1990
1991
$208 $316 $448
$586 $778 $1,057 $1,410 $1,866
CAGR
CAGR
1985-1987 1988-1992
1992
46.8%
Source:
33.6%
Dataquest
October 1988
Figure 2
Worldwide DSP Revenue Forecast
Millions of Dollars
2000
1985
1986
1987
1988
1989
1990
1991
1992
guest
Source: Dataquest
)ber 19
October
1988
© 1988 Dataquest Incorporated October
SIS DSP
0001470
DSP Forecast Summary
Product Segmentation
Dataquest segments DSP products into four distinct categories, as shown below:
•
DSP microprocessors (DSMPUs)
•
DSP building blocks (MPDSPs)
•
Special-function DSP ICs (SFDSPs)
•
ASIC DSP ICs (ASDSPs)
A full description of the products in each category is given in the section entitled
"DSP Family Tree."
Estimated worldwide revenue growth segmented by DSP product category for 1985
through 1992 is shown in Table 2 and Figure 3. Of particular significance is the DSMPU
category, which is expected to grow faster than other product categories, with a
projected CAGR of 47.5 percent through 1992. Contributing to continued DSMPU
growth is a maturing of DSP as a technology and the simultaneous maturing of these
high-performance, flexible, and programmable devices. The market potential for
DSMPUs used for real-time signal processing applications is analogous to the market
impact of general-purpose microprocessors to data processing applications in the 1980s.
Table 2
Worldwide DSP Revenue Forecast by Product Category
(Millions of Dollars)
Forecast
Actual
1995 1999 1997 1999 1989 199Q
1991
DSMPU
MPDSP
SFDSP
ASDSP
$ 34 $ 62 $ 98 $147 $221 $
97 111 139 150 161
34
75 113 158 215
43
68
98 131 181
320 $ 485 $
177
170
310
415
250
340
1992
695
160
550
461
Total DSP
Revenue $208 $316 $448 $586 $778 $1,057 $1,410 $1,866
CAGR
CAGR
1985-1987 1988-1992
69.8%
19.7'<b
82.3%
51.0%
47.5%
1.6%
36.6%
37.0%
46.8%
33.6%
Source:
SIS DSP
0001470
© 1988 Dataquest Incorporated October
Dataquest
October 1988
DSP Forecast Summary
Figure 3
Worldwide DSP Revenue Forecast by Product Category
Millions of Dollars
20001800
2
MPDSP
1S00
1400-
DSMPU
SFDSP
^3^^
ASDSP
12001000
iO&A
tm
4da^
2^
0
1985
19S6
1987
1988
19SS
1990
m
1991
1992
Source: Daiaquest
October 198S
ASICs and SFICs that perform DSP functions are also expected to experience rapid
growth. These devices are generally used in systems requiring optimized architectures
to solve high-performance signal processing problems. Interestingly, these devices are
generally not directly competitive with DSMPU products, as their more optimized and
less general-purpose architectures usually allow performance that exceeds that of the
DSMPU.
The market growth for the MPDSP market is also of interest, but not because of
expected high growth rates. MPDSP products should experience sluggish growth through
the early 1990s because of competitive products in three different areas:
•
Very fast and flexible DSMPUs
•
Incorporation of MPDSP cells into standard cell libraries,
customization of architectures in silicon instead of on a board
•
Emergence of RISC microprocessors that will displace many MPDSP products
in high-performance control applications
© 1988 Dataquest Incorporated October
allowing
SIS DSP
0001470
DSP Forecast Summary
Market Segmentation
Dataquest segments semiconductor applications into six primary markets as shown
below:
Military
Communications
Industrial
Computer
Automotive
Consumer
A description of the uses of DSP within these applications markets is provided in the
section on Markets and Applications. Table 3 summarizes the usage of DSP within these
six markets.
Table 3
Worldwide DSP Revenue Forecast by Aj^lication
(Millions of Dollars)
Military
Communications
Industrial
Computer
Consumer
Automotive
$ 88 $121 $161 $201 $242 $
94
1
21
4
136
16
26
16
0
Q
CAGR
CAGR
1985-1987 1988-1992
Forecast
Actual
1995 1986 X997 1989 3,999 1990
187
22
37
41
196
51
46
92
pa
, 1,3
230
80
59
163
3t9
X991
1992
289 $
344 $
403
ss.a-tb
19.0%
272
143
88
254
13
323
217
135
366
25
352
335
212
511
53
40.7%
435.4%
30.9%
207.5%
15.8%
60.3%
47.0%
53.7%
154.6%
Total DSP
Product
Revenue $208 $315 $448 $587 $778 $1,059 $1,410 $1,866
N/C
46.6%
33.6%
N/C = Not Computed
Source:
SIS DSP
0001470
© 1988 Dataquest Incorporated October
Dataquest
October 1988
Market Dynamics
DSP Markets and Applications
INTRODUCTION
Understanding the growth prospects for DSP products requires a deep understanding
of the size, growth rates, and trends of the end markets served by the products. It
further requires an understanding of where this technology called digital signal
processing (DSP) fits into the end applications that comprise the market.
While other sections in this report focus on DSP products, this section focuses on
end DSP markets and applications independent of the specific products. Perspective is
provided toward understanding the various applications and market forces that influence
the growth of semiconductor DSP products. Three global market trends are compounding
the rapid growth of DSP products:
•
Growth prospects for existing markets already served by DSP technology
•
Identification of emerging market opportunities that benefit from DSP
technology
•
Transitioning from older analog signal processing technology to digital signal
processing technology
END MARKET DEFINTTION
As shown in Table 1, Dataquest divides semiconductor applications into six major
market areas: communications, industrial, military, computer, consumer, and
automotive. Under the six major categories are subdivisions that are (or will be) well
served by DSP products.
SIS DSP
0001325
© 1988 Dataquest Incorporated September
DSP Markets and Applications
Table 1
Six Semiconductor Applications Markets
with Subcategories Served by DSP Technology
Communications
Industrial
Military
Modems
DTMF receivers
Transmultiplexers
Speech synthesis
Speech recognition
Speech compression
Mobile communications
Video teleconferencing
Test equipment
Medical equipment
Office automation
Inspection equipment
Remote monitors
Robot systems
Global pos satellites
Motor control
Radar
Sonar Navigation
Fuses
Communication
Reconnaissance
Computer
Consumer
Transportation
Compact disc players
Digital video
Electronic cameras
Toys
Antilock brakes
Distance sensors
Lane sensors
Arithmetic accelerators
Array processors
Image processing
Graphics
Geophysical processing
Source:
Dataquest
September 1988
Table 2 shows Dataquest's revenue forecast for DSP products in the six major
market categories. Notice that in the mid-1980s, military and communication
applications were by far the largest consumers of DSP semiconductors. While these two
segments of the market will remain important over time, Dataquest expects that
consumer applications will be the largest users of DSP products by 1992. This rapid
growth will be fueled by entertainment applications such as digital television and digital
audio.
© 1988 Dataquest Incorporated September
SIS DSP
0001325
DSP Markets and Applications
Table 2
Worldwide DSP Revenue Forecast by Application
(Millions of Dollars)
CAGR
1985-1987
CAGR
1988-1992
403
352
335
212
511
53
35.8%
40.7%
435.4%
30.9%
207.5%
N/C
19.0%
15.8%
60.3%
47.0%
53.7%
154.6%
$587 $778 $1,059 $1,410 $1,866
46.6%
33.6%
Actual
1985 1986 1987
Military
Conununications
Industrial
Computer
Consumer
Automotive
Total DSP Product
Revenue
$ 88 $121 3161
94
1
21
4
0
136
16
26
16
187
22
37
41
g 0.1
$208 $315 $448
1988 1989
Forecast
L991
1990
$201 $242 $
196 230
51
80
46
59
92 163
1.3 3.8
289 $
272
143
88
254
13
344 $
323
217
135
366
25
1992
N/C = Not Computed
Note: Columns may not add to totals shown because of rounding.
Source:
Dataquest
September 1988
Figure 1 provides a snapshot look at the same revenue data for the years 1987 and
1992, expressed as a percentage of total DSP revenue. This perspective illustrates the
growing importance of DSP technology, especially to the industrial and consumer
markets.
Figure 1
Revenue for DSP Products
in Six Major Semiconductor Application Markets
Communications
^m
IVIilltary
l^'Ji^^l Consumer
I
^ H
I Computer
Industrial
2.8%
Automotive
Negligible
%
1987
1992
Source: Dataquest
September 1988
SIS DSP
0001325
© 1988 Dataquest Incorporated September
DSP Markets and Applications
MARKET ANALYSIS
Introduction
The market for DSP products was essentially created in 1977 by TRW with the
introduction of the first monolithic multiplier, originally targeted for internal TRW use.
Soon thereafter, the first monolithic video-speed flash converter was also introduced by
TRW, and the combination of the two made high-performance, real-time DSP operations
a reality. The military market quickly embraced these new products, and still remains
one of the largest (and most sophisticated) users of DSP technology.
The early 1980s saw the introduction of single-chip DSPs such as TI's TMS32010,
AMI'S 28211, and NEC's 7720. These processors found tremendous success initially in the
communications market for applications in modems, DTMF receivers, and speech
applications.
Today, the military and communicationis markets are still the two largest users of
DSP semiconductor products. However, DSP techniques are being used in a plethora of
new ways, spurring growth in the industrial, computer, consumer, and automotive
markets.
Military
As shown in Table 3, Dataquest expects revenue growth for military applications to
increase at a compound annual growth rate (CAGR) of 19 percent through 1992. Military
expenditures for DSP products should total the second largest of the six markets in 1992
(see Table 2). However, the percentage of military DSP revenue drops from 42.3 percent
in 1985 to 21.7 percent in 1992, compared with total DSP revenue. This is attributable
to the much higher DSP growth rates expected in the commercial marketplace.
The most important applications using DSP technology in the military market
include the following:
Radar
Sonar
Navigation
Fuses
Communications
Reconnaissance
© 1988 Dataquest Incorporated September
SIS DSP
0001325
DSP Markets and Applications
Table 3
DSP Revenue Forecast for the Military Market
(Millions of Dollars)
Actual
1986 1987
Product
CateqocY
1985
DSMPUs
HPDSP
ASOSP
SFDSP
$16
43
20
9
$ 26 $ 35
55
65
28
44
12
17
$88
$121
Total
Military
DSP Revenue
$161
Military as Percent
of Total DSP
Revenue
42.3% 38.3« 35.9%
Forecast
1989 1990 1991
1992
CAGR
1985-1987
CAGR
1988-1992
$ 49 $ 64
73
81
58
72
21
25
$ 84 $110
92
87
106
88
36
30
$139
96
128
42
48.0%
23.0%
48.4%
37.7%
29.8%
7.1%
21.9%
19.0%
$201
$289
$344
$405
35.8%
19.0%
1988
$242
34.3% 31.1% 27.3% 24.4% 21.7%
Source:
Dataquest
September 1988
As in other market segments, DSMPU products will experience the largest part of
the market growth; nearly 30 percent through 1992.
Microprogrammable products enjoy their largest market penetration in military
applications. In 1985, 44 percent of microprogrammable revenue was funded by military
applications. In 1992, Dataquest expects that 60 percent of MPDSP revenue will be
funded by military applications. This growth is caused by two primary factors:
•
Because of shorter design cycles in the commercial market, MPDSP usage is
being displaced by the other three DSP product categories, slowing total
MPDSP revenue growth.
•
Many military projects take from three to seven years to enter production;
hence, products designed using MPDSP components a number of years ago are
only now beginning to enter production.
Communications
Usage of digital signal processing components in the communications market
experienced significant growth in the early 1980s. As shown in Table 4, this was spurred
by two primary applications: modems and DTMF receivers. These two applications have
historically used both first-generation DSMPU and special-function devices almost
exclusively. Very few applications in the communications market use MPDSP
components.
SIS DSP
0001325
© 1988 Dataquest Incorporated September
DSP Markets and Applications
Table 4
DSP Revenue Forecast for the Communications Market
(Millions of Dollars)
Application
Cateqory
Modems
DTMF Rxr
Speech Synth.
Speech Recog.
Speech Compres.
Mobile Conunun.
Video Telecon
Other
1985
Actual
1986
$45.0
25.0
2.0
0.7
1.0
0.3
1.4
19.0
$ 70.0 $101.0
35.0
30.0
6.0
4.0
1.0
0.9
3.4
2.0
1.0
0.5
2.0
1.7
37.0
27.0
$94.3
$136.3 $186.6
45.3%
43.1% 41.7%
1987
CAGR
1985-1987
CAGR
1988-1992
$106.0 $127.0 $151.0 $171.0 $161.0
18.0
20.0
23.0
27.0
31.0
26.0
40.0
16.0
11.0
8.0
7.0
10.0
4.4
2.6
1.6
18.0
25.0
14.0
9.6
5.7
12.0
20.0
6.0
4.0
2.0
5.0
7.0
3.0
2.4
3.8
70.0
65.0
54.0
46.0
39.0
49.7%
18.3%
73.2%
19.5%
34.4%
82.6%
19.5%
40.7%
11.1%
(12.5%)
49.5%
58.1%
44.7%
77.8%
30.7%
15.8%
$195.9 $229.8 $272.0 $323.4 $351.9
40.7%
15.8%
1988
1989
Forecast
1991
1990
1992
Total
Communications
DSP Revenue
Communications as
Percent of Total
06P Revenue
Note:
33.4%
29.5%
25.7%
22.9% 18.9%
Columns may not add to totals shown because of rounding.
Source:
Dataquest
September 1988
To add perspective to the revenue significance of modems and DTMF receivers, in
1985 these two applications accounted for more than 33 percent of all DSP revenue.
Partly because of both price erosion and proliferation of DSP technology into other
areas, the percentage of DSP revenue accounted for by both of these applications in 1988
should be down to 23 percent.
Modems will continue to contribute significantly to total communications revenue
through 1992. This is attributed in part to new generations of modems defined by the
Consultive Committee for International Telephony and Telegraphy (CCITT) such as the
CCITT V.22bis (2400 bps full duplex) and CCITT V.32 (9600 bps full duplex). These
modems use more sophisticated signal processing algorithms (such as adaptive equalizers)
than previous generations of slower-speed modems. In the case of the V.32, echo
cancellation algorithms are also required. Both adaptive equalizers and echo cancellers
require DSP algorithms for implementation. In addition, these modems are beginning to
use second-generation DSMPUs, which also command higher prices than first-generation
processors.
Revenue for DTMF receiver applications is expected to decline at a CAGR of about
12.5 percent through 1992, due mainly to price declines for the first-generation DSMPUs
that normally implement this function. This is in contrast to a positive CAGR of
4.1 percent in the number of installed central office and PBX lines through 1992.
© 1988 Dataquest Incorporated September
SIS DSP
0001325
DSP Markets and Applications
Dataquest expects remaining communications applications using DSP technology to
begin experiencing impressive revenue gains through 1992 as shown in Table 4. Voice
messaging applications using speech synthesis and speech compression algorithms will
help drive this revenue growth. Mobile communications is another large growth area
with enormous potential for DSP.
Industrial
Dataquest expects the industrial market to turn in impressive DSP revenue gains as
shown in Table 5, with 1992 revenue at nearly $335 million, more than six times the 1988
revenue. This number is even more impressive given the small revenue number
attributed to this market in 1985.
As shown in Table 5, the test equipment category of the industrial market consists
of applications such as the following:
•
Digital oscilloscopes
•
Spectrum analyzers
•
Vibration analysis
•
Recording instruments
•
Automatic test equipment
In general, test equipment applications tend to use DSMPU and SFDSP products for
operations such as fast Fourier transforms (FFTs) and digital filters. Digital
oscilloscopes and spectrum analyzers account for about 70 percent of the test equipment
revenue in 1988. In 1992, this percentage should decline to about 60 percent, as the
other equipment applications incorporate larger amounts of DSP.
As shown in Table 5, the medical equipment category of the industrial market
consists of applications such as the following:
•
Computed tomography (CT) scanners
•
Magnetic-resonance imaging
•
Doppler ultrasound
•
Hearing aids
•
Vision aids
SIS DSP
0001325
© 1988 Dataquest Incorporated September
DSP Markets and Applications
Table 5
DSP Revenue Forecast for the Industrial Market
(Millions of Dollars)
Application
Category
1985
Actual
1986 1987
1988
1989
Ebcecast
1990
1991
1992
CAGR
1985-1987
CAGR
1988-1992
$13.0 $15.0 $ 17.0 $ 20.0 $ 23.0
4.4
5.2
6.0
7.0
8.2
4.7
4.8
5.0
5.2
5.6
0.8
1.0
1.2
1.4
1.6
0.8
1.0
1.2
1.4
1.6
0.4
0.5
0.7
0.8
1.0
4.8
1.8
2.4
3.0
3.8
N/C
N/C
N/C
N/C
N/C
N/C
N/C
15.3%
16.8%
4.5%
18.9%
18.9%
25.7%
27.8%
17.0
2.3
6.0
2.0
7.0
24.0
2.5
8.0
2.4
11.0
N/C
N/C
N/C
N/C
N/C
47.9%
13.6%
41.4%
27.8%
106.9%
33.0
5.0
3.0
25.0
51.0
10.0
5.0
36.0
85.0
20.0
10.0
55.0
N/C
N/C
N/C
N/C
98.3%
111.5%
111.5%
92.6%
6.0
10.0
10.0
26.0
29.0
10.0
20.0
16.0
40.0
44.0
16.0
40.0
24.0
56.0
67.0
41.4%
108.2%
216.2%
347.2%
435.4%
127.0%
100.0%
56.5%
53.8%
60.3%
$50.6 $80.3 $143.0 $217.0 $335.0
435.4%
60.3%
Test Equipment
Oscilloscope
Spec Analysis
Vlb Analysis
Record Instr.
ATE
Other
N/C $ 9.4 $ 9.1
N/A
3.4
3.8
N/A
4.0
2.4
M/A
0.4
0.6
N/A
0.4
0.6
N/A
0.3
0.2
N/A
1.4
1.0
Medical
CT Scanners
MRI
Ooppler Ultr
Hearing Aids
N/C
N/A
N/A
N/A
N/A
1.8
0.9
0.5
0.4
0
3.2
1.2
1.2
0.6
0.2
5.0
1.5
2.0
0.9
0.6
7.8
1.8
2.8
1.2
2.0
12.0
2.1
4.2
1.6
4.0
Office
Image Compr
Copiers
Laser Print
0
0
0
0
0.2
0.1
0
0.1
0.4
0.2
0
0.2
5.5
1.0
0.5
4.0
13.0
2.0
1.5
9.0
Inspection
Remote Monitor
GPS
Motor Control
Other
0.1
0.3
0.1
0.1
0.2
0.1
0.6
0.5
0.5
3.3
0.2
1.3
1.0
2.0
4.3
0.6
2.5
4.0
10.0
10.0
2.0
5.0
6.0
16.0
16.0
Total Industrial
DSP Revenue
$0.8 $16.4 $21.5
Industrial as
Percent of Total
OSP Revenue
0.3%
5.2%
4.8%
8.6% 10.3%
13.5%
15.4%
18.0%
N/C ' Not Computed
N/A > Not Available
Note: Columns may not add to totals shown because of rounding.
Source:
8
© 1988 Dataquest Incorporated September
Dataquest
September 1988
SIS DSP
0001325
DSP Markets and Applications
Dataquest expects much of the revenue growth in the medical market through 1992
to come from a dramatic increase in the number of hearing aids. By 1992, 46 percent of
medical revenue could come from hearing aids. This compares with less than 7 percent
in 1988.
According to the Ear Research Institute in Southern California, there are two
million people in the United States who are either totally deaf or lack the ability to
detect speech without some form of external aid. Techniques in the form of either
electrical stimulation of the cochlea by implanted electrodes or tactile cutaneous
stimulation are often used in order to simulate hearing for the deaf. There exist another
12 million hearing impaired individuals who suffer from serious hearing loss, often
treated using conventional analog hearing aids. A number of different digital signal
processing techniques are being developed to aid both groups of people, providing a level
of hearing quality not achievable today using analog technology.
As shown in Table 5, the office equipment category of the industrial market consists
of applications such as the following:
•
Image compression
•
Copiers
•
Laser printers
Dataquest expects the office category of the industrial market to achieve the most
significant revenue growth of any category in the industrial market. We expect revenue
to reach nearly $85 million by 1992, from essentially no revenue in 1985. Furthermore,
there is additional upside potential in this category, as all of the applications are
currently growing at impressive rates.
Image compression techniques will be used to compress both gray-scale and color
images for storage and transmission requirements. Examples include transmission of
images over telephone lines using group IV facsimile and storage of high-quality images
on computer hard discs. A number of different DSP-based compression techniques
currently exist, including:
•
Discrete (or fast) cosine transform
•
Adaptive differential pulse-code modulation (ADPCM)
•
Vector quantization
Laser printers are beginning to incorporate DSP components for generating fonts
and graphics. This is important, particularly for page-description languages (PDLs) such
as Adobe Systems' "PostScript." PDLs such as this allow integration of both texts—with
a wide range of different font types and sizes—and graphics for desktop publishing
applications. Digital copiers are also beginning to appear on the market using DSP
operations such as digital filters.
SIS DSP
0001325
© 1988 Dataquest Incorporated September
DSP Markets and Applications
An interesting application area that deserves some discussion is global positioning
satellites (GPS). GPS is a three-dimensional coordinate system using triangulation
techniques from satellites orbiting the earth. The GPS system has both military and
commercial uses. Military uses include:
•
Replacement of terrain maps for cruise missiles with GPS location-finding
electronics
•
Soldier, transportation, and equipment location and placement on battlefields
Commercial applications include:
•
Surveying equipment
•
Aircraft and boat location
•
Navigation
•
Terrain maps for transportation, including automobiles
Computer
As shown in Table 6, DSP revenue in the computer market as a percentage of total
DSP revenue is expected to remain relatively stable at about 10 percent through 1992.
As can be seen in Table 6, applications that serve the computer market include:
•
Arithmetic accelerators
•
Array processors
•
Graphics
•
Image processing
•
Geophysical processing
Applications within this market are slightly different from DSP applications within
other markets. Many of the applications in the computer market require
general-purpose arithmetic (numerical processing) in addition to more classical digital
signal processing functions. Historically, hardware implementations have relied
principally upon MPDSP products because of the higher speed and inherent architecture
flexibility allowed by these products. This market is now beginning to migrate toward
usage of second- and third-generation DSMPU products because of the higher integration
and lower cost of implementation.
10
© 1988 Dataquest Incorporated September
SIS DSP
0001325
DSP Markets and Applications
Table 6
DSP Revenue Forecast for the Computer Market
(Millions of Dollars)
Application
Cateqory
Actual
1986
1985
Acith. Accel.
Array Proc.
Image Proc.
Graphics
Geo Process
Other
1987
1989
$ 0.5
12.0
Forecast
1990
1991
1992
CAGR
1985-1987
CAGR
1988-1992
$ 70.0
26.0
44.0
29.0
N/C
41.4%
41.4%
21.1%
4.0
6.0
8.4
10.0
13.0
15.0
$ 2.0
15.0
12.0
18.0
0.1
5.3
0.1
7.3
0.2
9.1
0.5
0.7
1.0
1.3
12.0
18.0
27.0
43.0
30.9%
244.0%
21.3%
51.3%
16.9%
59.7%
47.0%
$135.0
$212.0
30.9%
47.0%
0
0
0
$ 5.0
$ 7.0
$10.0
3.0
9.0
0.1
4.3
1988
$10.0
18.0
20.0
21.0
Total Computer
OSP Revenue
$21.4
$26.4
$36.6
$45.5
$59.1
$87.6
Computer as Percent
of Total DSP
Revenue
10.3«
8.4%
8.2%
7.8%
7.6%
8.3%
$ 30.0
22.0
30.0
25.0
9.6%
0
11.4%
N/C > Not Computed
Notes Columns may not add to totals shown because of rounding.
Source:
Dataquest
September 1988
It is important to also understand that the applications shown above for this market
are not entirely orthogonal. In other words, arithmetic accelerators and array processors
are often used as OEM products within image processing or geophysical processing
stations. Dataquest has taken care to reduce the risk of double counting these products.
Certain parts of this market, primarily image processing and graphics, also use both
SFDSP and ASDSP devices because of the very high processing rates required. This trend
is expected to continue.
Consumer
As shown in Table 7, Dataquest expects the consumer market to be the single
largest market for DSP products by 1992. DSP technology is currently being developed
for inclusion in applications such as the following:
•
Audio compact disc players
•
Digital television
•
Electronic cameras
•
Toys
SIS DSP
0001325
© 1988 Dataquest Incorporated September
11
DSP Markets and Applications
Table 7
DSP Revenue Forecast for the Consiuner Market
(Millions of Dollars)
Application
Category
1985
Actual
1986
CO Players
Digital Vid
Elec Cameras
Toys
Other
$3.0
0
0
0.5
0.9
Total Consumer
OSP Revenue
Consumer as Percent
o£ Total OSP
Revenue
1forecast
1990
1991
CAGR
1988-1992
1992
CAGR
1985-1987
$ 45.0 $ 64.0 $ 86.0
80.0 130.0 190.0
1.0
10.0
4.0
7.0
4.0
5.0
73.0
33.0
51.0
$110.0
270.0
20.0
9.0
102.0
158.2%
N/C
N/C
144.9%
207.5%
38.4%
61.2%
185.7%
31.6%
53.7%
$163.0 $253.0 $366.0
$511.0
207.5%
53.7%
1987
1988
1989
$10.0
2.0
0
1.0
3.3
$20.0
10.0
0.1
3.0
8.3
$30.0
40.0
0.3
3.0
18.0
$4.4
$16.3
$41.4
$91.6
2.1«
5.2«
9.2%
15.6%
21.0%
24.0% 26.0%
27.4%
NC > Not Computed
Source:
Dataquest
September 1988
Digital-audio applications such as audio compact disc (CD) players provide an
example of the huge volume potential of DSP technology. Dataquest estimates that
nearly 10 million audio CD players will be shipped in 1988, with nearly half of them using
DSP technology. DSP is used for implementing digital interpolation filters in all CD
players that "oversample." Because of the high volume of this application, SFDSP
devices are used to optimize the architectiore and reduce costs as much as possible.
Dataquest estimates that nearly $30 million worth of digital filters will be shipped in
audio CD players in 1988.
A similar growth scenario holds true for digital television sets, which hold the
promise for improved video quality. DSP technology will be important in digital
television regardless of the direction of standards for future high definition (HDTV),
extended definition (EDTV), or improved definition (IDTV) television. As in audio CD
players, DSP-based digital TV chip sets will be designed as SFDSP products.
A third consumer growth area is for digital cameras which hold the promise of
"filmless" cameras for the future. A picture will be taken by a user in the same fashion
as today, but the image will be stored electronically in the camera. The photographer
will have the options of connecting the camera directly to his VCR for viewing, or having
a "hard copy" (photograph) generated with quality equivalent to that achievable using
today's cameras with film. Experimental cameras are available for newspaper
photographers today, although the technology is still a few years away for widespread
commercial use.
12
© 1988 Dataquest Incorporated September
SIS DSP
0001325
DSP Markets and Applications
Sophisticated toys will also provide a ready application area for DSP technology. In
fact, Dataquest believes the "Julie" doll introduced by Worlds of Wonder, represents the
single largest order ever received for a DSMPU product.
Automotive
As shown in Table 8, the automotive market has historically been the smallest user
of DSP technology. This trend is expected to continue through 1992. However, there are
a number of automotive applications in which DSP will begin to appear, including:
•
Antilock brakes
•
Distance sensors
•
Lane sensors
Table 8
DSP Revenue Forecast for the Automotive Market
(Millions of Dollars)
Application
Cateqory
Antilock Brakes
Distance Sens
Lane Sensors
Other
Total Automotive
DSP Revenue
Automotive as
Percent of
Total OSP Revenue
NC - Not Computed
1985
Actual
19B6
0
0
0
1992
CAGR
1985-1987
CAGR
1988-1992
$20.0
0
0
5.0
$40.0
1.0
1.0
10.5
N/C
N/C
N/C
N/C
151.5%
N/C
N/C
154.6%
N/C
154.6%
Forecast
1990
1991
1987
1988
1989
$0.1
0
0
0
$1.0
0
0
0.3
$3.0
0
0
0.8
$10.0
0
0
2.5
0
0
0
0
£
0
0
$0.1
$1.3
$3.8
$12.5
$25.0
$52.5
0
0
0
0.2%
0.5%
1.2%
1.8%
2.8%
•
Source:
SIS DSP
0001325
© 1988 Dataquest Incorporated September
Dataquest
September 1988
13
DSP Markets and Applications
Antilock brakes are becoming a standard feature on many cars today. Many of the
hardware implementations use DSMPUs. This application continues to represent a
significant growth opportunity.
The car that can drive itself has been dreamed about for many years now. DSP
technology promises to allow some of those features to be incorporated within the next
decade. Technology to implement distance sensors to detect moving objects (other cars)
and stationary objects (like telephone poles) is relatively straightforward, using simple
radar techniques. Lane sensors that track visual lines on the highway or some other form
of reference can also be built using existing DSP technology.
Certainly the technology to allow a car to drive itself down long, lonely stretches of
highway is here today. Adapting that technology to cars that can drive themselves
within an urban environment will require additional technology that is probably further in
the future. Certainly other DSP-intensive applications that have been discussed
previously will ultimately be important in cars of the future, such as navigation using
global positioning satellites and communications using cellular telephones.
14
© 1988 Dataquest Incorporated September
SIS DSP
0001325
DSP Market Strategies
FACTORS WORKING FOR DSP
The digital signal processing (DSP) market is reaping the benefits of the
long-term conversion from a purely mechanical world to a digital world—a
process that started in the early 1950s with the development of the first
digital computers. One by one over the years, as semiconductor technology
has improved, applications that previously could be implemented only in
analog circuitry have been converted to digital.
Analog first gave way to digital in non-real-time applications such as
accounting. Real-time applications required a speed and performance level
not available initially. Advances in semiconductor design and fabrication
are only just now making some of these real-time applications feasible
through technologies such as DSP.
Already, most low-frequency industrial control applications can be
totally implemented digitally. Inroads are also being made into the audio
frequency realm in areas such as telecommunications.
Eventually, DSP
products will replace most, if not all, analog circuitry in this frequency
range.
Even some of the simpler video frequency applications, such as
non-real-time enhancement of pictures from spacecraft, are being implemented
using DSP techniques. DSP implementations of more difficult applications,
such as real-time programmable image processing, are just around the corner
as a new generation of DSP processors comes off the drawing board. These new
DSPs are expected to be in production by the early 1990s.
Other factors are coming into play to enhance growth in the DSP market.
One of the major factors is the reduced cost of DSP products, particularly
digital signal microprocessors (DSMPUs). The price of DSP technology is no
longer prohibitive compared to the performance enhancements DSPs provide.
The new availability of DSMPUs will also contribute to the growth of the
DSP market as engineers become more familiar with DSP technology. DSMPUs
actually make new products possible, like low-cost, high-bit-rate dial-up
modems.
Perhaps the single largest factor working in favor of the DSP market is
the conversion of the worldwide telephone system from analog to digital
(ISDN).
This is a massive market. By 1990, Dataquest expects the U.S.
telecommunications market alone to consume approximately $45 billion of
equipment. Of this amount, $301 million is expected to be spent on the
purchase of DSP integrated circuits.
Furthermore, the military requirements for DSP technology will continue
to increase, as the quest for military superiority demands the development of
more "intelligent" weapons.
SIS DSP
© 1987 Dataquest Incorporated April
DSP Market Strategies
FACTORS WORKING AGAINST DSP
Several factors are expected to slow the growth of the DSP market over
the next five years. These include a lack of DSP engineers, a lack of
adequate development tools, and competing signal processing techniques.
Digital signal processing is new to the vast majority of analog engineers
who are used to operational amplifiers, capacitors, inductors, transformers,
diodes, and electrically induced noise and cross talk. They have solved
signal-processing problems by the same methods for years and are comfortable
with analog analysis techniques.
The new DSP techniques are very complicated, and understanding the
algorithms requires a degree of mathematical expertise that is not widespread
in the industry. Universities are just now creating DSP courses. Professors
who have had experience using DSP circuits are rare, and there are only a few
textbooks on the subject. It could be as long as four years before enough
engineers have graduated and are employed designing commercial products to
have any impact on the industry.
The lack of adequate development tools, particularly in-circuit emulators
and inexpensive design software, may severely hamper the growth of this new
industry. Today, the starting package cost to begin a DSP design is in the
range of $5,000 to $10,000. This provides an engineer with an evaluation
board, a simulator, an assembler, and an in-circuit emulator. The engineer
is expected to provide the personal computer. These support tools are also
somewhat primitive. Dataquest expects a retrenchment period in DSP market
growth as support tools come up to speed.
A third reason for slow acceptance of DSP is the development progress of
competing technologies.
Advances in analog circuit technology, switched
capacitors, surface acoustic wave filters, and optical filtering are impeding
the transition from traditional signal processing to digital signal
processing. Switched capacitors, although difficult to design and lacking
the precision and programmability of DSMPUs, are still more familiar and
cost-effective than DSP circuits and will continue to steal high-volxime
applications in the audio spectrum.
DSP MARKET STRATEGIES .
The DSP market is in a rapid transition phase, and the relative positions
of suppliers to this market are highly volatile. Early DSP microprocessors
such as Intel's 2920 and AMI' s 28S211, have disappeared. TRW, which used to
dominate the multiplier market, has lost most of its market share to an
oligarchy of CMOS vendors.
© 1987 Dataquest Incorporated April
SIS DSP
DSP Market Strategies
New entrants are announcing products every day, and the market is undergoing rapid technological change both in architecture and in device speed.
Customers are still willing to buy a new manufacturer's DSP if its specifications look good, rather than seeking the safety of a brand name. Suppliers
can quickly establish themselves as major market participants with just one
revolutionary product. In addition, few defined second-sourcing relationships exist today. This leaves open the possibility of powerful market
alliances that could radically change the face of the DSP market. Such
turmoil presents fertile ground for clever strategic planners.
Foreign Alliances
Some of the most advanced DSP parts are coming from Europe and Japan.
Europe is particularly strong in telecommunications, and Japan is strong in
video technology, machine vision, and pattern recognition. Given that the
television industry exists almost exclusively in the Far East, it is likely
that Japan will continue to lead the world in video technology.
These advanced video and telecom chips are needed by the U.S. military
for new weapon systems, and military applications represent 30 to 40 percent
of the total DSP market in the United States. Foreign producers cannot
effectively sell into this market, however, as the military cannot rely on
parts that may not be procurable during conflict or during times of trade
restrictions. Therefore, the potential exists for alliances between U.S.
semiconductor manufacturers and foreign manufacturers to effectively tap this
very lucrative market.
The Fast Microprocessor
Because of the speed they offer, DSP microprocessors are used in many
applications besides signal processing. Speed is exactly the same reason
engineers have been using bit slices for the past 10 years. DSMPUs are now
being designed into some of the same applications previously implemented with
bit slices, but this may be overkill. The bulk of these applications using
DSMPUs for speed could probably do without the multiplier. A stripped-down
version of a Harvard-architecture DSMPU would offer an attractive alternative
for these types of applications.
The market potential for a fast microcomputer that uses the architectural
tricks of a DSMPU is huge. Dataquest estimates that the total microcomputer
market will be about $4 billion in 1990. If only 10 percent of this market
chooses a Harvard-architecture pipelined reduced instruction set computing
(RISC) machine, this would present a $400 million opportunity.
SIS DSP
© 1987 Dataquest Incorporated April
DSP Market Strategies
Product Families
No I/O standard currently exists for DSP microprocessors and filters. In
most cases, glue chips are required for interfacing with the real world.
These glue chips can be as simple as SSI packages between a DSMPU and its
host microprocessor, or they can range from filters and amplifiers for A/D
conversion to expensive dual-port memories.
No families of parts in today's DSP market have uniform part numbers and
interfaces.
A typical design combines some parts from a microprocessor
supplier, some from an analog supplier, and some from a memory supplier.
There is no one-stop shopping.
A standard DSP product family has yet to emerge. Any company that tries
to set the standard needs to display a commitment to a line of parts, to
provide information to customers on the next generation planned, and to be
able to convince the engineering community that its interface architecture
will endure. Texas Instruments is the company that is coming the closest to
this goal.
The DSP Solution Company
DSP applications, because of the need for high speed, are more economical
on one chip than spread over many chips. The DSP microprocessor offers a
one-chip solution, but it is limited because application customization may
only be done via software algorithms.
Hardware solutions can significantly outperform software mappings.
Custom silicon hardware takes longer from product conception to market,
however, and hardware limits the functional complexity.
For the past
10 years, engineers have been using a combination of bit slices and software
mappings.
When more processing power was needed, additional bit-slice
components were added. Control of both the hardware architecture and the
software architecture is invaluable to the designer for the best trade-off
between flexibility and cost.
Equivalent bit-slice circuits are now available as standard cells in
semicustom implementations. One possible strategy using these cells would be
to target the DSP market specifically with DSP functional building blocks
plus DSP simulation and analysis software. This may prove to be a good niche
in the crowded field of ASIC suppliers.
The revenue from gate arrays and standard cells used strictly for
DSP applications will be about $200 million in 1990. An ASIC supplier with
DSP standard cells and DSP software support would gain a sizable share of
this niche.
© 1987 Dataquest Incorporated April
SIS DSP
DSP Market Strategies
Preprogrammed DSP Microprocessors
DSP microprocessors are difficult and expensive to program. First, there
is the initial $10,000 price tag for software, evaluation boards, and
emulators. Then, there is the cost of programming time. Programming a DSMPU
costs $10 to $30 per byte, fully supported. At 3,000 source code lines per
year, 2 bytes per source code line on the average, costs could run as high as
$180,000 a year.
Some customers will choose to buy a preprogrammed chip instead of making
the investment required to program their own. We expect this to be the case
especially with applications using proprietary algorithms that would
ordinarily involve considerable software design and development time.
Consequently, we believe that there will be a large market for preprogrammed
chips.
The functions of these chips will mirror the DSP market.
For
instance, these chips might involve speech-recognition algorithms and
high-speed modem standards.
Dataquest estimates that 15 percent of the forecasted DSMPU revenue will
be from preprogrammed products, representing approximately $40 million in
revenue in 1990. The development of these chips .can be done most successfully by companies that are experts in a particular vertical market. Any
company focused on a vertical market is best in command of the proprietary
algorithms, specifications, and key features required by that market.
SIS DSP
© 1987 Dataquest Incorporated April
Family Tree and Definitions
DSP Family Tree
OVERVIEW
The digital signal processing (DSP) semiconductor market is characterized by a
variety of different products which address a diverse set of markets and applications.
Dataquest partitions DSP products into four categories, as shown in Figure 1. Each
product category is quite distinct from the others in terras of functionality,
performance, and end application.
Figure 1
Segmentation of DSP Products
D S P Products
DSP
Microprocessors
(DSMPU)
General-Purpose
Special-Purpose
Mlcroprogrammable
DSP(MPDSP)
Multipliers (Mpy)
l^ilultlpllerAccumulators
(MACS)
Mlcroprograrrwnable
Arithmetic
Units (MAUs)
Special-Function
DSP (SFDSP)
Modems
Codecs
Speech Processors
Digital Television Circuits
Digital Filters
Fast Fourier Transform
Chips
AopllcatlonSpeciflc
DSP (ASDSP)
Cell-Based ICs
Gate Arrays
Source: Dataquest
September 1988
SIS DSP
0001216
© 1988 Dataquest Incorporated September
DSP Family Tree
DSP MICROPROCESSORS
DSP microprocessors (DSMPUs) are general-purpose, programmable devices similar
in many ways to conventional microprocessors. Their distinction is characterized by
clever architectural modifications which make them efficient for implementing the
repetitive multiplications and additions required by DSP algorithms. DSMPUs are
available as standard catalog products from manufacturers.
In some cases, manufacturers will mask-program a general-purpose device and
resell it as a function-specific device such as a modem. In this case, the product will not
be counted as a DSP microprocessor, but instead as a special-function (i.e., modem) DSP
device. The reasons behind this are:
•
Often the end user is unaware that the device is really a DSP microprocessor
because it is sometimes put into a different package with a different pin
configuration.
•
When the product is resold it fits the definition of a special-function device.
It is useful to group all special-function devices into the same category
regardless of whether a general-purpose or custom architecture was designed
to implement the function.
•
An analysis is also provided to obtain the importance of DSMPU architectures
to the special-function DSP market.
MICROPROGRAMMABLE DSPs
The category of microprogrammable DSP (MPDSP) products encompasses the
traditional "bit-slice" or "building-block" components, such as
multipliers,
multiplier-accumulators (MACs), and arithmetic-logic units (ALUs). These products are
designed to allow high-performance, modular DSP architectures to be designed using
standard off-the-shelf components. Dataquest uses the term "microprogrammable" to
describe these products instead of "bit-slice" or "building block". These latter terms are
misleading and are not truly descriptive of the applications for these devices.
Dataquest believes that it is important to analyze the microprogrammable market
separately from the mainstream DSP market. Many of the products in this category are
used in applications such as graphics and controllers, which do not specifically fit the
definition of digital signal processing. In order to keep this category distinct, only
products traditionally considered as arithmetic building blocks are included in the
revenue forecasts. These range from older 4-bit ALUs and 8-bit multipliers to the
newer 32- and 64-bit multipliers and ALUs.
© 1988 Dataquest Incorporated September
SIS DSP
0001216
DSP Family Tree
Systems designed using MPDSP products also require additional support circuitry
such as sequencers, FIFOs, registers, and static memories. These peripheral products are
not specifically DSP products, and will not be counted in the DSP revenue forecasts.
Additionally, they are generally counted in other semiconductor services within
Dataquest. However, because of their applicability to DSP systems, forecasts will be
provided for these products in cooperation with the other appropriate Dataquest
semiconductor service sections.
SPECIAL-FUNCTION DSPs
Special-function DSPs (SFDSPs) include products that are built using DSP techniques
and architectures, but which are designed for specific functions. Examples of these
include modems, codecs, speech processors, digital television circuits, digital filters, and
fast Fourier transform chips. DSP technology is inherent to all of these devices, but
they are not designed to be general-purpose in nature. Generally these devices cannot
be programmed by users to perform operations other than their defined functions; i.e., a
DSP modem cannot be programmed to do fast Fourier transforms.
Occasionally, conflicts may arise regarding the category into which some products
fit. SFDSP products retain the following characteristics:
•
They are designed to implement specific functions.
•
Their architecture implementation uses DSP techniques.
•
The product is not reprogrammable by the user to implement functions other
than those intended.
•
They are available as standard catalog items.
APPLICATION-SPECIFIC DSPs
Products in the application-specific DSP (ASDSP) category are custom devices
designed using primarily gate-array or cell-based IC techniques. In order to avoid
confusion with some products which may be considered SFDSP, ASDSP products retain
the following characteristics:
•
All products are targeted at DSP applications.
•
Products are not available as standard catalog items, but are instead
proprietary to the user. If the products do become available as standard
catalog items, they will be moved to the most appropriate of the other DSP
product categories.
•
Full-custom devices are also included in the ASDSP category as long as they
meet the above two criteria.
SIS DSP
0001216
© 1988 Dataquest Incorporated September
3:
DSP Definitions
A/D (analog to digital). A circuit that transforms a linear (analog) signal to a digital
representation. The digital representation is usually in the binary format of Is and Os.
analog. A circuit or system in which the output signals bear a continuous relationship to
the input signals, as opposed to a digital circuit.
ASIC (application-specific integrated circuit). An integrated circuit designed or adapted
for a specific application.
ASP (average selling price).
bit slice. A section or slice of an ALU, typically 4 bits wide.
CAGR (compound annual growth rate).
CBIC (cell-based integrated circuit). ASIC design technique utilizing nonfixed width or
height megacells.
CODEC (coder/decoder circuit). An integrated circuit that codes a voice signal into a
binary waveform or decodes a binary waveform into a voice signal. Such circuits are
now used in digital communications applications.
D/A (digital to analog). A circuit that transforms a digital representation of a waveform
to a linear (analog) waveform.
digital. A circuit or system whose values or levels are binary.
digital signal processing (DSP). The manipulation of the digital representation of an
analog waveform, the digital data being obtained by sampling the analog waveform often
and converting the sampled data to digital via an A/D converter.
DSMPU (digital signal microprocessing unit). A single-chip microprocessor that
performs the special vector and array manipulations necessary for digital signal
processing.
DSP building blocks. High-speed arithmetic "pieces," such as bit slices, multipliers,
multiplier-accumulators, and registers, which can be combined to form a high-speed
computing system capable of performing the special vector and array manipulations
necessary for digital signal processing.
FACT. A Fairchild Semiconductor
Technology.
trademark denoting Fairchild Advanced CMOS
FAST. A Fairchild Semiconductor trademark denoting Fairchild Advanced Schottky TTL.
SIS DSP
0001732
© 1988 Dataquest Incorporated November
DSP Definitions
FFT (fast Fourier transform).
mflops (million floating-^x>int operations per second).
GaAs (gallium arsenide).
gate array. An IC consisting of a structured pattern of logic devices that are customized
to meet each customer's requirements.
IC (integrated circuit).
linear. A semiconductor circuit whose output varies directly with the input.
analog.)
(See
SFIC (special function DSP integrated circuit). A category of DSP chips performing a
specific function (i.e., not general-purpose DSMPUs) such as speech sjmthesis.
© 1988 Dataquest Incorporated November
SIS DSP
0001732
w
DSMPU—Executive Summary
SUMMARY
The general-purpose digital signal processing microprocessor (DSMPU) is the fastest
growing of the four categories of DSP products. Fueled by new applications,
replacement of older analog technology, and the simplicity of reprogrammability,
revenue for DSMPU products should continue to grow significantly over the next five
years. To date, most DSMPU revenue comes from the more established base of 16-bit
integer processors. However, Dataquest sees a significant trend towards 32-bit
floating-point processors for next generation designs.
DSMPU REVENUE FORECASTS
The following should be noted about the size of the DSMPU market:
•
DSMPU revenue was $98 million in 1987, a 58 percent increase over 1986.
This number is on target with Dataquest's projection made in June 1987.
•
Worldwide DSMPU revenue is forecasted to be $147 million in 1988, up
50 percent over 1987.
•
CAGR should continue through 1992 at nearly 48 percent, reaching annual
revenue of $695 million.
•
DSMPU products in 1987 represented 22 percent of total DSP revenue. By
1992, Dataquest expects DSMPU products to represent over 37 percent of
total DSP revenue.
DSMPU APPLICATIONS
The following are key points regarding applications:
•
Communication applications currently represent the largest market
DSMPU products, accounting for nearly 70 percent of all DSMPU revenue.
•
The two largest-volume DSMPU applications are currently modems and DTMF
receivers. Of the $98 million in 1987 DSMPU revenue, about $60 million went
into these two applications.
•
The military and industrial markets represent the largest growth opportunities
for DSMPU products through 1992.
SIS DSP
0001217
© 1988 Dataquest Incorporated September
for
DSMPU—Executive Summary
DSMPU COMPETITIVE ENVIRONMENT
The following are key points regarding the competitive environment of the DSMPU
market:
•
There are currently 14 different manufacturers of DSMPU products.
•
The three leading DSMPU manufacturers (worldwide) continue to be Texas
Instruments, NEC, and Fujitsu, respectively, totalling 72 percent of DSMPU
revenue. Individual market share estimates for these three manufacturers are
shown in Figure 1
Figure 1
Worldwide Market Share Estimates—
Three Largest DSMPU Manufacturers
Fuiltsu
'
12%
Others
28%
\
\
NEC
15%
Texas tnstruments
/
45%
Total = $98 Million
Source: Dataquest
September 1988
© 1988 Dataquest.Incorporated September
SIS DSP
0001217
DSMPU—Executive Summary
Dataquest expects a DSMPU vendor shakeout over the next several years,
similar to the shakeout that occurred in the microprocessor marketplace (for
similar reasons) in the late 1970s and early 1980s. National Semiconductor
made a visible exit from the DSMPU business in 1987.
DSP system design issues, such as availability of development tools and
application support from IC manufacturers, are becoming as important to the
DSMPU selection process as which architecture is "optimal" for the
application.
Because of the large investment and time required to develop and support new
DSMPU architectures, coupled with the expected shakeout in the next few
years, Dataquest expects that it will be difficult for new manufacturers to
enter this market successfully.
DSMPU TRENDS
Products
The following are key points regarding products:
•
DSMPU products were essentially the first RISC processors, characterized
predominantly by single-cycle instructions. Most of the devices have at least
some elements of Harvard architectures, allowing simultaneous instruction and
data fetch, overlapped with instruction execution. This trend continues to
evolve with new-generation designs.
•
Sixteen-bit integer DSMPU products currently represent more than 90 percent
of DSMPU revenue in comparison with other arithmetic formats.
•
Dataquest expects floating-point DSMPUs to begin winning significant
numbers of design slots across portions of the military, industrial, computer,
and communication markets. We expect floating-point DSMPUs to achieve
nearly a 35 percent share of the market by 1992.
Technology
The following are key points regarding technology:
•
A trend is beginning to evolve whereby DSMPU manufacturers are migrating
the architectural "cores" of existing first- and second-generation DSMPU
products into cell libraries.
•
Dataquest believes that CMOS will continue to be the dominant process
technology for DSMPU designs.
SIS DSP
0001217
© 1988 Dataquest Incorporated September
DSMPU—Executive Summary
Pricing
The following are key points regarding pricing:
•
First-generation DSMPUs in the class of the TMS32010 (Texas Instruments),
7720 (NEC), and 8764 (Fujitsu) are currently selling in volume from $5 to $10.
Dataquest does not expect these prices to decrease significantly below $5.
•
Second-generation products such as the TMS320C25 (Texas Instruments),
MC56000 (Motorola), and ADSP-2100 (Analog Devices) were introduced with
low-quantity prices of $200 to $500.
Volume prices on second-generation processors are now below $50.
Dataquest expects these prices to be in the $10 to $25 range within
two years.
•
Third-generation floating-point DSMPUs are being introduced at sample prices
of $500 to $1000.
© 1988 Dataquest Incorporated September
SIS DSP
0001217
DSMPU—Forecast
INTRODUCTION
Digital Signal Processing (DSP) microprocessors (DSMPUs) are without question the
fastest-growing category of all DSP products. Other sections within this report will deal
specifically with the issues of why DSMPUs are growing so rapidly. This section will
concentrate specifically on the DSMPU unit and revenue forecasts and the assumptions
that were used in creating the forecasts.
DSMPU PRODUCT SEGMENTATION
Dataquest segments DSMPU products into three different categories as defined
below. While it is difficult to create categories that all products fit into perfectly,
general characteristics for the processors are defined in each section.
First-Generation DSMPU Products
Characteristics of first-generation DSMPU products are:
Generally 16-bit integer processors
Instruction execution times of greater than or equal to 200 nsec
Single instruction, single data (SISD) von Neumann architectures
Original products designed in NMOS technology (some have since been
redesigned in CMOS)
Volume prices generally $5-$ 15
Representative products:
TMS3201X (Texas Instruments)
UPD7720 (NEC)
Second-Generation DSMPU Products
Characteristics of second-generation DSMPU products are:
•
Generally 16-bit integer processors; some other arithmetic formats also exist
such as block floating-point, 24-bit integer, 22-, 24-, and 32-bit floating point
•
Instruction execution times of 80-200 nsec
SIS DSP
0001326
© 1988 Dataquest Incorporated September
DSMPU—Forecast
•
Adoption of (modified) Harvard and/or RISC architecture characteristics
•
More efficient architectures for interfacing with external memory
•
Volume prices generally $15-$40
•
Representative products:
TMS320C25 (Texas Instruments)
ADSP-2100 (Analog Devices)
DSP56000 (Motorola)
WE DSP-16/32 (AT&T)
Third-Generation DSMPU Products
Characteristics of third-generation DSMPU products are:
•
Generally 32-bit floating-point processors
•
Instruction execution times of 50-100 nsec
•
Harvard and/or RISC architecture characteristics
•
Large (greater than 24-bit) external data address space
•
Efficient subroutine and looping architectures
•
Sample prices generally $500-$1300 (no production yet)
•
Representative products:
TMS320C30 (Texas Instruments)
DSP96000 (Motorola)
WE DSP-32C (AT&T)
ZR34325 (Zoran)
© 1988 Dataquest Incorporated September
SIS DSP
0001326
DSMPU—Forecast
REVENUE AND UNIT FORECAST
Revenue
Table 1 and Figure 1 show Dataquest's forecast for the DSMPU market through
1992. The following bullets summarize highlights of the figures from the table:
•
All segments of the market should experience revenue growth rates exceeding
30 percent through 1992.
•
Revenue for the entire DSMPU market through 1992 should experience a
CAGR of nearly 50 percent.
•
DSMPU products as a percentage of total DSP revenue should reach over
37 percent in 1992, up from about 16 percent in 1985.
Of special significance is the expected revenue increase for 3rd-generation
DSMPUs, This market is expected to grow at an enormous CAGR of about 285 percent
through 1992, at which point revenue is expected to be larger than that of Ist-generation
DSMPUs. As can be observed in later tables, this is partly attributable to the much
higher average selling prices (ASPs) expected for these devices, compared with
Ist-generation DSMPUs.
Table 2 shows Dataquest's forecast for DSMPU unit shipments through 1992. The
following bullets summarize highlights of the figures from the table:
•
Dataquest expects total DSMPU unit shipments to increase at an impressive
CAGR of over 77 percent per year through 1992.
•
Unit shipments for both 2nd- and 3rd-generation DSMPUs are expected to
grow at rates significantly above that of the rate for the aggregate DSMPU
market.
•
Unit volume for 2nd-generation DSMPUs will probably not exceed volume for
Ist-generation DSMPUs until 1992, even though revenue for 2nd-generation
products should be higher in 1988.
Notice from both Tables 1 and 2 that the CAGR for unit shipments through 1992
(76.8%) is much higher than the CAGR for revenue growth through the same period
(47.5%). This is attributable to the significant decrease in ASPs for all DSMPU products
as shown in Table 3. Third-generation DSMPUs are expected to have much higher ASPs,
but will not be able to keep overall DSMPU ASPs high in the near term because of the
relatively small unit penetration of these devices in their first few years of production.
Figure 2 shows the percentage of market penetration by all three generations of
devices.
SIS DSP
0001326
© 1988 Dataquest Incorporated September
DSMPU—Forecast
Table 1
Worldwide DSMPU Revenue Forecast
(Millions of Dollars)
Forecast
1990
CAGR
1985
ictual
1986
1987
1988
1989
$ 34
0
$ 45
17
$ 51
47
$ 60
86
$ 83
135
$
105
191
24
$
148
250
87
$
176
310
209
22.0%
N/C
N/C
30.8%
37.8%
285.1%
S 34
i 62
$ 98
$147
$221
$
320
$
485
$
695
69.8%
47.5%
Total DSP
Products Revenue
$208
$316
$448
$586
$778
$1,057
$1,866
46.8%
33.6%
DSMPU as Percent
(\) of Total
DSP Revenue
16.3%
19.7%
21.9%
25.1%
28.4%
Ist-Generation
2nd-Generatioa
3rd-Geiieration
Total DSMPU
Revenue
30.3%
1991
$1,410
34.4%
;995-1997
1992
CAGS
1988-1992
37.2%
N/C s Hot Computed
Source:
Dataquest
September 1988
Figure 1
Revenue Growth for DSMPU Products
Revenue in Millions of Dollars
800
720-
T^^
640
3rd Generation
2nd Generation
1st Generation
mm
560
480400320
240
160
80
0
_E:
1985
1986
1987
1988
1989
1990
1991
1992
Source: Dataquest
September 19S8
© 1988 Dataquest Incorporated September
SIS DSP
0001326
DSMPU—Forecast
Table 2
Estimated DSMPU Unit Shipments
(Millions of Units)
1985
Ist-Generation
2nd-Generation
3rd-Generation
Total DSMPU
Units
;ictuBi
1986
1987
1988
1989
5.5
2.5
0.002
9.2
6.1
0.01
7.9
15.4
0.6
0
1.3
0.1
3.2
0.8
_a
_a
_a
0.6
1.4
4.0
PareEsat
1990
CAGa
1985-1987
CAGR
1988-1992
1991
1992
15.0
11.9
0.14
24.7
22.7
0.87
35.2
38.7
3.48
136.2%
H/C
N/C
59.4%
99.3%
584.8%
27.1
48.3
77.4
164.1%
75.8%
Mote: Columns may not add to totals shown because of rounding.
H/C - Hot Computed
Source:
Dataquest
September 1988
Table 3
Estimated DSMPU Average Selling Price
(Dollars)
1985
Actual
1986
1987
1988
1989
Forgcaat
1990
1991
1992
CfcCS
1985-1987
CAGE
1988-1992
Ist-Generation
2nd-Generation
3 rd-Generation
$60
0
0
$ 35
$150
0
$16
$60
0
$ 11
$ 35
$600
$ 9
$ 22
$300
$ 7
$ 16
$180
$ 6
$ 11
$100
$ 5
$ 8
$60
(48.4%)
H/C
H/C
(17.9%)
(30.9%)
(43.8%)
ASP
$60
$ 44
$25
$ 19
$ 14
$ 12
$ 10
$ 9
(35.7%)
(16.6%)
H/C = Hot Computed
Source:
SIS DSP
0001326
© 1988 Dataquest Incorporated September
Dataquest
September 1988
DSMPU—Forecast
Figure 2
Percentage Penetration of 1st-, 2nd-, and 3rd-Generation DSMPUs
to Total DSMPU Market
1988
3rd Generation
0.002 Million Units
0.02%
Total DSMPU Units = 8 Million
Source: Dataquest
September 1988
1992
3rd Generation
3.5 Million Units
4.5%
Total DSMPU Units = 77.4 Million
Source: Dataquest
September 1988
© 1988 Dataquest Incorporated September
SIS DSP
0001326
Product Analysis
DSMPU—Product Comparison
ARCHITECTURES
A digital si^ml processing microprocessor (DSMPU) is a single-chip microcomputer
designed for fast execution of signal jwocessing algorithms. The DSMPU contains its own
arithmetic and logic units, multiplier, local data storage registers, instruction sequencer,
and instruction decoder.
There are three styles of DSMPUs being used today:
•
General-purpose DSP microprocessor
•
Parallel-array microprocessor
•
Data-flow microprocessor
Genra^l-Purpose DSMPUs
Most current DSMPU designs use a variation of the Harvard architecture. A
Harvard architectiire differs from the classical von Neumann architecture in that the
program and data memories lie in two separate spaces, permitting full overlap of the
instruction fetch and execution. The von Neumann model, i^ed in standard
microprocessors, relies on a step-by-step sequence of fetch and execute cycles. The
Harvard modification allows data and program accesses to execute in parallel, rather
than simultaneously. Despite the implied parallelism, the Harvard design is still
considered a single- instruction, single-data-path (SISD) device. Figure 1 illustrates a
typical Harvard architecture used in DSMPUs.
The Harvard architectural differences have one punxjse—speed. A typical Harvard
architecture DSMPU can calculate complex vector and matrix mathematical functions
approximately 10 times faster than a standard microprocessor,
These DSMPUs are harder to program than the standard microprocessors, however.
Although their instruction sets are very similar to those of standard microprocessors, the
implied parallelism of these DSMPUs substantially increases the burden placed on the
programmer.
Parallel-Array Micn^rocessors
An9ther type of DSMPU architecture is the parallel-array microprocessor. Parallel
processor architectures fall into two categories:
•
Single Instruction Multiple Data (SIMD)
•
Multiple Instruction Multiple Data (MIMD)
SIS DSP
0001733
© 1988 Dataquest Incorporated November
DSMPU—Product Comparison
Figure 1
Typical Harvard Architecture
Program RAM
1
]
Proflram Counter
1
• ^
Program Bus
Data Bus
Data Bus
Source:
0001733-1
© 1988 Dataquest Incorporated November
Dataquest
November 1988
SIS DSP
0001733
DSMPU--Product Comparison
The systolic-array microprocessor, shown in Figure 2, is an example of an SIMD
machine. A systolic-array microprocessor is a matrix of processors, each with its own
local memory and communication links to neighboring processors. Each processor
calculates its data in parallel with the others. The advantage of this architecture is the
amount of parallel processing that is achieved. If there are N processors, then the
number of instructions being executed is N times that of a single processor.
Figure 2
Systolic-Array Microprocessor
To Rest of Grid
To Rest of Grid
Source: Dataquest
November 1988
0001733-2
SIS DSP
0001733
© 1988 Dataquest Incorporated November
DSMPU—Product Comparison
NCR introduced the first commercial systolic chip, known as the geometric
arithmetic parallel processor (GAPP), in 1984. The GAPP consists of a 6 x 12 arrangement of bit serial processor cells, with each processor connected to its four nearest
neighbors.
The second systolic chip, the MSM6956, known as an adaptive-array processor, was
introduced by Oki Semiconductor earlier this year. The array grid comprises 8 x
8 single-bit processing elements, each communicating with its eight neighbors. This chip
has a hierarchical bypassing scheme that allows data to skip rows in the matrix if no
operation is to be performed. Both the MSM6956 and GAPP arrays may be expanded to
handle larger processing problems by linking identical processors together.
Another type of parallel processor to consider is the Transputer from Inmos. The
Transputer is a 32-bit microcomputer with its own local memory and with links to
connect one transputer to another in a multiprocessor (MIMD) mode. The transputer is
supported by its own communication language known as Occam.
Transputers may be linked together to form networks of programmable components
in any topology desired, from as simple as a binary tree to as complex as a hypercube.
Data-Flow MicroprocessOTS
The last major type of DSMPU is the data-flow microprocessor. Currently, the only
example of this type of architecture is the NEC uPD7281, known as an image-pipelined
processor. The uPD7281 employs a token-based data flow and pipelined architecture to
achieve a very high throughput rate. Although it is tailored for image processing, the
UPD7281 is considered a general-purpose DSP.
The token-based data-flow architecture used by the uPD7281 is very similar to that
used by token-based communication systems. Specially formatted input tokens are
downloaded from the host processor to the link and function tables of the uPD7281. The
contents of the link and function tables are closely related to a computational graph.
The computational process may be represented graphically by a directed data-flow
graph. In this type of graph, entries into the link table are represented by the arcs, and
entries into the function table are represented by the nodes. The arc between two nodes
has a data value and is known as a token. Each node signifies an operation.
A real-world analogy would be the automobile assembly line, with the assembly line
representing the data-flow path and the iiKiividual assembly stations being the functional
nodes. Identifying information (in the form of tokens) travels with each vehicle as it
moves down the line. This information details what type of equipment is to be installed
at each location. At each station, these "travelers" are consulted, the necessary
operation is performed, and the automobile is sent on down the line. This type of flow
enables automobiles to he produced at the rate of one per minute.
© 1988 Dataquest Incorporated November
SIS DSP
0001733
DSMPU—Emerging Technology and Trends
This section describes the major technical issues of today's digital signal processing
microprocessor (DSMPU) architectures. In it, we discuss the different approaches chip
makers are employing to solve digital signal processing (DSP) problems and offer pros
and cons for each.
INSTRUCTION SETS
General-Purpose versus Specikl-Purpose
A typical DSP system combines very high speed signal processing with low-speed
control fimctions. For example, in a modem, the low-speed functions include scanning of
the front panel buttons, checking for carrier presence, making decisions on whether to
retransmit a block of data, and dialing a telephone number. The program loops for these
functions execute from 10 to 1,000 times per second. DSP loops, on the other hand,
occur at rates of 25,000 to 1 million instruction executions per second. There is a wide
discrepancy between the processing ix)wer needed for low-speed "housekeeping"
functions and that needed for high-speed signal processing functions.
To perform housekeeping functions, a processor must have timers, interrupts, and an
easy-to-use instruction set, all of which impede fast DSP algorithms. Consequently,
many designs employ two processors—a DSMPU to handle the fast signal processing
algorithms and a general-purpose microcontroller for the housekeeping functions.
Separation of the housekeeping functions and the signal processing functions into
separate processors minimizes their interaction, consequently reducing system design
time and improving system reliability.
Dataquest believes that the two-processor solution will be the predominant system
design for the next 5 to 10 years. DSMPU users tend to favor a specific microcontroller,
the one for which they have development systems and programming experience. In
addition, because prices of the most popular microcontrollers—including the 8051, the
TMS7000, and the 6801—are approximately 20 percent of the price of a DSMPU, there is
no pressure to eliminate the general-purpose microcontroller.
Instruction Word Width
There are two major approaches to handling instruction word width. In the first, the
DSMPU architecture provides a wide program instruction width. For instance, the
Philips PCB5010 has a 40-bit instruction word. This wide instruction word is similar to
the microcoded instruction sets found in minicomputers. The instruction has multiple
fields, and each field controls a portion of the chip, such as the arithmetic logic unit
(ALU). One instruction executes a number of parallel operations, such as gating data
onto buses, setting up the ALU, multiplying, and conditionally branching to the next
instruction.
SIS DSP
0001734
© 1988 Dataquest Incorporated November
DSMPU—Emerging Technology and Trends
When using a wide instruction word, the programmer must refer to a block diagram
of the chip. Using the diagram, the programmer will decide which data items gate onto
the buses and which counters increment. While this process affords the most flexibility,
it also creates a very difficult programming task. The programmer must keep track of a
number of operations occurring in parallel while being aware of pipelining delays.
Some manufacturers, such as Texas Instruments, take a different approach to
instruction word width. They are using small instructions, typically 16 bits wide. The
chip designer preselects a number of the possible operations and makes these into
instructions. Each 16-bit instruction fans out through a decoder into the individual
control signals for the buses and the adders. This approach is considerably easier to
program. An assembly language programmer has no trouble learning to use this type of
DSP instruction set.
Data Word Width
A data path 16 bits wide, with its 96dB dynamic range, has proven to be adequate
for most audio and telecommunications applications. Approximately 70 percent of all
commercial DSMPUs have 16-bit data paths. ALUs, however, must be larger than
16 bits.
Mathematical operations induce noise in the signal being processed. To minimize
this noise, more bits are used. Toshiba engineers, for example, found that they needed a
32-bit machine to do a fast Fourier transform (FFT) on 16-bit input data. When a 16-bit
machine was used, round-off and truncation errors in the FFT built up and destroyed the
significance of the output. By using a 32-bit ALU, the Toshiba engineers were able to
keep 16 bits of significant data.
Approximately 25 percent of all DSMPU offerings have 32-bit word lengths. Most
of these chips are also floating-point machines. Dataquest believes that there will be a
long-term trend toward 32-bit floating-point DSMPUs. Until 1995, however, the
workhorse will continue to be the 16-bit fixed-point DSMPU.
INPUT/OUTPUT
DSMPU communication paths typically resemble the application problem. For
instance, the Inmos transputer has four serial channels for communication between
adjacent chips. As a result, it can be put into a two-dimensional topology in which there
is another transputer to the north, south, east, and west. Image-processing problems are
two-dimensional problems. Therefore, a typical application for the transputer is image
enhancement, where a two-dimensional picture is sliced into rectangles and each
transputer operates on one piece of the picture. In a system with 16 transputers, each
transputer would process one-sixteenth of the image.
© 1988 Dataquest Incorporated November
SIS DSP
0001734
DSMPU—Emerging Technology and Trends
Telecommunication problems, on the other hand, are linear. There are two
communication paths—one going in and one going out. A linear string of DSMPUs is,
therefore, the most typical configuration. A good example of a telecom-oriented
DSMPU is the Philips PCB5010. It has two full-duplex serial ports that facilitate
stringing the processors along the signal path, as shown in Figure 1.
Figure 1
Serially Connected DSMPUs in a Telecom Application
Serial In
Serial Out
Serial In
Serial Out
Telephone
Line Side
User
Side
Serial Out
Serial In
Serial Out
Serial In
0001734-1
Source: Dataquejt
November 1988
NUMBER OF INTERNAL BUSES
The Harvard architecture is the most widely used structure for DSMPUs. It has at
least two buses, one for instructions and one for data. An example of a two-bus DSMPU
is Texas Instruments' TMS320C25.
Because of the nature of DSP operations, two buses can limit the data transfer
rate. During the "multiply-accumulate" instruction, the instruction sequencer fetches
two data operands. One operand is the signal value, and the other is a filter coefficient.
Because the instruction and each data operand requires a bus, a two-bus DSMPU is
insufficient.
The solution in a two-bus DSMPU is to repeat the multiply-accumulate instruction.
During the instruction repeat, both buses transfer data. The instruction needs to be
fetched only once, after which it is held in a register while a counter counts the
repetitions. In this way, the buses are freed from multiple instruction fetches during a
repetitious operation.
SIS DSP
0001734
© 1988 Dataquest Incorporated November
DSMPU—Emerging Technology and Trends
The other solution, used in approximately half of all DSMPUs, is to separate the
instruction path from the data buses. An example is the Philips PCB5010. Because
instruction fetches do not interfere with data transfers, this type of DSMPU can fetch a
new instruction while simultaneously fetching two data operands.
Both of these solutions work well. At present, the number of buses may be pointed
out in the marketing literature of these products, but ultimately this is an unimportant
feature critical only to the chip designers themselves.
ARITHMETIC
Floating-Point versus Fixed-Point
In highly iterative algorithms, the round-off error problem can be reduced by using
floating-point arithmetic. At present, Dataquest estimates that only 5 percent of all
applications use floating-point arithmetic. Dataquest forecasts that this percentage will
rise to 25 percent by 1990 and to 50 percent by 1995 because of floating-point
arithmetic's ability to increase dynamic range, simplify DSP algorithms, and shorten
time to market.
Figure 2 contrasts the signal-to-noise ratio of a 22-bit floating-point multiplier
with a 22-bit fixed-point multiplier. A 22-bit fixed-point format guarantees a
96dB signal-to-noise ratio over the range of 96dB to 132dB. A 22-bit floating-point
format does not reach as high a maximum signal-to-noise ratio. It does, however,
maintain a 96dB signal-to-noise ratio from 96dB to 380dB. When the 36dB fixed-point
dynamic range is contrasted with the 284dB floating-point dynamic range, the difference
is substantial.
The greater dynamic range of floating-point arithmetic is an advantage in radar
signal processing. The strength of incoming radar signals may vary from a signal that
saturates the A/D converter to a signal that barely registers. Using a floating-point
format, the signal power level differences can be easily handled without losing the signal
to arithmetic overflows.
Not only does floating-point arithmetic have a greater dynamic range than
fixed-point, but it is also easier to use to implement filters. In DSP filter algorithms,
fixed coefficients are multiplied against the signal. These coefficients are often very
close to one or to zero. This is analogous to a high Q resonator in the analog world being
very close to oscillation. While a fixed-point coefficient has many ones or zeros after
the decimal point, the number of significant digits huddled over to the right is small.
The lower the number of significant digits used in a multiplication, the less significant
the result. Maintenance of significance using fixed-point arithmetic requires the use of
scaling tricks and careful selection of coefficients.
© 1988 Dataquest Incorporated November
SIS DSP
0001734
DSMPU—Emerging Technology and Trends
Figure 2
Effect of Fixed-Point versus Floating-Point Arithmetic
on Signal-to-Nosie Ratio
Signal-to-Noise Ratio (dB)
150
125 -
100 -
75 -
^y>,^.•^^yv'.•^•yy :"'>..
22-ait Fixed Point JC'-,:
50 -
25 -
25
50
75
100
125
150
175
200
223
Signal Strength (dB)
Source: Dataquest
November 1988
0001734-2
Typically, multidimensional adaptive filters require floating-point arithmetic. In
this type of computation, numbers may occasionally become very large or very small on
the way to a final value. Fixed-point arithmetic would overflow during computation
unless scaling of the data is done before arithmetic calculations are attempted.
Floating-Point Performance Trade-Offs
Floating-point units present some performance trade-offs in speed and in size to the
user who selects them to take advantage of the wide dynamic range already discussed.
Floating-point units are more complex and, consequently, larger than their fixed-point
counterparts. As shown in Figure 3, a floating-point multiplier requires an additional
adder and extra registers.
SIS DSP
0001734
© 1988 Dataquest Incorporated November
DSMPU—Emerging Technology and Trends
Floating-point units use more chip area. Floating-point processors have a larger die
size because of the additional circuit elements required to implement their superior
data-processing capabilities. The extra functional blocks shown in Figvire 3 actually
have a less substantial impact on overall chip size than the space taken up by the
additional RAM required to process the greater data load of floating-point operations.
Figure 3
Block Diagram of a Floating-Point Multiplier
Exponent
Mantissa
Exponent
Mantissa
Multiplier
Exponent
Mantissa
Source: Dataquest
November 1988
0001734-3
Floating-point chips are approximately 50 percent slower than fixed-point chips for
two reasons. The first reason is that extra operations are needed. A floating-point
"multiply" operation consists of a "multiply" followed by an additional "add" to process
the exponent. Similarly, a floating-point "add" operation requires an extra "shift"
operation followed by an extra "add." A second reason that floating-point chips are
slower is that they usually have longer data word lengths than their fixed-point
counterparts. A 32-bit "add" in a floating-point chip is going to take more time than a
16-bit fixed-point "add."
© 1988 Dataquest Incorporated November
SIS DSP
0001734
DSMPU—Emerging Technology and Trends
Additionally, in spite of its many advantages, floating-point arithmetic can be
inexact. The amount of noise added during arithmetic operations is not an absolute
percentage of the maximum input signal level. Instead, the noise added during
floating-point operations varies, depending on the values being added or multiplied.
Since the amount of noise being added varies from operation to operation, it is difficult
to mathematically predict the performance of a floating-point algorithm. In 16- and
22-bit formats, floating-point arithmetic requires as much care as fixed-point
arithmetic.
Smaller-size floating-point formats, such as the Hitachi format, can cause
signal-to-noise problems just like those caused by small fixed-point formats. Using
small word lengths requires detailed analysis, but floating-point arithmetic is more
complex than fixed-point arithmetic and, consequently, harder to analyze. Therefore,
most floating-point DSMPUs have at least a 32-bit data path. With the longer word
length, the noise added in each operation is small and its effect on the algorithm can be
neglected. Basically, applications needing 16 to 32 bits of precision can use fixed-point
DSMPUs; those needing 32 to 64 bits of precision must use floating-point DSPMUs.
Normalization
When floating-point numbers are normalized, they retain their significance. The
normalization operation shifts the data bits to eliminate zeros after the decimal point.
The result is a word in which all the bits are significant. Because floating-point
arithmetic automatically performs normalization, the resulting data are nicely packed
with significant data. The programmer is not required to check the magnitude of the
data and then scale it in order to pack the most significance into each word, as is
necessary in fixed-point arithmetic.
A floating-point DSMPU can lighten a programmer's burden by automatically
normalizing after each mathematical operation. However, this raises an architectural
issue as to whether the DSMPU should, in fact, do this postoperation normalization or
leave it up to the programmer. The main argument against automatic postoperation
normalization is the time added to each operation.
The NEC 77230 DSMPU does not do postoperation normalization. As a result, it is
faster than other floating-point DSMPUs that do normalize after each "multiply."
Because a filter calculation normally requires only one normalization, the NEC 77230
can eliminate automatic postoperation normalization by having an expanded 55-bit ALU
to accumulate the intermediate results. The wider format eliminates round-off and
truncation errors. A single "normalize" instruction is sufficient to return the data to a
32-bit floating-point format after as many as 32 filter taps.
Dataquest expects fixed-point DSMPUs to maintain their majority share of DSMPU
sales through 1995. Their faster speed and lower cost will facilitate design-ins into
high-volume applications, such as consumer electronics. Dataquest anticipates the
inherent advantages of floating-point DSMPUs to open up new market opportunities,
such as complex modulation communication techniques in telecommunications.
SIS DSP
0001734
© 1988 Dataquest Incorporated November
DSMPU—Emerging Technology and Trends
Standards for Floating-Point Formats
Floating-point formats have been a controversial issue since the early mainframe
days and the minicomputer days. The initial controversy stemmed from the fact that
each manufacturer used its own format, the result being that data could not be
interchanged easily between computers. The controversy over floating-point format
continues into the DSP arena today. Standardization of formats is perhaps more
important for DSMPUs than it was for large computers. With DSMPUs, data must be
transferable from chip to chip without intervening translators that would increase cost
and slow operation.
There are several floating-point formats competing to become the standard. NEC
has a 32-bit two's-complement format. TRW has offered products with a 22-bit
floating-point format, and Hitachi introduced a 16-bit format in the HD61810. Analog
Devices and AT&T are using the IEEE 32-bit format.
The main argument against using the IEEE 32-bit format is that it is difficult to
implement. The IEEE format contains features to please everyone. Signal processing
does not need some of the features intended primarily for scientific data processing,
such as "-" infinity, "+" infinity, and seminormalized data formats. These extra features
add cost and slow down operations. Since the main design goals of DSMPUs are high
speed and low cost, many companies are choosing to eliminate unneeded features and use
simpler floating-point formats for DSP.
For computational efficiency, NEC chose a two's-complement 32-bit floating-point
format for its 77230. NEC acknowledged the IEEE format for data interchange,
however, by including one instruction to convert between their two's-complement
format and the IEEE 32-bit standard. As a result, data computed by the 77230 can be
prepared for output with a single instruction.
TRW is championing a 22-bit format in its DSP building block product line. TRW's
format has its advantages: It has a lower cost than the IEEE format, and it operates at a
higher speed. It is TRW's contention that 22 bits is more than sufficient to satisfy the
signal-to-noise requirements of typical applications, such as radar processing.
The difference in chip area between 22 bits and 32 bits is approximately one-third,
which amounts to nearly a factor of 2 in die cost. Packaging is also less costly with a
22-bit format. It is conceivable that IC cost in a 22-bit system might be as low as
one-half that of a 32-bit system.
Dataquest forecasts that both the 22-bit TRW format and the 32-bit IEEE format
will succeed in the marketplace. Worldwide DSMPU shipments are forecast to grow at a
42.3 percent compound annual growth rate (CAGR) from $62 million in 1986 to
$254 million in 1990. DSMPUs will gradually replace analog devices at higher and higher
bandwidths, as they are now replacing analog devices in the audio spectrum. It is only a
short time before DSP moves into an intermediate frequency spectrum below
1-MHz bandwidth. As DSP grows, there will be a greater opportunity for floating-point
formats that are sufficiently differentiated from one another to provide unique features
for specific applications.
S:
© 1988 Dataquest Incorporated November
SIS DSP
0001734
DSMPU—Emerging Technology and Trends
Bit-Serial Arithmetic
An internal implementation issue for DSP chips is whether to use parallel or
bit-serial arithmetic. In parallel arithmetic, which is used by every DSMPU except the
NCR GAPP processor, the entire word is added or multiplied at one time; consequently,
throughput is very fast. An add or multiply operation typically takes 100ns, which is a
rate of 10 MHz. In bit-serial arithmetic, each bit is processed sequentially, a relatively
slow process. For example, adding two 16-bit numbers serially would take 16 clocks, one
clock for each bit. If the clock frequency is 10 MHz, the total operation would require
1.6 microseconds, which is equivalent to a rate of 600 kHz.
Even though bit-serial arithmetic may be slower than the parallel arithmetic, it can
be used successfully in a fixed algorithm. Examples of fixed algorithms are those used in
hi-fi graphic analyzers and nonadaptive telephone line filters. In these applications, a
number of simplifications can be made to the algorithms, such as left-shifting a data
word to do a multiply-by-two operation. The left-shift and the multiply operations are
equivalent mathematically, but a shift takes much less circuitry than a multiplier.
By simplifying the algorithm to conform to operations that can be implemented
effectively by bit-serial hardware, it is possible to drastically reduce the amount of chip
area used. For instance, a hi-fi graphic equalizer implemented in bit-serial arithmetic
would be approximately 20 percent of the chip size of a parallel arithmetic version.
Historically, bit-serial machines have not fared well. Before the 8-bit microprocessor was available, many people made bit-serial processors out of standard TTL.
The popularity of 4-, 8-, and 16-bit parallel microprocessors attests to the fact that
bit-serial arithmetic is not seriously considered for standard microprocessor designs.
Dataquest does not expect bit-serial DSMPUs to be offered. Bit-serial algorithms
are difficult to change and, for substantial chip-area savings, multiplication coefficients
must be powers of two. These restrictions make bit-serial arithmetic unsuitable for
general-purpose DSMPUs. For special-function DSP circuits (SFICS), however, bit-serial
arithmetic may be a viable approach. If the algorithm can be transformed into a
bit-serial implementation and if the sampling rate bandwidth is low enough, then a
bit-serial architecture is worth considering. The potential 80 percent reduction in chip
area more than outweighs the difficulties.
SIS DSP
0001734
© 1988 Dataquest Incorporated November
m-
DSP Building Blocks—Product Comparison
OVERVIEW
Bit slices, multipliers, multiplier-accumulators, and filters are types of DSP building
blocks. They are an integration level below the DSP microprocessor (DSMPU) and are
used where performance or speed is needed. When using the building block approach, the
designer creates his own DSP processor. This is a very flexible approach because the
hardware can be selected to fit the processing task. For higher-performance
applications, a number of arithmetic units can be cascaded together. For
lower-performance applications, cost can be saved by multiplexing the use of one
arithmetic unit.
Additionally, building blocks are often available in faster technologies, such as
gallium arsenide (GaAs) or bipolar emitter coupled logic (ECL), than DSP
microprocessors. Thus, to achieve ECL speeds, the engineer is constrained to using
building blocks. However, as DSMPUs become available in a technology, they displace
building blocks, despite the higher performance available in building blocks.
DSP BUILDING BLOCKS
Bit Slices
Bit slices consist of an arithmetic logic unit (ALU), microsequencers, and
miscellaneous components, such as register sets. Since its introduction, the bit slice
approach has developed into a major segment of the building block IC market. There are
three major bit slice product lines: AMD's 2901, the Texas Instruments 74AS88 series,
and Fairchild's FAST line. Of these, AMD has the predominant share. AMD's original
2901 design is widely sourced, with several vendors having converted it into standard
cells.
The bit slice approach is midway between DSP microprocessors and custom designs
in flexibility and performance. Bit slices give the designer the ability to manipulate the
hardware topology of the algorithm as well as the microcode. But, with DSP
microprocessors increasing in speed and CAD tools making custom designs easier, is
there a place for bit slices?
Dataquest believes that 4-bit ALU slices will continue to evolve toward wider
microprogrammable products that are sliced on a functional level rather than the bit
level. These products, although noncascadable, offer more complex internal functions
and faster system speeds. Total hardware flexibility may be sacrificed, but perhaps
better software tools will be gained.
SIS DSP
0001735
© 1988 Dataquest Incorporated November
DSP Building Blocks—Product Comparison
Semicustom vendors offering high-speed microprogrammable cells will also
gradually erode the 4-bit ALU slice customer base by offering the desired hardware
flexibility combined with the speed advantages of a single piece of silicon, not to
mention the board space savings.
Multipliers
The DSP IC industry was born with the invention of the high-speed multiplier. In
the early 1970s, TRW's bipolar high-speed parallel multipliers were the clear standard.
Pin-compatible circuits abounded. However, TRW was not fast enough in switching to
CMOS, with its inherent power savings and cost advantages. Consequently, several small
companies with CMOS technologies were able to enter the market. There are now more
than 16 suppliers of multipliers, with CMOS products offered by companies such as
Analog Devices, IDT, Logic Devices, and Weitek.
The multiplier industry structure is quite flat and characterized by many suppliers,
small market shares, a lack of a dominant supplier, low margins, and low prices. There
have been many market entrants, and the market is currently oversupplied. Eventually,
multipliers should become identified with a major bit slice product line and will be
supplied as part of a chip set. An industry fallout should result in a readjustment of
market shares.
Multiplier-Accumulators
Since most DSP algorithms require a multiply followed by an ADD, another DSP
building block is the multiplier-accumulator. The block diagram of a multiplieraccumulator is shown in Figure 1.
The fastest multiplier-accumulators are bipolar ECL devices, such as those offered
by Bipolar Integrated Technology. A 16 by 16 multiply can be performed by CMOS
multiplier-accumulators in 60 nanoseconds, by CMOS/SOS in 30 nanoseconds, by TTL
bipolar in 25 nanoseconds, and by ECL bipolar in 8 nanoseconds.
The current multiplier-accumulator market is crowded; the number of participants
is large, at approximately 18, and market shares are low. No dominant supplier presently
exists.
Filters
The integration of the multiplier will continue to drive digital signal processing
products. For instance, the infant DSP filter industry was created by the ability to put
several multipliers on one chip. All high-quality filter applications are expected to
convert to digital, giving this market segment the highest compound annual growth rate
of any of the building block areas.
© 1988 Dataquest Incorporated November
SIS DSP
0001735
DSP Building Blocks—Product Comparison
Figure 1
Multiplier-Accumulator Block Diagram
Output = Accumulator + (Input A x Input B)
<l
Accumulator
2
Adder
^
J
^
Multiplier
7
1
I
Input A
0001735-1
SIS DSP
0001735
Source: Dataquest
November 1988
© 1988 Dataquest Incorporated November
Microprogrammable DSP Building Blocks—
Executive Summary
Building blocks are no longer the largest segment of the DSP IC market. Total
revenue is expected to decline modestly because of competing alternate solutions.
Although product lines are broad and the industry has many suppliers, product and
supplier successes are becoming increasingly rare.
To understand this unusual and dynamic situation, Dataquest has outlined three
primary objectives for use with the material accompanjdng these market estimates and
forecasts. The first objective is to establish clear product and market distinctions.
Building blocks are basic and serve broad markets, but now are more accurately seen as
specific products with narrow and only sometimes overlapping user markets. The second
objective is to understand the key product trends weighted by their market importance.
Numerous product directions currently exist, but some important user market trends may
eclipse these directives. The third objective is to clarify the real business opportunity.
The building block market is an easy-to-enter and highly visible market. Many
companies have to enter this market and their relative lack of success is not widely
understood.
FORECAST SUMMARY
Market Size
•
Actual DSP building block revenue in 1988 was $150 million.
•
The largest product revenue category in 1988, $85 million, was fixed-point
data path elements, but it is declining.
•
The largest market use in 1988 ($70 million) was in data
accelerators and coprocessors.
processing
Market Forecast
•
Expected building block revenue in 1993 is $149 million, representing a small
net decline.
•
In 1993 the largest product revenue category is expected to be floating-point
data path elements at $101 million, but it should experience slow growth.
•
The largest market use in 1993 will continue to be data processing
coprocessors and accelerators, at $100 million, but it is expected to
experience declining revenue.
SIS DSP
0004601
© 1989 Dataquest Incorporated August
Microprogrammable DSP Building Blocks—
Executive Summary
PRODUCT ANALYSIS SUMMARY
•
DSP building blocks can be clearly distinguished from processors and
special-function devices, even though the products exist in both processor
families and functional product lines.
A large overlap exists in the use of DSP and data processing building blocks.
The major product distinctions are in function, data format, and process and
interface technology.
Major building block processor family suppliers are AMD, Analog Devices, TI,
and Weitek.
Significant building block functional product line suppliers are BIT, IDT, Logic
Devices, and TRW.
Software development is a major problem for microcoded building block
processors, because only primitive tools can be supplied for general use.
MARKET ANALYSIS SUMMARY
Building blocks are used for architectural flexibility
high-performance needs not met by other standard products.
in
meeting
Single usage is at low to moderate levels, otherwise an ASIC investment would
be justified.
In general-purpose data processing, the major uses in order of size are
coprocessors, controllers, and central processing units (CPUs).
The largest use of coprocessors is for PCs and high-performance workstations.
In DSP, the largest user market is for special-purpose processors for radar,
sonar, imaging, and graphics.
PRODUCT, MARKET, AND BUSINESS TRENDS SUMMARY
The following conclusions are drawn from looking at how products and market uses
change over time:
•
The most important product and user market trend is the existence of
attractive alternate solutions: i.e., DSP processors (DSMPUs), special-function
DSPs (SFDSPs), and application-specific DSPs (ASDSPs).
© 1989 Dataquest Incorporated August
SIS DSP
0004601
Microprogrammable DSP Building Blocks—
Executive Summary
The most important technology trend is that high integration levels make
block partitioning unnecessary, and because of interface delays and software
support, ^undesirable.
The most important business trend is that because of easy market entry, many
have entered this market but few have been successful.
Building block opportunities will continue to exist at the edges of technology
and for immature functions.
The reasons for the creation of the building block market have largely
disappeared.
The building block skill set in the future will reside in application-specific
functional libraries.
SIS DSP
0004601
© 1989 Dataquest Incorporated August
Microprogrammable DSP Building BlocksExecutive Summary
(Page intentionally left blank)
© 1989 Dataquest Incorporated August
SIS DSP
0004601
DSP Building Blocks—Executive Summary
EXECUTIVE SUMMARY
The digital signal processing (DSP) building blocks category currently
enjoys the largest share of the DSP market. This category is also the most
populous in terms of products and competitors. Nevertheless, Dataquest is
forecasting a slowing in the growth rate of the DSP building block segment as
DSP microprocessors and application-specific DSPs take a bite out of this
market.
DSP BUILDING BLOCK MARKET
Market Size
The following should be noted about the size of the DSP building block
market:
•
DSP building block revenue was $131 million in 1986.
•
DSP building block unit shipments were 9.3 million in 1986.
Market Forecast
The Dataquest DSP building block market forecast is as follows:
•
Worldwide DSP building block revenue is forecast to be $139 million
in 1987.
•
Worldwide DSP building block revenue is forecast to be $177 million
in 1990.
Competitive Environment
The following are key points regarding the competitive environment of the
DSP building block market:
•
SIS DSP
The leading suppliers of microprogrammable arithmetic units, which
include bit slices, are AMD, Texas Instruments, and Fairchild,
respectively.
© 1987 Dataquest Incorporated June
DSP Building Blocks—Executive Summary
Leading suppliers of multipliers and multiplier-accumulators include
TRW, Weitek, IDT, AMD, Texas Instruments, and Analog Devices,
respectively.
Interesting new programmeible digital filters are being offered by
Gould, Fairchild, Inmos, Motorola, NCR, RCA, TRW, and Zoran.
DSP BUILDING BLOCK TRENDS
Product Trends
DSP building block product trends include the following:
•
Dataguest expects the DSP building block market to increase at a
modest compound annual growth rate (CAGR) of 7.8 percent through
1990, giving way to DSMPUs and ASICs.
•
Many applications currently using bit slices will shift to ASICs
constructed from ALUs, MACs, and registers currently in cell
libraries.
•
At least eight ASIC manufacturers are offering 2901 equivalent bit
slice cells.
•
New fast 16- and 32-bit ALUs are providing the designer with an
alternative to the traditional 4-bit slice.
Technology Trends
DSP building block technology trends include the following:
•
The bulk of all DSP building block products are still TTL bipolar.
•
ECL bipolar products offered by companies such as Bipolar Integrated
Technology and gallium arsenide products offered by companies such
as Vitesse will begin to take a bite out of the very high speed end
of this market.
© 1987 Dataguest Incorporated June
SIS DSP
Microprogrammable DSP Building Blocks—Forecast
INTRODUCTION
Component products can be defined either by their inherent function (product
distinction) or their end use (market distinction). Each is forecast separately in this
document. The difference is important because natural groupings of products by
function do not necessarily equal natural market groupings. This is particularly true with
digital signal processing (DSP) building blocks, as the following analysis sections
describe. Likewise, the term DSP, as opposed to general-purpose data processing, must
be looked at in both product and market domains. Figure 1 schematically illustrates
these and other distinctions used in this section.
Figure 1
Building Block Product and Market Distinctions
Building Block Product and Market Distinctions
Product (Function) Domain
l\/!arketinq (Use) Domain
Separate-Function Product Lines
General-Purpose
DSP
General-Purpose Systems
Controllers
Coproducers
CPUs
Processor Family Product Lines
General-Purpose
DSP
DSP Systems
Array processors
Special-purpose
Source: Dataquest
August 1989
0004604-1
PRODUCT DEFINITIONS
For forecast purposes, the building block processor families and separate-function
product lines are combined and broken into five segments, as follows:
•
Fixed-point data path elements—Fixed-point data path elements include all
shifters, ALUs, multipliers, and multiplier-accumulators with fixed-point only
formats. They also include bit-slice forms.
•
Floating-point data path elements—Floating-point data path elements include
all of the above functions that use floating-point formats. Fixed-point
formats also may be present, as well as register files.
SIS DSP
0004604
© 1989 Dataquest Incorporated August
Microprogrammable DSP Building Blocks—Forecast
Microcontrollers and address generators—Microcontrollers are microsequencers and special-purpose sequencers such as coprocessor controllers.
Address generators are used for data memories and include memory
management units as well as memory controllers.
Special memories—Special memories include multiported register files,
pipeline registers, and special DSP shift registers. FIFOs are not included.
Special-purpose elements—Special-purpose elements include filter, convolver,
correlator, and template-matching DSP building blocks.
MARKET DEFINITIONS
For forecast purposes, the DSP building block markets are divided into five
segments, as follows:
•
General-purpose data processing systems
Controllers—Controllers are any non-data-path processor used in a
system for any control purpose. The definition includes all peripheral
controllers except coprocessors or data path accelerators.
Coprocessors—Coprocessors are data processors that operate under
control of a CPU such as floating-point coprocessors, vector processors,
and graphics accelerators in a general-purpose system.
Central processing units (CPUs)—Central processing units are used for
general-purpose data processing.
•
Digital signal processing systems
Array processors—Array processors are programmable CPUs or
peripheral processors designed for digital signal processing in its broad
form.
Special-purpose—Special-purpose digital processing systems are used for
a narrow range of signal processing functions or problems and often are
not programmable.
FORECAST
Dataquest expects attractive alternative solutions to reduce the total DSP building
block revenue of $161 million forecast for 1989 to $149 million in 1993. This negative
2 percent compound annual growth rate (CAGR) compares with a positive 16 percent
growth during the previous four years (1985-1988). We believe that the largest revenue
2
© 1989 Dataquest Incorporated August
SIS DSP
0004604
Microprogrammable DSP Building Blocks—Forecast
growth by product type will continue to be achieved by the floating-point data path
elements, as shown in Table 1. The largest revenue growth by market use will be for
data processing coprocessors as Table 2 shows.
Table 1
Worldwide DSP Building Block Revenue Forecast by Product Type
(Millions of Dollars)
Product Type
Data path elements
Fixed-point
Floating-point
Actual
1985 1986 1987 1988
Forecast
1989 1990 1991 1992 1993
$83
4
$ 81 $ 77 $ 62 $ 43 $ 26
60
81
89
96 101
$ 86 $ 85 $ 85
12
37
45
1%
124%
(25%)
14%
7
3
12
44%
26%
15%
4%
(19%)
14%
$161 $177 $170 $160 $149
15.6%
(1.9%)
Microcontrollers &
address generators
Special memories
Special-purpose
Total revenue
6
4
7
4
^_1Q
$97
$111 $139 $150
C^GH
1985-1988 1989-1993
Source:
Dataques t
August 1989
Table 2
Worldwide DSP Building Block Revenue Forecast by Market Use
(Millions of Dollars)
Market Use
Data processing systems
Controller
Coprocessor
CPU
CAGR
Actual
1985 1986 1987 1988
1989 1990 1991 1992 1993
$39
14
11
$ 27 $ 25 $ 20 $ 17 $ 15
80
94 101 105 100
5
4
4
3
3
DSP systems
Array processor
Special-purpose
5
Total revenue
$97
$ 37 $ 34 $ 29
28
58
70
9
6
5
6
31
7
34
8
38
$111 $139 $150
1985-1988
(9%)
71%
(23%)
(14%)
6%
(12%)
13
18
17%
11%
10%
(18%)
$161 $177 $170 $160 $149
15.6%
(1.9%)
9
40
10
44
11
34
12
23
Source:
SIS DSP
0004604
1989-1993
© 1989 Dataquest Incorporated August
Dataguest
August 1989
DSP Building Blocks—Forecast
MARKET DEFINITION
Dataquest breaks the
subsegments, as follows:
DSP
building
block
market
•
Microprogrammable arithmetic units (MAUs)
•
Multipliers and multiplier-accumulators (MACs)
•
Digital filters
•
Other miscellaneous building blocks
segment
into
four
The MAU category includes bit slices, arithmetic logic units (ALU), and
special arithmetic support chips. The MAC category consists of independent
multipliers and acciunulators as well as multiplier-accumulators. The digital
filter category is self-explanatory.
All other building blocks, such as sequencers, register files, and barrel
shifters, are combined into the miscellaneous building block (Other) category.
FORECAST
Dataquest expects the continuing development of both integrated DSP
microprocessors (DSMPUs) and high-speed general-purpose microprocessors to
slow revenue growth of the DSP building block market segment. Table 1 shows
building block revenue growing from $131 million in 1986 to $177 million in
1990, which is equivalent to a compound annual growth rate (CAGR) of
7.8 percent.
Figure 1 compares the revenue growth for each of the four categories of
building blocks. The only category showing any significant growth is digital
filters, which are expected to grow from $9 million in 1986 to $28 million in
1990 at a CAGR of 32.8 percent.
Vigorous competition among building block manufacturers is expected to
drive price reductions in all building block categories. Dramatic reductions
in average selling prices (ASPs) as a result of increased competition and the
conversion from bipolar products to CMOS products will have the greatest
negative effect on revenue growth for the period.
As shown in Table 2, healthy growth in unit shipments is expected through
1990. Dataquest forecasts DSP building block unit shipments to grow at a
CAGR of 24.2 percent, from 9.3 million units in 1986 to 22.1 million units in
1990.
SIS DSP
e 1987 Dataquest Incorporated May
DSP Building Blocks—Forecast
Table 1
WORLDWIDE DSP BUILDING BLOCK REVENUE FORECAST
(Millions of Dollars)
1983
Actual
1984 1985
1986
1987
Forecast
1988 1989
1990
CAGR
83-86
CAGR
86-90
MAUs
MACS
Filters
Other
$32
37
0
25
$ 58
52
2
37
$ 43
36
4
27
$ 51
40
9
31
$ 52
40
14
33
$ 53
41
18
36
$ 56
42
24
19
$ 63
44
28
42
16.8%
2.6%
348.1%
7.4%
5.4%
2.4%
32.8%
7.9%
Total Bldg.
Block Revenue
$94
$149
$110
$131
$139
$148
$161
$177
11.7%
7.8%
Source:
Dataquest
May 1987
Figure 1
WORLDWIDE BUILDING BLOCK REVENUE GROWTH BY PRODUCT CATEGORY
Millions of Dollars
210-| I
[ Othar
ESO Altera
180^ ^ 1
$177
MACa
(A^I
MAUS
150
$161
$149
$131
120
$139
$148
$110
$94
90
fr^^^ ft'^---:^
60
•
Xvv
30-1
0
*^
f^
'^
1983
\vC
1 ,
1984
1985
v1986
W'
WW •'
BBBIB
•S^H
^MWiB
inHHumiW
^r^^P
To^^
1987
19S8
m^ w
i^
© 1987 Dataquest Incorporated May
1989
1990
Source: Diiaquett
M»y 1987
SIS DSP
DSP Building Blocks—Forecast
Table 2
WORLDWIDE DSP BUILDING BLOCK SHIPMENT FORECAST
(Millions of Units)
1983
MAUs
MACS
Filters
Other
2.1
1.1
0
1.6
Total Bldg.
Block
Shipments
4.8
Actual
1984 1985
4.3
1.6
0.1
J2JLA
8.4
3.8
2.1
0.2
JL^
8.5
1986
1987
4.1
2.4
0.3
5.1
2.8
0.5
3.2
11.6
J2^
9.3
Forecast
1988 1989
1990
6.3
3.6
0.9
4.0
8.1
4.5
1.4
4.8
9.8
5.0
1.9
5.4
14.8
18.8
22.1
CAGR
83-86
CAGR
86-90
25.0%
29.7%
210.7%
16.4%
24.3%
20.1%
58.6%
21.2%
24.7% 24,2%
Source:
SIS DSP
O 1987 Dataguest Incorporated May
Dataquest
May 1987
Product Analysis
Microprogrammable DSP Building Blocks—
Product Analysis
PRODUCT DISTINCTIONS
Processor Family versus Separate Function Product Lines
Historically, building blocks have been derived from a natural grouping of gates into
some function such as an adder or arithmetic logic unit (ALU). Product lines are formed
around blocks that fit together to form a processor, while other product lines are formed
around a particular function such as multiplication. Dataquest looks at them as two
different types of product lines; separate function or processor families, as shown in
Figure 1.
Figure 1
Building Block Product and Market Distinctions
Data Busses
Input/Output
ALU
Multiplier
Data
Memory
Control
Control
Address
Control
Control
I/O
Arithmetic
Address
Microcode Instruction Register
Program
Sequencer
Address
Microcode
Program Memory
0004603-1
Source: Dataquest
August 1989
Building Blocks versus Processors
As functional densities have increased, the definition of the term "building blocks"
may have blurred. Generally, however, if such products continue to have control lines
that must be sequenced externally at a basic clock rate, they are considered to be
building blocks. If a product operates on its own after loading control registers, then it
is considered to be a processor; even if multiple products fit together in block style, the
result is not considered a building block. Digital filters still illustrate the distinction:
filters may be designed using multiplier-accumulator array building blocks, which must
have data and controls sequenced to them (e.g., Zoran ZR33881) or standalone
SIS DSP
0004603
© 1989 Dataquest Incorporated August
Microprogrammable DSP Building Blocks—
Product Analysis
special-function digital signal processors (SFDSPs), which need only a data stream and
clock. The Motorola DSP 56200 and the Inmos IMS AllO, for example, are SFDSPs, not
building blocks.
Digital Signal Processing (DSP) versus General-Purpose Data Processing
Figure 2 shows the functional breakdown of a general-purpose data processor.
These functions are the basis for the partitioning of most building block processor
product lines today. Figure 3 is the same diagram modified to represent a typical digital
signal processor. More complex data memory address generation and the
multiply-accumulate data path element are the only significant differences.
Figure 2
Building Block Elements of a General-Purpose Data Processor
Data Busass
.
'
1
1
MultlpNer
Data
Memory
Input/Output
1
Control
'
,,
Address
ALU
Control
^
i
Address
Generator
1
Control
Program
Sequencer
)/o
Arithmetic
Address
Microcode Instruction Register
,1
Address
Microcode
d**
Source: Dataquest
August 1989
0004603-2
© 1989 Dataquest Incorporated August
SIS DSP
0004603
Microprogrammable DSP Building BlocksProduct Analysis
Figures
Building Block Elements of a Digital Signal Processor
Building Block Product and Market Distinctions
Product fFunction^ Domain
Marketina fUse^ Domain
Separate Function .Product Lines
General-Purpose
DSP
General-Purpose Systems
Controllers
Coproducers
CPUs
Processor Family Product Lines
General-Purpose
DSP
DSP Systems
Array processors
Special-purpose
0O04603-3
Source: Dataquest
August 1989
Forecast Distinctions
With this major overlap in usage, no distinction is made between general-purpose
and DSP building blocks for the processor product line market estimates and forecasts.
Separate functional blocks outside of the processor families, however, are included only
if they are DSP related. For example, correlators are considered functional DSP building
blocks, while FIFOs are not part of a processor family, are not uniquely DSP, and so are
not considered DSP building blocks. A/D and D/A converters, likewise, are not uniquely
DSP and therefore are not included.
SEPARATE-FUNCTION PRODUCT LINES
Functional product lines have grown up because suppliers are able to master
enabling technology and/or thoroughly understand the market requirements for that
function. For example, multipliers became a mainstay product line for TRW because the
highly regular design uses TRW's high-density bipolar process to advantage. Knowledge
of the market it created allowed TRW to broaden that market successfully for many
years by introducing many variations of multipliers. TRW never defined a complete
processor product line. Weitek engineers were experts in floating-point computations,
and they used those skills to develop a broad product line that only recently became a
processor family.
SIS DSP
0004603
© 1989 Dataquest Incorporated August
Microprogrammable DSP Building Blocks—
Product Analysis
Different product lines are distinctive in function, precision or data format, and
implementation technology. Dataquest only briefly examines these major distinctions
because market forces may make a careful understanding of product differences and
trends of little value. Major functional and precision distinctions and example products
are summarized in Tables 1 and 2.
Table 1
Examples of Building Blocks in Processor Families
from Major Siqjpliers
Building Blocks
AMD
Data Path Elements
Shifter
16-Bit
32-Bit
Floating Point
Exchange
ALU
4-Bit
8-Bit
16-Bit
32-Bit
Floating Point
MPY
8-Bit
12-Bit
16-Bit
32-Bit
Floating Point
MAC
8-Bit
16-Bit
32-Bit
Floating Point
ALU-MPY
16-Bit
32-Bit
Floating Point
ALU-MPY-RF
16-Bit
32-Bit
Floating Point
AD
TI
Weitek
29130
8838
8833
8839
2901
29501
29116
29332
888
8832
3221
29516
1080
1012
1016
2265
1616
3211
29509
29510
29323
29325
1008
1110
2516
2264
9510
8836
2010
1101
8847
29327
3264
8137
3364
(Continued)
© 1989 Dataquest Incorporated August
SIS DSP
0004603
Microprogrammable DSP Building BlocksProduct Analysis
Table 1 (Continued)
Examples of Building Blocks in Processor Families
from Major Siq^liers
Building Blocks
AMD
AD
TI
Weitek
Control
Program Sequencer
Crossbar
Controller
29331
29141
1402
8818
8841
8136
-
Address Generation
General
FFT
29540
1410
Memory
Register File
Pipeline Register
29334
29525
3128
897
8834
Source:
1066
Dataquest
August 1989
Table 2
Examples of Special-Function Building Blocks
Company
FIR Filters
TRW
Zoran
NEC
LDi
GE/RCA
Correlators
TRW
Product
Precision
(Bits)
MultiplierAccumulators
TDC1028
TMC2243
ZR33481
ZR33881
ZR33891
NCR45CF8
LMS12
TA13073A
4x4
10x10
8x8
8x8
9x9
8x9
12x12
8x8
8
3
4
8
8
4
1
20
TMC2023
TMC2220
TMC2221
64x1
32x4
128x1
64
128
128
Source:
SIS DSP
0004603
Dataquest
August 1989
© 1989 Dataquest Incorporated August
Microprogrammable DSP Building Blocks—
Product Analysis
Functional Distinctions
Figure 3 shows the functional blocks of a DSP processor; these correspond generally
with the functional product lines. This is a result of the building consisting of all
elements that lack internal control or sequencing, and the processors being defined from
their horizontal microcode control method.
Data path operations have the high functional complexity that makes building blocks
attractive. The basic adder function was enhanced with shifters and other logical
operations to be a full ALU. Only the multiplier had to be separate. The multipliers
combined to form an adder for the DSP multiply-accumulate function; then the full ALU
and multiplier were merged.
Simple register sets grew into multiported register files that supplied data to the
arithmetic functions. These register files then merged with the combined arithmetic
functions. Some blocks configured the memory and several multiplier-accumulators to
do filtering and correlation operations. Examples of these special DSP functions are
shown in Table 2.
The relatively simple program counter with branch capability grew into a
stack-and-interrupt controller for the time-critical task of generating program memory
addresses. Address generation for the data memory has received less attention over the
years and were left to the standard ALUs, although they were poorly suited. Only the
fast Fourier transform (FFT) address sequence received special attention.
Data Format Distinctions
As process density increased, data widths expanded from the initial artificially
narrow 4-bit "slice" to 8-, 12- and 16-bit complete words, which is the full precision
that many applications need. The 8-bit words were, and largely remain, the precision of
choice for radar, image, and video processing. Higher-resolution systems in these
applications now call for 12 bits in line with what is attainable from high-speed A/D and
D/A converters.
The highest-speed devices are often building blocks, and, at the highest speeds,
precision often is limited. The 16-bit word widths are natural for DSP at less than the
highest bandwidth signals, as well as in general data processing, so they now are the most
common. Fixed-point formats greater than 16-bit have had limited applications.
Two's-complements representation of negative numbers is almost universal in addition to
unsigned mode.
The addition of floating-point was a natural enhancement with the higher densities
of NMOS circuits, and the building-block format allowed it to be used most widely.
Because of the complex arithmetic and the operation sequence necessary, pipelined
operation was used to increase throughput. Initial products were 32-bit, but multiple
precisions are handled with repeated use of the basic internal arithmetic elements, with
a correspondingly longer processing time.
© 1989 Dataquest Incorporated August
SIS DSP
0004603
Microprogrammable DSP Building Blocks—
Product Analysis
Building blocks, like microprocessor coprocessors, have largely avoided the
floating-point format wars. The first products were compiled with the IEEE standard
754, and largely because of the PC revolution, this is all that has been required. The
Digital, IBM, and military 1750 formats all have been added to some building blocks, but
additional floating-point formats have not been important in determining the success of
these products. Complete IEEE exception handling and function coverage, however, has
been very important.
Technology Distinctions
Because of the simple design and the standard function, the first products in a
faster, denser, or better technology often are building blocks. Generally, the progression
mirrors the rest of the semiconductor industry in absorbing new technologies, but older
ones die more slowly because of their more basic functions and adaptability to special
needs. NMOS and CMOS displaced bipolar slowly, although power and speed
improvement were great. BiCMOS is expected to replace CMOS as I/O speeds become
more important in overall speed performance.
GaAs still has a high cost and limited density, but it is used for its high speeds,
particularly where it can reduce the basic microcode sequencer microcycle time. The
TTL I/O interface and Am 2900 series functions have been maintained.
As device geometries shrink, the on-chip delays get shorter, but off-chip I/O times
remain the same and become relatively larger in the complete system. BiCMOS can
provide faster bipolar I/O with a high-density CMOS core, which will reduce this
problem. The interface preferred still is 5V TTL but the superior speed and noise
properties of ECL I/O make it a new option. Many products with bipolar technology are
available with either TTL or ECL interfaces.
Bipolar processes continue to increase in density and to rival CMOS. Although
BiCMOS has lower power, these high-density bipxjlar processes are at little disadvantage
when used with ECL interfaces.
Examples of Separate-Function Product Lines
TRW's LSI Products Division originated building block circuits for DSP. For more
than 12 years, TRW has remained functional only. Starting with multipliers,
multiplier-accumulators, and memory products, TRW has increased speed and lowered
power requirements through many generations of process enhancement, both bipolar and
CMOS. Functionality increased slowly, and every precision need was met, generally
setting the industry standard for function and pinout.
Integrated Device Technology (IDT) used a higher-performance CMOS process and
aggressive pricing to become a significant supplier. Initial products were simply
higher-speed versions of industry-standard products. But IDT now is introducing new
products with improved functions as well.
SIS DSP
0004603
© 1989 Dataquest Incorporated August
Microprogrammable DSP Building Blocks—
Product Analysis
Logic Devices became a broad supplier by offering low prices and low-cost packages
on industry-standard parts. Product improvements on new proprietary parts have
expanded the line.
Bipolar Integrated Technology (BIT) used its bipolar process to make the fastest
parts commercially available. BIT was first with ECL interfaces as well as the standard
TTL.
PROCESSOR FAMILY PRODUCT LINES
Only four companies have produced product families with interlocked functional
blocks, making a consistent and, in a sense, complete processor. The latest and/or
highest-performing members of these families are shown in Figure 4. Of these, only
Analog Devices and Weitek concentrated on DSP in determining the initial functions.
Advanced Micro Devices
The whole building block concept started in earnest with the Am2900 bit-slice series
from AMD. The initial bipolar 4-bit slice has expanded to 16-bit words with the 29100
series and to 32-bit words with the CMOS 29300 series. Special functions for DSP, such
as multipliers, complex arithmetic ALUs, MACs, and FFT addressers were added in the
29500 series. The 29400 series are future ECL versions of the 29300 series. The 2900
and 29100 were the industry standards that made second-sourcing and enhancements to
individual products possible without producing a whole new product line. This AMD
processor family's steady growth in precision, performance, and technology has defined
building blocks' status. All other products are viewed in relation to their framework.
Texas Instruments
The current 32-bit CMOS TI 88XX family is an outgrowth of the 8-bit bipolar
88X family. Bipolar product availability often was late and was not leading edge in any
sense, so it never became a standard that others followed. The 88XX series is more
aggressive and has set the price/performance standard in single-chip floating-point
functions. The rest of the series faces the same uncertainties as all 32-bit building
blocks at this time.
Analog Devices Incorporated
The ADSP-1000 series started as a CMOS product line with 16-bit precision. It was
directly targeted for DSP and, as a result, paid attention to real-time control issues and
fast concurrent address generation. Its use of other modern computer science concepts
made it attractive for data processing in general. Major enhancements have been made
in the address generator and microsequencer, but the most rapid change has been in the
8
© 1989 Dataquest Incorporated August
SIS DSP
0004603
Microprogrammable DSP Building Blocks—
Product Analysis
floating-point data path blocks driven by the accelerator market. A general strategy has
been to maximize the function per chip rather than carefully selecting functions for the
lowest price and minimum package size.
Weitek
The 8XXX series started as an improved family of loosely coupled processor building
blocks but has become an array processor chip set with three different arithmetic
precisions. The change resulted from trying to solve the microcode problem. By limiting
the configurations to three with fixed instruction sets, higher-level languages could be
supported and microcode libraries could be developed. This strategy put Weitek in
competition with potential customers—board array processor suppliers—but more
accurately, this product family may- have been the first of the high-performance RISC
processors with a fast floating-point vector coprocessor. In any case, Weitek has
progressed out of the traditional building block market.
SOFTWARE SUPPORT
Microcode's flexibility is inherent in building block processor families, but this
flexibility translates into complexity in writing and maintaining the microcode. These
tasks generally are underestimated and constitute building block use's largest
disadvantage. The flexibility limits all but the most basic meta assembler from being a
useful software tool for many different processor designs. The two most widely used
assemblers today are ones provided by microprogram ROM simulator suppliers Step
Engineering and HiLevel Technology.
As building block complexity has grown, functional simulation models have become
necessary, particularly for floating-point data path elements. The IEEE floating-point
standard has elaborate exception handling that increases demands on the microcode
routines as well as improving arithmetic correctness. Because time to market is so
important, a functional simulator often is required before working silicon on new building
block products. Logic Automation has become the dominant supplier of these functional
models.
SIS DSP
0004603
© 1989 Dataquest Incorporated August
Microprogrammable DSP Building BlocksProduct Analysis
(Page intentionally left blank)
10
© 1989 Dataquest Incorporated August
SIS DSP
0004603
Microprogrammable DSP Building Blocks—
Emerging Technology and Trends
PRODUCT AND MARKET TRENDS
The most important market trend is that product alternates to building blocks are
now available to current users for new designs. Technology limits may also impact the
use of building blocks.
Alternate Solutions
The single-chip DSP microprocessor (DSMPU) is now a serious alternative to
building block processor families in many applications. The on-chip performance of
DSPMUs has long been nearly equal to building block performance. However, DSMPUs
now offer much greater I/O and support of external memory, so that total system
performance is now equal or greater. Benefits include lower cost and far easier software
development with a fixed instruction set and full hardware and software support
systems. New multiprocessor configurations promise an even greater performance/price
ratio in the future.
Special-function processors (SFDSPs) now exist for many of the tasks that before
could only be done efficiently with a special microcoded architecture. The complete
digital filter, FFT, and image-processing chips now available often provide higher
performance than building blocks as well as providing much easier use.
The microcoded building block solution has always offered architectural flexibility.
This flexibility now is available in a practical form in application-specific processor
(ASDSP) methodologies. The semicustom cells are often used like building blocks.
Functional simulation tools in this design environment already exceed what is available
for standalone building blocks.
The rapid increase in performance of general-purpose microprocessors also provides
competitive solutions. Not only have raw speeds approached that of DSP building blocks
but functionality has expanded to make general-purpose processors more suitable for
DSP applications. Vector processing and floating-point capability additions to
microprocessors have been the last large market for building blocks. These functions are
increasingly being added on-chip to the main processor. The wide markets for these
general-purpose microprocessors ensures that they will always have attractive, low-cost,
and widely available support systems and personnel.
Technology Limits
Building blocks have always been a way of partitioning functions to take advantage
of chip density while separating the control function to allow flexibility. With the
densities achievable today, fewer functions demand a full chip that cannot also
accommodate the control function. In addition, as on-chip speeds increase, keeping the
control function separate can reduce system speeds substantially as well as unnecessarily
increasing power consumption and packaging costs. I/O speed and power penalities are
increasingly high and unnecessary.
SIS DSP
0004600
© 1989 Dataquest Incorporated August
Microprogrammable DSP Building Blocks—
Emerging Technology and Trends
For new process technologies that offer unusual performance and that alter these
limits, building blocks still make sense. GaAs does not appear to alter these limits, but
rather confirms the interface problem. Its high speed is lost getting on and off the chip
and yet density is still low, so that functions must be partitioned into blocks.
For new functions in which system integration is immature and for very complex
functions, building blocks also may still be appropriate. But product planners should see
this as a near-term necessity rather than part of a positive strategy of a growing
building block product concept.
BUSINESS TRENDS
Market Entry
Becoming a producer of building blocks is relatively easy. There are established
functions, pin-outs, and simulation support so that if a company has an improved process,
industry standard building blocks are a natural vehicle for displaying its lower power or
higher speed. No costly software support is needed. Likewise, improved process density
allows the chip count within an accepted functional framework be reduced. An increase
in precision is a natural and often-used improvement.
Initially, the high profit margins were incentive to enter this market. And with the
low cost of entry, many suppliers entered the market. Building blocks became
commodity parts, and only the lowest-cost producers got any significant market share.
This situation remains current today and the cost of entry for building blocks in the
future also should remain low. As a result, regardless of other factors that are mostly
unattractive, building blocks should always attract competitors because of easy market
entry.
Business Success
Business success is the result of having profitable, producible products that meet a
market need. There can be little question that AMD, Analog Devices, IDT, Texas
Instruments, TRW, and Weitek have had the right products in healthy markets for periods
long enough to recover their investments. But such success is becoming less common,
and now the only successful efforts are for products that move into more complete
processors. The following experiences support those conclusions:
•
TRW, having stuck to function-only products, has suffered
maintained business volume only by doing custom designs.
•
AMD did not recover investment costs on its last expansion to 32-bits with the
Am29300 series. AMD's new thrust is the Am29000, a complete RISC
processor.
© 1989 Dataquest Incorporated August
losses and
SIS DSP
0004600
Microprogrammable DSP Building Blocks—
Emerging Technology and Trends
•
Weitek's profitability and volume growth has come from floating-point
coprocessors for personal computers and workstations. Special-purpose
controllers, like the XL-8200 laser printer controller, are the new company
thrust.
•
Analog Devices is putting new development efforts into DSMPUs, after it lost
out in most of the floating-point coprocessor wars.
•
Texas Instruments has one notable success in its 32-bit CMOS building block
line: the floating-point block used in coprocessors. This product is also part of
its SPARC RISC commitment.
•
IDT, a price and technology leader with a function-only DSP building block
product line, is making no future investments in building blocks.
The functional reasons for building blocks seem to be going away. Building blocks
were a transitional product in the increasing density curve of integrated circuits. The
business success of the suppliers now reflects this when looked at carefully. It will
become even more apparent in the future.
New Business Opportimities
As separate functions, building blocks will continue to offer opportunities,
particularly where other processor family building blocks are not required. Generally,
new building blocks must implement some new function that does not have a natural
connection with some existing microprocessor or else must use some new technology that
is not mature enough to have a natural connection into existing processing systems.
Even these opportunities are best viewed as transient. Building blocks are best viewed as
part of a larger cell-based ASIC design system product plan.
SIS DSP
0004600
© 1989 Dataquest Incorporated August
Microprogrammable DSP Building BlocksEmerging Technology and Trends
(Page intentionally left blank)
© 1989 Dataquest Incorporated August
SIS DSP
0004600
DSP Building BlocksEmerging Technology and Trends
GALLIUM ARSENIDE TRENDS IN DSP
Product development activity is accelerating in the gallium arsenide (GaAs)
industry. In the area of DSP building blocks, several companies are already offering
GaAs versions of the 2901 4-bit ALU slice.
Vitesse
Vitesse introduced the first commercially available LSI GaAs digital IC set, an ECL
lOOK-compatible 4-bit-slice family. The product family consists of the VE29G01 4-bit
microprocessor slice, the VE29G02 look-ahead carry generator, and the VE29G10
microsequencer. The VE29GXX family is expected to permit customers to achieve a
factor of 3x speed improvement over a CPU implementation using the AMD2901C, which
is a bipolar device. Interestingly, Vitesse has employed leaded chip carrier (LCC)
packaging to preserve the device speed advantage at the system level as much as
practical.
Gigabit Logic Inc.
Gigabit Logic (GBL) is also offering a family of 4-bit ALU slice products. They are
the 10G181 4-bit microprocessor slice, the lOGlOO 4-bit adder, and the lOGlOl carry
look-ahead unit. GBL's goal is to become the leader in the commercial GaAs IC
market. Currently, the company is focusing on the computer, instrumentation, defense,
and communications markets.
Adams Russell
Adams Russell offers a family of more than SO analog standard cells for custom and
semicustom applications. The parent company manufactures airborne antenna and cable
assemblies, microwave and RF components and subsystems, and special-purpose
high-speed DSP systems.
Sony Corporation
In 1983, Sony reported results of R&D and successful testing of a GaAs JFET chip
containing 511 DCFL gates organized as a 4x4 multiplier, lO-bit accumulator, and 8-bit
shift register.
SIS DSP
0001736
© 1988 Dataquest Incorporated November
DSP Building BlocksEmerging Technology and Trends
(Page intentionally left blank)
© 1988 Dataquest Incorporated November
SIS DSP
0001736
Microprogrammable DSP Building Blocks—
Application and User Issues
MARKET DISTINCTIONS
Initially, the uses of building blocks were as broad as the TTL market itself because
the primary benefit simply was higher density. As processor families began to form,
then programmable processors of all types, special-purpose or general data processing
became more popular. As microprocessors flourished and gate arrays became larger,
building blocks have been used increasingly only for special-purpose tasks involving
coprocessors, controllers, or digital signal processing (DSP), where the highest speeds are
necessary or the processing architecture is unusual. Yearly unit volume in these
applications generally is low or moderate; otherwise, an ASIC design would be justified.
The majority of building blocks still are used to build programmable processors, but the
trend is toward hardwired fixed-function or state machine designs.
GENERAL-PURPOSE DATA PROCESSING SYSTEMS
The naturally wide instruction sets of the building block families generally have not
aligned with the instruction sets of major central processing unit (CPU) suppliers.
Initially, however, microcode allowed designers to emulate existing minicomputer
instruction sets. In addition, microcoded systems could be enhanced readily as the
instruction set grew. Compared with the original complex-instruction-set computers
(CISC) and slow-core main memories, multiple-chip building block designs suffered no
speed penalty.
As fast, inexpensive semiconductor memories became available, the speed penalty
was noticeable unless the processor operated in its native microcode. Within general
data processing, native microcode operation was possible only in special attached
controllers or coprocessors. Typically, coprocessors were designed for floating-point
arithmetic, vector processing, simulation kernels, or graphics.
In minicomputer and mainframe systems, high-speed microcoded peripheral
controllers became the main use for building blocks because every system had multiple
disk and I/O channel controllers. The floating-point coprocessor often used the same
technology as the CPU as the coprocessor became more of a standard part of the
system. Also, non-IEEE formats had to be supported.
New superminicomputer and parallel processor systems adopted the vector and
general-purpose array processor requirement for simulation and graphics. The IEEE
floating-point format is the usual choice, so superminicomputers are making wide use of
the floating-point arithmetic building blocks, sequenced by a controller, and usually
designed in gate arrays.
SIS DSP
0004602
© 1989 Dataquest Incorporated August
Microprogrammable DSP Building Blocks—
Application and User Issues
In PC and microcomputer systems, simulation and graphics created a large market
for faster vector and floating-point calculations. The need for faster vector and
floating-point calculations was met in two ways, both of which used floating-point
building blocks:
•
Add-in accelerator boards
•
Higher-speed replacements for existing floating-point coprocessors such as
the 80X87 family
The high-performance engineering workstation is the complete answer to simulation
and graphics needs. As with microcomputers, special controllers are used rather than
others of the building block processor family. Today, this represents the largest single
use of building blocks, generally for floating-point ALUs, multipliers, and register files
in various combinations.
DIGITAL SIGNAL PROCESSING SYSTEMS
The array processor is the general-purpose computer for DSP. Array processors
always have been microcoded for speed and flexibility and thus were quick to use DSP
building blocks. Large array processors now are in decline as they are replaced by
superminicomputers. But the density of new building blocks has allowed moving a
function onto a single board for use in PCs, workstations, or VME bus microcomputer
systems. The end applications cover all digital signal processing applications in some
form. Most processors are broadly applicable, but image and graphics requirements are
different enough to require that special architectures be used. These array processors
are the purest application of the DSP building block processor families and are highly
visible, but they are not a large-volume market.
Many signal processing tasks are highly repetitive and do not need the flexibility of
programming. In addition, maximum speed is needed, so a hardwired special-purpose
processor is used. Data path building blocks and special-function blocks often can be
used directly, although some sequencing circuits may be required. Large-end
applications include radar, sonar, medical and military imaging, graphics, and
high-performance real-time simulation. Collectively, these represent the largest
market for DSP building blocks.
© 1989 Dataquest Incorporated August
SIS DSP
0004602
ASICs/SFICs Executive Summary
ByBPTTTTVE SUMMARY
The combined potential of application-specific DSP integrated circuits
(ICs) and special-function DSP ICs is tremendous. Dataquest believes that
application-specific ICs (ASICs) and special-function ICs (SFICs) will
account for 44 percent of all DSP revenue by 1990.
ASIC/SFIC MARKET
Market Size
The following should be noted about the size of the ASIC/SFIC market:
•
Application-specific DSP IC revenue was $68 million in 1986.
•
Application-specific DSP IC unit shipments were 9.8 million in 1986.
•
Special-function DSP IC revenue was $50 million in 1986.
•
Special-function DSP IC unit shipments were 3.8 million in 1986.
Market Forecast
The Dataquest ASIC/SFIC market forecast is as follows:
•
Worldwide application-specific
$98 million in 1987.
DSP
IC
revenue
is
forecast to be
•
Worldwide application-specific DSP
$192 million in 1990.
IC
revenue
is
forecast
•
Worldwide special-function
$73 million in 1987.
DSP
IC
revenue
is
forecast
to
be
•
Worldwide special-function
$146 million in 1990.
DSP
IC
revenue
is
forecast
to
be
SIS DSP
© 1987 Dataquest Incorporated June
to be
ASICs/SFICs Executive Summary
Competitive Environment
The following are key points regarding the competitive environment of the
ASIC/SFIC market:
•
The ASIC marketplace is crowded, to say the very least. A minimum
of eight vendors offer 2901 bit-slice cells in their ASIC
libraries.
Many more vendors are capable of duplicating the
function in bipolar gate arrays—even in ECL bipolar gate arrays.
•
The SFIC market is diverse by definition, with pockets of heavy
competition in certain areas such as 300-baud modems.
ASIC/SFIC TREKDS
Product Trends
ASIC/SFIC product trends include the following:
•
Dataquest expects the application-specific DSP market segment to
increase at a compound annual growth rate (CAGR) of 29.6 percent
through 1990 as discrete DSP building block designs convert to ASIC.
•
We expect the special-function DSP segment to grow at a CAGR of
30.7 percent, fueled by the growing digital-telecommunications and
graphics-generation markets.
Technology Trends
The availability of high-speed, low-power CMOS standard cells is expected
to have a significant impact on the ASIC DSP market.
Application Trends
ASIC/SFIC application trends include the following:
•
Special-function DSP ICs are expected
important role in graphics generation.
to
•
Certain general signal processing applications in telecommunications, such as modems, will also evolve from general-purpose DSP
microprocessors to special-function DSP ICs.
© 1987 Dataquest Incorporated June
play
an
increasingly
SIS DSP
ASICs/SFICs—Forecast
DSP APPLICATION-SPECIFIC IMTEGRATED CIRCUITS (ASICs)
Market Definition
Dataquest recognizes four types of products as belonging to the
application-specific integrated circuit (ASIC) product category, as follows:
•
Programmable logic devices (PLDs)
•
Gate arrays
•
Cell-based integrated circuits (CBICs)
•
Full-custom logic designs
Figure 1 shows the entire ASIC family tree segmented as either semicustom
or custom devices.
Figure 1
ASIC FAMILY TREE
ASIC Family Tree
1
1 •
Program mable
Log ic
Semicustom
Custorn
1
Geite
Arr ays
Cell-Base<d Des gns
(CB Cs)
Fj l l
Cus tom
Source; Dataquest
June 1987
SIS DSP
© 1987 Dataquest Incorporated June
ASICs/SFICs—Forecast
ASICs are used in many digital signal processing (DSP) applications to
replace a combination of small-scale integration (SSI) devices, medium-scale
integration (MSI) devices, and DSP building blocks.
Signals from one
component to another frequently require slight modification. SSI and MSI are
the glue that cements the building blocks together. The number of gates
required to glue building blocks together is usually small. SSI and MSI have
only a few gates per chip; therefore, the number of individual SSI and MSI
chips needed is large, typically accounting for 50 percent of the printed
circuit board area. By using an ASIC design that incorporates many of the
glue-chip functions, the total chip count of a DSP design can be
significantly decreased. The reduction in chip count results in a cost
advantage that, at high production quantities, outweighs the higher
development cost of ASICs.
It should be pointed out that there is no one best design technique.
ASICs, building blocks, and DSMPUs all excel in certain areas.
ASICs
generally are preferred for very high volume applications.
Forecast
Dataquest estimates that gate arrays and cell-based ASIC designs used in
DSP applications will grow at a 29.6 percent compound annual growth rate
(CAGE), from $68 million in 1986 to $192 million in 1990.
Table 1
illustrates the relative growth of DSP ASIC revenue by end-use market. We
believe that the use of ASICs will grow most rapidly in high-volume
communications and consumer applications. Military use of ASICs will remain
quite strong, with a CAGE of 28.0 percent through 1990.
Because of size and performance benefits, as well as ease of
implementation, DSP ASIC customers will increasingly turn to cell-based
integrated circuit (CBIC) designs. Dataquest expects that 90 percent of ASIC
DSP designs will be cell based by 1990.
© 1987 Dataquest Incorporated June
SIS DSP
ASICs/SFICs—Forecast
Table 1
WORLDWIDE DSP ASIC REVENUE BY END-USE MARKET
(Millions of Dollars)
Actual
1983 1984 1985 1986
Data Processing
Communications
Industrial
Consumer
Military
Transportation
Total DSP
ASIC Revenue
Forecast
1987 1988 1989 1990
CAGE
83-86
CAGR
86-90
0
$ 5
20
2
1
23
0
$ 7
17
3
2
18
0
$ 9
28
4
2
25
0
$12
40
6
3
36
1
$ 15 $ 18 $ 21
84
71
55
11
9
8
7
5
4
67
56
45
1
2
1
44.2%
32.6%
58.7%
26.0%
24.4%
0
23.6%
31.6%
28.8%
36.8%
28.0%
111.5%
$30
$51
$47
$68
$98
$128 $160 $:192
31.4%
29.6%
$ 3
12
1
1
13
Source:
Dataquest
June 1987
DSP SPECIAL-FUNCTION INTEGRATED CIRCUITS (SFICs)
Market Definition
Dataquest defines special-function DSP integrated circuits (SFICs) as
commercially available, standard semiconductor products that are designed to
perform a specific function and employ digital signal processing. DSP
capability is not the exclusive function of these products; rather, DSP is
primarily an implementation technique allowing the circuit to perform its
targeted function.
These dedicated or specialized chips typically originate as a custom
design for a specific customer or group of customers. As the particular end
market grows, demand for these and other specialized chips to serve the
application also grows. Eventually, merchant semiconductor companies will
offer standard semiconductor products to fill these needs.
SIS DSP
© 1987 Dataquest Incorporated June
ASICs/SFICs—Forecast
Special-function circuits employing DSP will replace general-purpose DSP
designs as the size of each particular market warrants it. Examples of SFICs
include the following:
Modems
CODECS
Music-synthesizer chips
Speech-synthesizer chips
Speech-recognition chips
Forecast
SFICs have applications in a wide variety of end equipment markets
including the telecommunications, consumer, and military markets. Dataquest
forecasts that the consumption of DSP SFICs will grow at a 30.7 percent CAGR
from $50 million in 1986 to $146 million in 1990. As shown in Table 2, we
believe that rapid growth will occur in the communications, industrial, and
military segments.
Table 2
WORLDWIDE DSP SFIC REVENUE BY END-USE MARKET
(Millions of Dollars)
Actual
1983 1984 1985 1986
Data Processing
Communications
Industrial
Consiomer
Military
Transportation
Total DSP
SFIC Revenue
Forecast
1987 1988 1989 1990
CAGR
83-86
CAGR
86-90
$ 1
9
1
0
10
0
$ 3
14
2
1
15
0
$ 5
13
2
1
13
0
$ 7
20
3
2
18
0
$ 9
31
4
2
26
1
$11
41
6
3
35
1
$ 15 $ 16
64
56
7
9
4
5
44
50
2
2
91.3%
30.5%
44.2%
171.4%
21.6%
0.0%
23.0%
33.8%
31.6%
25.7%
29.1%
111.5%
$21
$35
$34
$50
$73
$97
$128 $ 146
33.5%
30.7%
Source:
© 1987 Dataquest Incorporated June
Dataquest
June 1987
SIS DSP
Competitive Analysis
ASICs/SFICs—Suppliers
APPLICATION-SPECIFIC DSP INTEGRATED CIRCUITS (ASICs)
ASICs offer a designer the opportunity to fully integrate a buildingblock system onto a single piece of silicon. Several ASIC vendors now offer
cell equivalents of AMD's 2900 microprogrammable family. Table 1 highlights
some of these vendors. Table 2 gives a more complete list of worldwide ASIC
suppliers by region, by technology, and by type of product offered.
SIS DSP
© 1987 Dataquest Incorporated June
ASICs/SFICs—Suppliers
Table 1
BEPRESENTATIVE ASICs SUPPORTING
MICSOPROGRAMMABLE COMPONENTS
ASIC PART NUMBER
MICROPROGRAMMABLE
ELEMENTS SUPPORTED
IN LIBRARY
(SEE NOTE 1)
iC
PROCESS
2901 A.B-TO-Y
DELAY (nSEC)
NRE CHARGES
PART COST
CALIFORNIA
DEVICES
CHA3200
2901, 2910
CMOS
51
$18k TO S24k
$7 (10,000) IN AN
84-LEAD PLCC
GOULD INC
MEGACELL
2901, 2902, 2904, 2909,
2910, 2911
HCMOS
72
(SEE NOTE 2)
(SEE NOTE 2)
INTEGRATED
LOGIC SYSTEMS
CA2000
2901. 2902, 2904 2910
CMOS
67
S25k TO S5Sk
$22 77 (5000) FOR A
CA2110 IN AN
84-LEAD PLCC
LSI LOGIC
LSA2005
2901, 2902, 2903. 29203,
2904. 2909, 2910, 2910A,
16-BIT 2910. 2911,29116,
29117. 29501
HCMOS
56
$40k'nOS80k
$50 TO $150
COMMERCIAL.
BASED ON PACKAGE
AND VOLUME
LSA2006
SAME AS LSA200S
HCMOS
66.1
SAME AS LSA200S
SAME AS LSA200S
LSA2011
SAME AS LSA200S
HCMOS
56
SAME AS LSA2005
SAME AS LSA2005
NATIONAL
SEMICONDUCTOR
SCX6200
2901. 2909. 2911
CMOS
29
$15k TO $70k
(SEE NOTE 3)
SCL
2901. 2909. 2911
CMOS
29
SSOkTOSTOk
(SEE NOTE 4)
UNICORN
COMPILE
29C01, 29C03. 29C10.
29C14
CMOS
lOfPCP)
FROMSSOk
(SEE NOTE 5)
MANUFACTURER
VTC
WAFERSCALE
NOTES:
1 KEY TO 2901
2901
2902
2903, 29C03
29203
2904
VL2000
2901, 2902
CML
3 25
S40k TD S50k
(SEE NOTE 6)
MODULAR-CELL
8-, 16-, 32-BIT 2901s.
29iaA
CMOS
30 (4-BIT)
S60k TO $12Sk
(SEE NOTE 7)
FAMILY;
4-BIT ALU SLICE
LOOKAHEAO-CARRY GENERATOR
4-BIT EXPANOED-FUNCTION ALU SLICE
4.BIT ALU SLICE WITH BCD ARITHMETIC
STATUS AND SHIFT-CONTROL UNIT
2909
2910, 29C10, 2910A
2911
29116
29117
29501
HYPOTHETICAL GOULD MEGACELL ASIC:
FUNCTION
MICROCODE ENGINE COMPONENTS:
QUANTITY DEVICE
1
2910 MICROSEQUENCER
4
2901 4-BIT SLICE ALU
1
2902 LOOKAHEAD CARRY
REGISTERS. MUXES (1000-GATE EQUIVALENT)
MICROCODE STORAGE:
Ik-WORDx 60-BIT ROM
AUXILIARY STORAGE:
1k-WOR0x16-aiT ROM
512-WORD X 16-BIT RAM
4-BIT MICROSEOUENCER SLICE
12-BIT MICROSEQUENCER
4-BIT MICROSEOUENCER SLICE
16-BIT ALU
2P0HT, 16-8IT ALU
MULTIPORT. PIPELINED. 8-BIT ALU SLICE
% AREA CHIP USAGE
SUBTOTAL
22%
SUBTOTAL
24%
15%
PROPRIETARY DIGITAL COMPONENTS: 2000-QATE EQUIVALENT RANDOM LOGIC
ANALOG INTERFACE CIRCUITS:
29%
FILTER. SAMPLE AND HOLD ANALOG SWITCHES.
COMPARATOR. VOLTAGE DOUSLER. CLOCK CIRCUITS SUBTCfTAL
SUBTOTAL
TOTAL
10%
100%
USER PROVIDES GOULD WITH A SEMIVALIDATEO NET LIST USING GOULD LIBRARIES NRE COST LESS THAN $50,000; PART COST; LESS
THAN $30 (10.000/YEAR) IN 68-LEAD PLCC,
A NATIONAL SEMICONDUCTOR SCX6287 GATE ARRAY WITH A 16-BIT ALU REPRESENTING APPROXIMATELY 25% OF THE DIE AREA,
PACKAGED IN AN 84-LEAD PLCC, WOULD COST APPROXIMATELY $39.
A NATIONAL SEMICONDUCTOR SCL SEMICUSTOM ASIC WITH A 16-BIT ALU REPRESENTING APPROXIMATELY 25% OF THE DIE AREA,
PACKAGED IN AN 84-LEAD PLCC, WOULD COST APPROXIMATELY $32
A UNICORN ASIC GENERATED FROM THE COMPILE LIBRARY WITH A 16-8IT ALU REPRESENTING APPROXIMATELY 25% OF THE DIE AREA,
PACKAGED IN A 68-PIN PLCC. WOULD COST APPROXIMATELY $14 (50.000).
A VTC VL2000 SEMICUSTOM ASIC WITH A 16-BIT ALU REPRESENTING APPROXIMATELY 25% OF THE DIE AREA. PACKAGED IN AN 84-LEAD
PLCC. WOULD COST APPROXIMATELY $50 (500Q/YEAR)
A WAFERSCALE MODULAR-CELL ASIC WITH A 32-BIT ALU. A lk-WOR0x64-BIT EPROM. A 6k-BIT RAM, AND A 5910 MICROSEOUENCER IN A
t32-PIN PLASTIC PGA COSTS APPROXIMATELY $200 (5000/YEAR)
Source: EDN
1987 Dataguest Incorporated June
SIS DSP
ASICs/SFICs—Suppliers
Table 2
WORLDWIDE ASIC SUPPLIERS
Companies
by Region
MOS
PLDS
Bipolar
Gate Arravs
Bipolar
MOS
MOS
CBICs
Bipolar
16
7
83
42
80
13
15
North American Companies
0
AT&T Technologies
0
Acumos
1
Advanced Micro Devices
1
Altera
Applied Micro Circuits Corp. 0
7
0
0
1
0
0
51
0
1
0
0
1
28
0
0
1
0
1
55
1
0
0
0
1
10
0
0
0
0
1
Atmel
Barvon Research
California Device Inc.
California Micro Devices
Calmos Systems
1
0
0
0
0
0
0
0
0
0
0
1
1
1
1
0
1
0
0
0
0
1
1
1
0
0
0
0
0
0
Cherry Semiconductor
Circuit Technology Inc.
Cirrus Logic
Commodore Semiconductor
Custom Arrays
0
0
0
0
0
0
0
0
0
0
0
1
0
1
0
1
0
0
0
1
0
1
1
1
0
0
0
0
0
0
Custom Integrated
Circuits (CIC)
Custom Silicon
Cypress Semiconductor
Data Linear
Exar Integrated Systems
0
0
1
0
0
0
0
0
0
0
0
1
0
0
1
1
1
0
1
1
1
1
0
0
1
0
0
0
1
0
Exel Microelectronics
Fairchild Semiconductor
Ferranti Interdesign
GTE Microcircuits
General Instruments
1
0
0
0
0
0
1
0
0
0
0
1
1
1
1
0
1
1
0
0
0
1
1
1
1
0
0
1
0
0
Gould Semiconductors
Harris Semiconductor
Holt Integrated Circuits
Honeywell
Hughes Solid State
1
1
0
0
0
0
1
0
0
0
1
1
0
1
1
0
1
0
1
0
1
1
1
1
1
0
0
0
0
0
Worldwide Total
(Continued)
SIS DSP
1987 Dataquest Incorporated June
ASICs/SFICs—Suppliers
Table 2 (Continued)
WORLDWIDE ASIC SUPPLIERS
Companies
by Region
MOS
Gate Arrays
MOS
Bipolar
PLDs
Bipolar
MOS
CBICs
Bipolar
ICI Array Technology
ITT VLSI
Integrated Circuits Systems
Integrated Logic
Systems (ILSI)
Integrated Microcircuits
0
0
0
0
0
0
1
1
1
0
0
0
1
1
1
0
0
0
0
0
0
0
1
1
0
0
0
0
0
0
Intel
Intercept Microelectronics
Interconics
International CMOS
Technology
International
Microcircuits Inc. (IMI)
1
0
0
0
0
0
1
1
1
0
0
1
0
0
0
International
M i c r o e l e c t . P r o d u c t s (IMP)
LSI L o g i c
Laseroath
L a t t i c e Semiconductor
Linear Technology I n c .
1
1
0
0
:1.
0
0
0
1
0
0
0
0
0
0
0
0
0
1
0
1
1
0
0
0
1
0
1
1
1
0
1
0
0
0
0
0
1
1
1
1
1
0
0
0
0
0
0
0
1
1
1
1
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
Micro-Rel
M i c r o c i r c u i t s Technology
Mitel
Mitron
M o n o l i t h i c Memories
0
0
0
0
1
0
0
0
0
0
Motorola
0
0
1
1
NCM
NCR
0
0
0
0
1
1
1
0
National Semiconductor
Pacific Microcircuits Ltd.
0
0
0
1
0
0
0
0
1
0
0
0
0
1
1
MCE S e m i c o n d u c t o r
Matra Design Semiconductor
McDonnell Douglas
Corporation
Micro L i n e a r
M i c r o Power S y s t e m s
1
0
0
0
0
0
0
0
1
1
Q
0
0
0
0
1
1
1
0
p
i
1
1
1
1
1
(Continued)
1987 D a t a q u e s t I n c o r p o r a t e d
June
S I S DSP
ASICs/SFICs—Suppliers
Table 2 (Continued)
WORLDWIDE ASIC SUPPLIERS
PLDs
Bipolar
Gate Arrays
MOS
Bipolar
CBICs
Bipolar
Companies
by Region
MPS
Panasonic
Polycore Electronics
RCA
Raytheon
S-MOS Systems (Seiko)
0
0
0
0
0
0
0
0
0
0
1
0
1
1
1
0
1
0
1
0
1
0
1
0
1
0
0
0
0
0
Sierra Semiconductor
Signetics
Silicon Systems
Siliconix
Siltronics
0
1
0
0
0
0
1
0
0
0
0
0
0
1
0
0
1
1
0
0
1
1
1
0
1
0
0
0
0
1
Soraque Solid State
Standard Microsystems Corp.
Supertez
Tektronix
Texas Instruments
0
0
0
0
1
0
0
0
0
1
0
0
0
0
1
0
0
0
1
1
1
1
1
0
1
0
0
1
0
0
United Microelectronics
Corp.
United Technologies
VLSI Technology Inc.
VTC
Vatic
0
0
1
0
0
0
0
0
0
0
1
1
1
1
0
0
0
0
1
1
0
0
1
1
0
0
0
0
1
0
WaferScale Integration
Western Digital
Xilinx
Zymos
0
0
1
0
0
0
0
0
0
1
0
0
0
1
0
0
1
1
0
1
0
0
0
0
Japanese Companies
Asahi
Fujitsu
Hitachi
Matsushita Electronics
MitsTibishi Electronics
1
0
0
0
0
0
0
0
0
0
0
0
11
1
1
1
1
1
5
0
1
1
0
1
11
0
1
0
1
1
0
0
0
0
0
0
MOS
(Continued)
SIS DSP
© 1987 Dataquest Incorporated June
ASICs/SFICs—Suppliers
Table 2 (Continued)
WORLDWIDE ASIC SUPPLIERS
Companies
by Region
MOS
NEC
Oki
Ricoh-Panatech
Sanyo
Seiko Epson
0
0
Sharp
Toshiba
Yamaha
PLDs
Bipolar
Gate Arrays
MOS
Bipolar
MOS
CBICs
Bipolar
0
1
X
1
6
0
i
1
X
1
1
60
0
0
0
I
0
0
1
1
0
0
#
b
0
0
1
1
0
a
t
0
9
1
1
0
European Companies
0
ASEA HAFO Inc.
Austria Microsystems (Gould) 0
Electronic Technology
Corp. (ETC)
0
0
European Silicon Structure
Eurosil GMBH
0
Q
a
a
a
18
0
1
8
0
0
14
1
1
p
0
1
0
1
1
0
0
1
1
0
0
0
0
Ferranti Electronics
HMT Microelectronics Ltd.
Heuer Microtechnology (HMT)
Marconi Electronic Devices
Matra-Harris Semiconductors
0
0
0
0
0
0
0
0
0
0
1
1
1
1
1
1
0
0
0
0
1
0
0
1
0
1
0
0
0
0
Micro Circuits
Engineering (MCE)
Mietec
Newmarket Microsystems
Philips
Plessey
0
0
0
0
0
0
0
0
0
0
1
0
1
1
1
0
0
1
0
1
1
1
0
0
1
0
0
0
0
1
0
0
0
0
0
0
0
0
0
1
1
1
1
0
0
1
0
0
1
1
0
0
1
0
&
0
1
1
1
0
R.T.C. La
Radiotechnique-Compelec
Recal Electronics Ltd.
SGS Semiconductor
Siemens
Silicon Microsystems
LTD. (SMS)
6
0
0
b
b
i
0
0
q
Q
^
•
b
b
(Continued)
© 1987 Dataquest Incorporated June
SIS DSP
ASICs/SFICs—Suppliers
Table 2 (Continued)
WORLDWIDE ASIC SUPPLIERS
Companies
by Region
MOS
Smiths
Swindon Silicon Systems Ltd.
Thomsen Semiconductor
(inc. Mostek)
ROW Companies
AWA Microelectronics
Gold Star Semiconductor
Micro Electronic Ltd.
PLDs
Bipolar
Gate Arrays
MOS
Bipolar
0
0
0
0
1
0
0
0
0
0
0
0
0
0
3
1
1
1
MOS
0
0
1
1
0
0
0
1
0
0
0
0
0
0
0
0
0
0
Source:
SIS DSP
© 1987 Dataguest Incorporated June
CBICs
Bipolar
Dataquest
June 1987
ASICs/SFICs—Suppliers
(Page intentionally left blank)
1987 Dataguest Incorporated June
SIS DSP
DSP—Technical Overview
DEFINinON
Digital Signal Processing (DSP) is a circuit technique whereby a digital
representation of an analog waveform is manipulated using computer techniques. Digital
processing allows operations upon signals that are difficult and complex (or sometimes,
impossible) with conventional analog methods.
DIGITAL SIGNAL PROCESSING ADVANTAGES
The advantages of digital signal processing techniques affect system performance,
engineering design, manufacturing, and maintenance. Cost reductions may occur, but
that will depend upon the complexity of the final system.
Performance Advantages
Precision and Accuracy
Wider word lengths improve the signal-to-noise ratio. It is neither limited nor
changed by the noise of its individual transistors. This allows flexibility in choosing the
best word length for each application.
With sufficient processing speed, a DSP system will reproduce exact filter
characteristics. Also, it is easy to design a phase linear filter (using a finite impulse
response filter). This is difficult in analog circuits.
Temperature Insensitivity
The properties of a DSP processor do not drift with changes in the temperature of
its environment. This is a frequent reason for performance shortcomings in analog
circuits. Compensating for temperature drift in circuits is a difficult engineering task
and is more of an art than a science.
Vibration Insensitivity
Physical vibrations cause an effect called microphonics in circuits that use inductors
and capacitors. Microphonics causes undesirable changes in performance during high
G-force situations and periods of vibration, such as in military applications. DSP
components do not change when shaken; thus, vibrations do not affect the circuit output.
SIS DSP
0001731
® 1988 Dataquest Incorporated November
DSP—Technical Overview
Design Advantages
A DSP program (called an algorithm) usually consists of addition and multiplication
operations that are performed on binary words that have up to 32 bits of accuracy. With
DSP techniques, designers have the flexibility to choose the word length and
computational speed to fit each application.
Programmability
In analog circuits, changing the filter characteristics can require changing the
component values. In digital signal processing only the coefficients in the equations need
to be changed. Thus a filter can adapt to signal line conditions. Similarly, a DSP
processor can be completely reprogramraed to perform a different set of functions.
Quick Implementation
Since software determines the characteristics of the DSP designs, circuits can be
quickly prototyped, changed, and upgraded. DSP software is, however, more difficult
than non-real-time microcomputer programming.
Integration
Capacitors and inductors necessary for analog processing are difficult to integrate
on silicon. For instance, when IC features are smaller than 5 microns, capacitors
become nonlinear. This causes implementation problems for conventional analog
techniques such as switched capacitor filters.
DSP circuits, on the other hand, using basic on-off circuitry, are amenable to
shrinking. In fact, 0.5-micron parts are now working in engineering laboratories.
Shrinking the chip does not change the circuit function, only its speed and size. Thus
DSP has a clear future as transistors get smaller and faster.
Pinouts
Analog integrated circuits require pins to attach the capacitors, resistors, and
inductors that are not part of the chip. Even with technologies such as switched
capacitors, which eliminate much of the need for external components, there are always
pins for components that alter and customize the characteristics of the part (slew rate,
band pass, etc.). Future DSP chips could have a single pin to allow bit serial
programming from a microcomputer. Additionally, DSP circuits for audio frequencies
can also use bit serial signal paths. As silicon dice continue to shrink, packaging costs
and board space are ultimately the determining factors in system costs. Therefore DSP
will eventually be cheaper than analog just based on pinouts.
© 1988 Dataquest Incorporated November
SIS DSP
0001731
DSP—Technical Overview
Manufacturii^ Advantages
R^>eatability
Each DSP system coming off the assembly line can have the same algorithm as any
other. Analog components, however, are manufactired with a tolerance on the
parameter values. Because of the cumulative effects of these tolerances, circuit designs
that use many analog components require individual adjustments on every circuit to meet
critical performance specifications.
Device-to-device conformity is desirable in communication systems where optimum
information transfer depends on the matching of transmitter and receiver characteristics.
Maintenance Advantages
Stability
Since the DSP algorithm does not change with temperature or age, once it is fixed
into read only memory, it stays fixed. There are fewer capacitors, inductors, or resistors
to drift out of alignment on equipment using digital signal processing ICs. Consequently,
the systems need less routine maintenance to adjust and recalibrate the circuits. Since
the performance-critical features of the circuit are embodied in software, routine
maintenance can be performed by less-skilled laborers.
Cost
Cost saving is dependent upon the complexity of the final system. In simple systems
with modest performance requirements, the costs of DSP components are unacceptable
at the present state of the art. When designing a single filter, for example, it is usually
cheaper to use operational amplifiers instead of a signal processor.
With high-performance applications, however, the cost advantages are clear. In
fact, some systems have performance requirements that are very difficult to achieve
without using DSP techniques. For example, one DSP chip can replace an entire board of
audio frequency analog circuits. A DSP processor can implement 100 filters by
multiplexing the use of a single multiplier and adder. An analog implementation, by
comparison, would need 100 individual sets of hardware. In a DSP circuit, each signal
requires only a memory location, while an analog implementation requires an actual line
for each signal.
SIS DSP
0001731
© 1988 Dataquest Incorporated November
DSP—Technical Overview
Life costs are lower since the DSP algorithm does not change as devices age. This
helps to minimize "downtime" required for routine maintenance because the DSP circuits
do not need frequent realignment. In many military and telecommunications projects
maintenance and total life costs are primary decision criteria.
DIGITAL SIGNAL PROCESSING DISADVANTAGES
Lack of Knowledge
The slow acceptance of DSP is in part due to the need for education. Digital signal
processing is a new technique compared to conventional analog methods. Consequently,
fewer engineers are skilled in DSP compared to those who understand alternative
techniques such as operational amplifiers. Furthermore, many of the design engineers
experienced with DSP work on sensitive or classified military projects. For security
reasons they are unable to publish as frequently or in as much detail as their industrial
counterparts. This means that detailed application knowledge as well as the clever ideas
and tricks-of-the-trade will take longer to diffuse throughout the industry.
Retraining
Test and assembly technicians will require retraining in the digital signal processing
technology. This involves considerable expense and time in a large manufacturing
organization. Technicians who have accumulated years of intuitive know-how will find a
portion of their skill obsolete with the conversion to DSP technology.
Instrumentaticm
Manufacturing operations will have to retrofit test and assembly stations with some
new equipment and other instruments will become obsolete. This also applies to field
repair depots.
Obsolescence
New parts in this rapidly changing technology are being introduced every month.
But standards have not yet emerged. As IC manufacturers leam more about application
requirements, they will continue to upgrade their component offerings. The result will
be early obsolescence of first-generation parts. The risk is greater for equipment
manufacturers because the semiconductor industry has not developed second sources for
the new DSP chips.
© 1988 Dataquest Incorporated November
SIS DSP
0001731
DSP—Technical Overview
National Defense Issues
Military applications have been the driving force behind development of digital
signal processing methods. Many of the components available today are critical in
defense systems. From time to time this nation expresses concern about the export of
such advanced technology. We know of no activity at present to limit the export of
these devices. However, such actions could be considered in response to concerns about
the si4)remacy of the United States' global military position. Such action would affect
the international marketing strategy of U.S. semiconductor manufacturers.
A SIMPLE DSP CIRCUIT
A block diagram of a simple digital signal processing system is shown in Figure 1.
This circuit is too simple to perform a useful function but will serve to illustrate some
digital signal processing fundamentals. The basic elements are an analog-to-digital
convertor (A/D), a digital section that performs operations using computer techniques,
and a digital-to-analog convertor (D/A). The digital section is usually more complex
than the simple register shown here. At each clock pulse, the A/D samples the input
waveform and converts that voltage to a binary number. At the end of each clock period
the number from the A/D is stored in the register. The output of the register is
converted back to an analog signal that has different characteristics from the original
iiq)ut signal.
SIS DSP
0001731
© 1988 Dataquest Incorporated November
DSP—Technical Overview^
Figure 1
Block Diagram of a Simple DSP System
Input
Signal
Analog-toOlgital
Converter
Raglster
Digltal-toAnalog
Cofivertsr
Output
Signal
Part A
Input
Signal
Wavsform
Sample
Pulse
Waveform
Part B
Output
Signal
Waveform
0001731-1
Source: Dataquest
November 1988
© 1988 Dataquest Incorporated November
SIS DSP
0001731
DSP—Technical Overview
Sampling Rate
Note that although the input signal, as shown in part A of Figxire 1, is a smooth
continuous signal, the output signal, as illustrated in Part B of Figure 1, looks more like a
set of stairs. With a faster sampling rate, the output signal will more closely
approximate the input. This points out the importance of the sampling rate on the
performance of the circuit. In general, a faster sampling rate is more desirable. There
are, however, both theoretical and practical constraints upon the sampling rate.
The rate at which samples are taken is called the sampling frequency, usually
expressed in megasamples per second (MSPS). From a theoretical basis, waveform
sampling must occur at least twice as fast as the highest frequency compHDnent in order
for the system to captin*e the information contained in that waveform. This is known as
the Nyquist criterion. A sampling rate that is too low will create unwanted frequency
components. A second consideration is the amount of time required for the various
circuit elements to perform their part of the operation. This places a restriction on the
maximum sample rate. A new sample cannot be taken, for example, until the
A/D convertor finishes with the previous sample.
Word Length
At each tick of the sample clock, the circuit converts the analog input signal to a
digital number. The number of bits in this number is the word length. This is important
in determining how well the output represents the desired signal. Analog signals are
continuous and have an infinite number of values, but digital signals only have a discrete
and finite number of values. The apparent stair-step effect introduces noise in the
output signal. The noise is called the quantization error. Therefore the finer the
stair-step, the lower the noise component of the output signal.
Generally there is a 6dB of signal-to-noise ratio per bit. An 8-bit word will provide
48db of signal-to-noise ratio, which is telephone-line quality. A 16-bit word will provide
96db signal-to-noise ratio, which is superb high-fidelity quality. Processing speed is
dependent upon the word length. A longer word length will require more time to perform
the signal conversions and calculations. Consequently, each digital system designer must
weigh the trade-off between speed and word length.
Distortion
There are two main reasons for unwanted signal components to be present in the
output. First, if the input contains frequencies that are faster than the Nyquist rate,
then false signals (called alias signals) will be present in the output. The solution is to
increase the sampling rate or to eliminate the undesirable component from the input
signal. Second, the stair-step-like output signal contains an infinite set of harmonics.
SIS DSP
0001731
© 1988 Dataquest Incorporated November
DSP—Technical Overview
Low-pass filters, as shown in Figure 2, are used to eliminate these unwanted
distortions. A low-pass filter placed in front of the A/D converter will eliminate the
aliasing effect. This iiput filter is called an anti-aliasing filter. Another low-pass filter
after the D/A converter will eliminate the quantization error. The addition of these two
filters eliminates the unwanted distortion from the output signal.
Figure 2
A Simple DSP System with Anti-Aliasing Filters
Low-Pasa
Flitsr
Analog-teDlgftd
Converter
Regjater
Digltal-teAnatoo
Convener
Low-Pass
Filter
Output
Signal
Output Signal
Source: Dataquest
November 1988
0001731-2
a
© 1988 Dataquest Incorporated November
SIS DSP
0001731
DSP—Technical Overview
Where it might take 10 transistors and some coils and capacitors to design an analog
filter, with DSP it might take 10,000 transistors. Thus it a^jears at first glance that
DSP is an unwarranted luxury. However, DSP has a number of advantages, and in a world
where transistors are free, DSP is steadily replacing analog.
A LOW-PASS FILTER: A PRACTICAL EXAMPLE
A block diagram for a simple DSP lov^-pass filter is shown in Figure 3. An input
register takes the sample word from the A/D converter at each sample clock. A second
register, the result register, feeds the D/A converter with the output signal. This
hardware implements a simple equation that computes an output number as a
combination of the input number and the previously computed number.
This implementation is the equivalent of a simple resistor, capacitor low-pass filter
shown in Figure 4. The output register of the DSP low-pass filter acts like the
capacitor, storing an average of all the previous samples. The constants, .3 and .7,
determine the time constants.
Figures
Block Diagram fOT a Low-Pass Filter
0.3
Input
Signal
Analea-toOlQltat
Convsrtsr
Input
-h
n«gl»t»r
Output
RtQlctar
Dlgltal-teAnalog
Convert nr
&
0.7
Source: Dataquest
November 1988
0001731-3
SIS DSP
0001731
© 1988 Dataquest Incorporated November
DSP—Technical Overview
Figure 4
Analog Implementation of a Low-Pass Filter
Input
Signal
Delay
Un«
AAA.
^VVV
»
©V*"".*
Signal
a:
0001731-4
Source: Dataquest
November 1988
The input register is tied to a multiplier that calculates the product 0.3 times the
value of the input signal and has this number available at its output at all times. For
instance, if the input register contained the number 10, then the multiplier would have
the number 3 at its output.
The result register is similarly tied to a multiplier that calculates the product
0.7 times the value in the result register and has this number available at its output. An
adder sums the two multiplier outputs and provides a new input for the result register,
thus completing the calculation for the current sample.
At each clock period the value in the input register is multiplied by 0.3 and is added
to the value in the result register multiplied by 0.7. The result of this operation is stored
in the result register, replacing the previous number. The final value in the result
register at the end of the clock period is 0.3 x input + 0.7 x result. The content of the
result register is then converted back to an analog value by the D/A converter.
10
© 1988 Dataquest Incorporated November
SIS DSP
0001731
DSP—Technical Overview
One common way to characterize the operation of a filter is to instantaneously
change the input signal from zero to a constant value and observe the behavior of the
output. Such an input is called a step function. The intermediate calculations and the
final response to this step function are shown in Figure 5. This low-pass filter circuit
smooths the waveform by averaging the input register and the result register. The
sequences of numbers illustrate this averaging process by showing the values in the input
and result registers at each sample period.
The first time period begins with the input signal equal to zero. After the A/D
conversion process, the circuit stores the resulting binary 0 in the input register. The
initial values of the registers are shown in the row labeled sample period 0. Before the
next sample period, the input changes to 100 volts. At the next sample clock, the A/D
converter puts the number 100 in the input register. During this period, the circuit
computes the value 100 x 0.3 plus 0 x 0.7 and transfers the answer to the result register.
As this process continues, the circuit generates a new output for each new value of the
input signal.
When the input register equals 100, the output register approaches 100
exponentially. Likewise, when the input register equals zero, the output register
approaches zero exponentially.
Note, however, that the output register approaches but never reaches the value
100. It a{:q}ears to stop at 97 because of rounding error. Since this circuit uses integer
arithmetic, the remainder of each multiplication operation is discarded. For instance,
.3 X 100 + .7 X 97 • 97.9. The integer value of 97 is kept while the decimal part of 0.9 is
dropped. In some applications this will c:ause unacceptable errors, so more complex
floating point operations are employed.
SIS DSP
0001731
© 1988 Dataquest Incorporated November
11
DSP—Technical Overview^
Figure 5
Step Function Re^wnse of a Digital Low-Pass Filter
Input 100Signal
Sanripj*
Puis*
Output
Signal
Flttarad
Output
Signal
*i
aock
Partod
Input
Ragiatar
RaauK
Ragiatar
Output
Ragiatar
0
1
2
3
4
5
6
7
8
9
10
11
12
.3x0
.3x100
. 3 x 100
.3x100
.3 X 100
.3 X 100
. 3 x 100
. 3 x 100
.3 X 100
. 3 x 100
. 3 x 100
. 3 x 100
.3 X 100
.7x0
.7x0
.7x30
.7x51
.7x65
.7x75
.7x82
.7x87
.7x90
.7x93
.7x95
.7x96
.7x97
0
30
51
65
75
82
87
90
93
95
96
97
97
Source: Dataquest
November 1988
0001731-6
12
© 1988 Dataquest Incorporated November
SIS DSP
0001731
DSP—Technical Overview
More complicated filters are possible with digital signal processing techniques.
Common filter applications include the Biquad filter, the Finite Impulse Response filter,
and many nonlinear filters.
The block diagram of a DSP implementation of a Biquad filter is shown in Figure 6.
It is like an analog filter with capacitors, inductors, operational amplifiers, and
feedback. The Biquad has both poles and zeros and can oscillate.
Figure 6
DSP Implem^itation of a Biquad Filter
Output
Input
0001731-
SIS DSP
0001731
Source: Daiaquest
November 1988
© 1988 Dataquest Incorporated November
13
DSP—Technical Overview
An important DSP filter is the Finite Impulse Response (FIR) filter shown in
Figure 7. One of its nice features is its stability—there is no feedback. It is ideal for
designing phase linear filters. A phase linear filter changes the amplitude of the
frequency spectrum of a signal but does not affect its phase spectrum.
The FIR filter is especially important in communications and television. For
instance, in television the phase of the subcarrier signal contains information about the
picture color. Changes in the phase relationship will distort the picture color. By using
the FIR, the composite television signal can be filtered without changing the color.
DSP can do other things that are difficult to do in the analog world. Decision
algorithms such as pitch extraction, data compression, and speech compression are not
simple filtering; they resemble computer programs with flowcharts and decision points,
A DSP processor can do more than filtering; it can act like a computer by performing
logical decisions.
Figure?
Finite Impulse Response Filter
Output
Constant /
\
Conitant /
\ Constant A
^
^
Unit
D«lay
^
Untt
"^ Dslay
^ Constant
Constant
X
Input
UnK
Ddsy
Unit
Dslay
Source: Dataquest
November 1988
0001731-7
14
- ^ oto.
© 1988 Dataquest Incorporated November
SIS DSP
0001731
"
-
^
-
*
'
-
^
• • • ^
. .
^
•
.
.
.
Potential Users
Automatic Call Distributors (ACDs)
COMPANY DIRECTORY
AT&T Information Systems
100 Southgate Parkway
Morristown, NJ 07960
Ring Communications
35 Pinelawn Road
Melville, NY 11747
Automation Electronics Corporation
344_40th Street
Department A
Oakland, CA 94609
Rockwell International
P.O. Box 10462
Dallas, TX 75207
ROLM Corporation
4900 Old Ironsides Drive
Santa Clara, CA 95050
DND Teletronics, Inc.
235 Elizabeth Street
Utica, NY 13501
Dacon Electronics, Inc.
7 Industrial Avenue
Upper Saddle River, NJ 07458
Digital Transmission, Inc.
315 Eisenhower Lane South
Lombard, IL 60148
Fujitsu Systems of America, Inc.
12670 High Bluff Drive
San Diego, CA 92130
GTE Switching & Telephone Products
One Stamford Forum
Stamford, CT 06904
Solid-state Systems
1990 Delk Industrial Blvd.
Marietta, GA 30067
Teknekron/Infoswitch
1784 Firman Drive
Richardson, TX 75081
Telcom Technologies
3072 East "G" Street
Ontario, CA 91764
Telephonic Equipment Corporation
1704 Armstrong Avenue
Irvine, CA 92714
Isotek Communications, Inc.
6 Thorndal Circle
Darien, CN 06820
SIS DSP
© 1987 Dataquest Incorporated June
Directory of Potential Users
CAD/CAM Equipment
COMPAMY DIRECTORY
IBM Corporation
Armonk, NY 10504
Adage, Inc.
One Fortune Drive
Billerica, MA 01821
Lexidata Corporation
755 Middlesex Turnpike
Billerica, MA 01865
Applicon, Inc.
4251 Plymouth Road
Ann Arbor, MI 48106
The MacNeal-Schwendler Corp.
815 Colorado Boulevard
Los Angeles, CA 90041
Auto-trol Technology
Corporation
Denver, CO 80233
CADLINC, Inc.
700 Nicholas Boulevard
Elk Grove, IL 60007
McDonnel Douglas Corporation
Automation Company
P. 0. Box 516
St. Louis, MO 63166
CADNETIX Corporation
5757 Central Avenue
Boulder, CO 80301
Mentor Graphics Corporation
8500 S.W. Creekside Place
Beaverton, OR 97005
Calay Systems, Inc.
Irvine, CA
PDA Engineering Inc.
1560 Brookhollow Drive
Santa Ana, CA 92705
Calma Company
501 Sycamore Drive
Milpitas, CA 95035
Personal CAD Systems, Inc.
981 University Avenue
Los Gatos, CA 95030
Computervision Corporation
201 Burlington Road
Bedford, MA 01730
Prime Computer, Inc.
Prime Park
Natick, MA 01760
Control Data Corporation
8100 34th Avenue South
Minneapolis, MN 55420
Racal-Redac Limited
Tewkesbury
Gloucestershire, England
GL20 8HE
Daisy Systems Corporation
700 Middlefield Road
Mountain View, CA 94039
Scientific Calculations, Inc.
7635 Main Street
Fishers, NY 14453
Hewlett-Packard Company
1501 Page Mill Road
Palo Alto, CA 94304
Intergraph Corporation
One Madison Industrial Park
Huntsville, AL 35807
SIS DSP
Silvar-Lisco
1080 Marsh Road
Menlo Park, CA 94025
Synercom
P. 0. Box 27
Sugarland, TX
1987 Dataquest Incorporated June
7747 8
Directory of Potential Users
CAD/CAM Equipment
COMPANY DIRECTORY (Continued)
Tektronix, Inc.
4900 S. W. Griffith Drive
Beaverton, OR 97077
Telesis Systems Corporation
21 Alpha Road
Chelmsford, MA 01824
Valid Logic Systems, Inc.
2820 Orchard Parkway
San Jose, CA 95134
© 1987 Dataquest Incorporated June
SIS DSP
Directory of Potential Users
Carrier Equipment
COMPAMY DIRECTORY
ADC Teleconmiunications, Inc.
4900 W 78 Street
Minneapolis, MN 55435
Pulsecom
2900 Towerview Road
Herndon, VA 22071
American Telephone
and Telegraph
550 Madison Avenue
New York, NY 10022
R-TEC
2100 Reliance Parkway
P.O. Box 919
Bedford, TX 76021
Ericsson Communications
7465 Lampson Avenue
Garden Grove, CA 92641
Rockwell Telecommunications, Inc.
Wiscom Telephone Products Division
8245 Lemont Road
Downers Grove, IL 60515
GTE Communications Systems
2800 Utopia Road
Phoenix, AZ 85027
International Telephone
and Telegraph
Network Systems Division
3100 Highwoods Boulevard
Raleigh, NC 27604
San Bar Corporation
Transmission Systems Division
9999 Muirlands Parkway
Irvine, CA 92718
Seiscor, Inc.
P.O. Box 1590
Tulsa, OK 74102
Lynch Communication Systems, Inc.
204 Edison Way
Reno, NV 89520
Tellabs, Inc.
4951 Indiana Avneue
Lisle, IL 60532
NEC America, Inc.
Radio & Transmission Division
2740 Prosperity Avenue
Fairfax, VA 22031
Transcom Electronics, Inc.
1170 E. Maine Road
Portsmouth, RI 02871
Northern Telecom, Inc.
1555 Roadhaven Drive
Stone Mountain, GA 30083
SIS DSP
© 1987 Dataquest Incorporated June
Directory of Potential Users
Carrier Equipment
(Page intentionally left blank)
1987 Dataguest Incorporated June
SIS DSP
Directory of Potential Users
CAT Scanners
COMPANY DIRECTORY
DiasonicS/ Inc.
Milpitas, CA
Siemens
Iselin, NJ
Toshiba Medical Systems
Tustin, CA
General Electric Medical
Systems
Milwaukee, WI
Picker International
Highland Heights, OH
SIS DSP
© 1987 Dataquest Incorporated June
Directory of Potential Users
CAT Scanners
(Page intentionally left blank)
1987 Dataguest Incorporated June
SIS DSP
Directory of Potential Users
Cellular Mobile Radio
COMPANY DIRECTORY
Astronet Corporation
400 Rinehart Road
Lake Mary, FL 32746
Hitachi America, Ltd.
2990 Gateway Dr., Suite 1000
Norcross/Atlanta, GA 30071
AT&T Technologies, Inc.
3800 Golf Road
Rolling Meadows, IL 60008
ITT Telecon
3100 Highwoods Blvd.
Raleigh, NC 27604
CTI, Inc.
P.O. Box 71, Hwy. 45
South Corinth, MI 38834
Kokusai Electronic Company Ltd.
America
363 Coral Circle
El Segundo, CA 90245
E.F. Johnson Company
299 Johnson Avenue
Waseca, MN 56093
Ericsson Radio Systems
730 International Parkway
Richardson, TX 75081
Fujitsu America, Inc.
10 East 53rd Street
New York, NY 10022
GTE Communication System
Dept. 582/A6
400 North Wolf Road
Northlake, IL 60164
General Electric Company
Mountain View Road
Lynchburg, VA 24502
Glenayre Electronics
12 Pacific Highway
Blaine, WA 98230
Harris Corp.
RF Communications Group
1680 University Avenue
Rochester, NY 14610
SIS DSP
Mitsubishi International
Corporation
879 Systems Drive
Bewsenville, IL 60106
Northern Telecom, Inc.
1201 East Arapaho Road
Richardson, TX 75081
NovAtel Communications
2820 Peterson Place
Norcross, GA 30071
OKI Advanced Communications
1 University Plaza
Hackensack, NJ 07601
Panasonic Industrial Company
Telecommunications Division
1 Panasonic Way
Secaucus, NJ 07094
Quintron Corporation
1 Quintron Way
P.O. Box 3726
Quincy, IL 62305
© 1987 Dataquest Incorporated June
Directory of Potential Users
Cellular Mobile Radio
(Page intentionally left blank)
1987 Dataquest Incorporated June
SIS DSP
Directory of Potential Users
Central Office Switching Equipment
COMPANY DIRECTORY
AT&T Network Systems
475 South Street
Morristown, NJ 07960
Northern Telecom, Inc.
4001 E. Chapel Hill-Nelson Hwy.
Research Triangle Park, NC 27709
CIT-Alcatel, Inc.
10800 Parkridge Blvd.
Reston, VA 22091
Stromberg-Carlson
400 Rinehart Rd.
Lake Mary, FL 32746
GTE Communications Systems
2500 W. Utopia Rd.
Phoenix, AZ 85027
NEC America, Inc.
1525 Walnut Hill Lane
Irving, TX 75062
ITT Telecom
Network Systems Division
3100 Highwoods Blvd.
Raleigh, NC 27604
SIS DSP
© 1987 Dataquest Incorporated June
Directory of Potential Users
Central Office Switching Equipment
(Page intentionally left blank)
1987 Dataquest Incorporated June
SIS DSP
Directory of Potential Users
Data Network Control Systems
COMPANY DIRECTORY
Accurate Electronics, Inc.
215 North Avenue
Bridgeport, CT 06606
Case Communications
2120 Industrial Parkway
Silver Spring, MD 20904
ADC
4900 W. 78th Street
Minneapolis, MN 55435
Codex Corporation
20 Cabot Blvd.
Mansfield, MA 02048
Advanced Data Support Systems
717 Mt. Vernon Ave., Suite 1
Bakersfield, CA 93307
Connections Telecommunications
130 Steamboat Rd.
Great Neck, NY 11024
Amdahl
2200 N. Greenville
Richardson, TX 75081
Dataswitch Corporation
444 Westport Ave.
Norwalk, CT 06851
Anritsu
128 Bauer
Oakland, NJ
Datacom Northwest
3303 112th St.
Everett, WA 98204
07436
AstroCom Corporation
120 W. Plato
St. Paul, MN 55107
Datacomm Management Sciences
25 Van Zant St.
East Norwalk, CT 06855-1790
AT&T Communications
202/206 North
Bedminster, NJ 07921
Datagram
11 Main St.
East Greenwich, RI
AT&T Information Systems
One Exchange Plaza
New York, NY 10006
Datatel
81 N. Oak Lane
Cherry Hill, NJ
Atlantic Research
7401 Boston Blvd.
Springfield, VA 22153
Digilog Inc.
1370 Welsh Rd.
Montgomeryville, PA
Avant-Garde Computing
8000 Commerce Parkway
Mount Laurel, NJ 08054
Digitech Industries
66 Grove St.
Ridgefield, CT 06877
Bytex Corporation
120 Turnpike Rd.
Southborough, MA 01776
DMW Group
2070 Hogback Rd.
Ann Arbor, MI 48104
C-Cor Electronics
1400 N.W. Compton Dr.
Beaverton, OR 97006
Domain Systems
241 Central Avenue
East Farmingdale, NY
SIS DSP
1987 Dataquest Incorporated June
02818
08003
18936
11735
Directory of Potential Users
Data Network Control Systems
COMPAMY DIRECTORY (Continued)
Dynatech Data Systems
7644 Dynatech Ct.
Springfield, VA 22153
Infinet
6 Shattuck Rd.
Andover, MA 01810
Electrodata, Inc.
23070 Miles Rd.
Bedford Heights, OH
Infotron Systems Corporation
Cherry Hill Industrial Center
9 N. Olay Ave.
Cherry Hill, NJ 08003
44128
Emcom Corporation
101 E. Park Blvd.
Piano, TX 75074
LPCom
21020 Homestead Rd.
Cupertino, CA 95014
Foremark Technologies
550 Old Country Rd.
Hicksville, NY 11801
Navtel
3331 N.W. 55th St.
Ft. Lauderdale, FL
General Datacomm
Route 63
Middlebury, CT 06762-1298
Giltronix
3780 Fabian Way
Palo Alto, CA 94303
70877
Hewlett-Packard
P.O. Box 7050
Colorado Springs, CO
NCC
9600 W. 76th St.
Minneapolis, MN 55344
NEC America
110 Rio Robles
San Jose, CA 95134
Hard Engineering, Inc.
2804 B Memorial Parkway
Huntsville, NY 35801
Hekimian Labs
9298 Gaither Rd.
Gaithersburg, MD
33309
80933
Halcyon
2121 Zanker Rd.
San Jose, CA 95131
IBM Corporation
1133 Westchester Ave.
White Plains, NY 10604
Idacom Electronics
Research Center One
9411-20th Ave.
Edmonton, Alberta, Canada T6W 1E5
Northern Telecom
Spectron Division
8000 Lincoln Dr.
Marlton, NJ 08053
NU Data
32 Fairview Ave.
Little Silver, NJ
07739
Ocean Data Systems
6000 Executive Blvd.
Rockville, MD 20852
Paradyne Corporation
8550 Ulmerton Rd.
Largo, FL 33541
Penril Datacomm
207 Perry Parkway
Gaithersberg, MD 20877-2197
© 1987 Dataquest Incorporated June
SIS DSP
Directory of Potential Users
Data Network Control Systems
COMPANY DIRECTORY (Continued)
Plantronics Wilcom
P.O. Box 508
Laconia, NH 03247
Questronics
3570 South W. Temple
Salt Lake City, UT 84115
Racal-Milgo
1601 N. Harrison Parkway
Sunrise, FL 33323
Telenex
502 Pleasant Valley Ave,
Moorestown, NJ 08057
Teleprocessing Products
4565 E. Industrial St.
Simi Valley, CA 93036
Tymnet
2710 Orchard Parkway
San Jose, CA 95014
VIR
2255 Pioneer Rd.
Hatboro, PA 19040
Symplex
2002 Hogback Road
Ann Arbor, MI 48107
T-Bar, Inc.
141 Danbury Rd.
Wilton, CT 06847
Tekelec, Inc.
26540 Agoura Rd.
Calabazas, CA 91302
Tektronix
1800 S.W. Merlord
Beaverton, OR 97075
Telecommunications Technologies
444 N. Fredrick Ave.
Gaithersburg, MD 20877
SIS DSP
© 1987 Dataquest Incorporated June
Directory of Potential Users
Data Network Control Systems
(Page intentionally left blank)
1987 Dataquest Incorporated June
SIS DSP
Directory of Potential Users
Data PBX
CrogAMY DIRECTORY
ATT Network Systems
475 South Street
Morristown, NJ 07960
Infotron Systems Corporation
Cherry Hill Industrial Center
Cherry Hill, NJ 08003
Bridge Communications
2081 Stierlin Road
Mountain View, CA 94043
Intelligent Business Communications
80 Oser Avenue
Hauppauge, NY 11787
CASE Communications
8310 Guilford Road
Columbia, MD 21046
M/A-COM Linkabit, Inc.
3033 Science Park Road
San Diego, CA 92121
Codex Corporation
20 Cabot Boulevard
Mansfield, MA 02048
METAPATH, Inc.
727 Lincoln Center Drive
Foster City, CA 94404
ComDesign Inc.
751 S. Kellogg Avenue
Goleta, CA 93117
Micom Systems, Inc.
20151 Nordhoff Street
Chatsworth, CA 91311
Commtex Inc.
24511 Crofton Lane
Crofton, MD 21114
Network Products
Research Triangle Park, NC
Develcon Electronics
744 Nina Way
Warminister, PA 18974
Equinox Systems, Inc.
12041 S.W. 144th Street
Miami, FL 33186-6108
Gandalf Data, Inc.
1019 S. Noel Avenue
Wheeling, IL 60090
SIS DSP
Sequal Data
975 Walnut Street
Cary, NC 27511
TELLABS, Inc.
4951 Indiana Avenue
Lisle, IL 60532
Western Telematic Inc.
2435 S. Anne Street
Santa Ana, CA 92704
© 1987 Dataquest Incorporated June
27709
Directory of Potential Users
Data PBX
(Page intentionally left blank)
© 1987 Dataquest Incorporated June
SIS DSP
Directory of Potential Users
Digital Radiography
COMPAMY DIRECTORY
Siemens
Iselin, NJ
ADAC Laboratories
Sunnyvale, CA
Squibb Corporation
Princeton, NJ
Andersen Group, Inc.
Bloomfield, CT
Technicare Corporation
Solon, OH
Diasonics, Inc.
Milpitas, CA
Toshiba Medical Systems
Tustin, CA
Picker International
Highland Heights, OH
Xonics
Des Plaines, IL
Philips Medical Systems, Inc.
She1ton, CT
Raytheon Medical Systems
Lexington, MA
SIS DSP
© 1987 Dataquest Incorporated June
Directory of Potential Users
Digital Radiography
(Page intentionally left blank)
© 1987 Dataguest Incorporated June
SIS DSP
Directory of Potential Users
Facsimile Equipment
COMPANY DIRECTORY
Murata Business Systems
4801 Spring Valley Road
Suite 108B
Dallas, TX 75240
Adler-Royal Business
Machines, Inc.
1600 Route 22
P.O. Box 1597
Union, NJ 07083
NEC America, Inc.
8 Old Sod Farm Road
Melville, NY 11747
Alden Electronics, Inc.
Washington Street
Westborough, MA 01581
Panafax Corporation
10 Melville Park Road
Melville, NY 11747
Amacom
5115 Calvert Road
College Park, MD 20740
Pitney Bowes
1515 Summer Street
Stamford, CT 06926
AT&T Information Systems
One Speedwell Avenue
Morristown, NJ 07960
Fujitsu Imaging Systems of
America, Inc.
Corporate Drive, Commerce Park
Danbury, CT 06810
Canon U.S.A., Inc.
One Canon Plaza
Lake Success, NY 11042
Teleautograph Corporation
8700 Bellanca Avenue
Los Angeles, CA 90045
Minolta Corporation
101 Williams Drive
Ramsey, NJ 07446
Mitsubishi Electric Sales
of America, Inc.
5757 Plaza Drive
Cypress, CA 90630
SIS DSP
Sanyo Business Systems
Corporation
51 Joseph Street
Moonachie, NJ 07074
Sharp Electronics Corporation
10 Sharp Plaza
Paramus, NJ 07652
Hitachi America Ltd.
2990 Gateway Drive
Suite 1000
Norcross, GA 30071
Muirhead Inc.
1101 Bristol Road
Mountainside, NJ 07092
Ricoh Corporation
7 Kingsbridge Road
Fairfield, NY 07006
Harris/3M Business Products, Inc.
2300 Parklake Drive, N.E.
Atlanta, GA 30345
Toshiba America
2441 Michelle Drive
Tustin, CA 92680
Xerox Corporation
Information Products Division
1341 West Mockingbird Lane
Dallas, TX 75247
© 1987 Dataquest Incorporated June
Directory of Potential Users
Facsimile Equipment
(Page intentionally left blank)
© 1987 Dataquest Incorporated June
SIS DSP
Directory of Potential Users
Front-End Processors
COMPAMY DIRECTORY
Amdahl Corporation
1250 E. Argues Avenue
P.O. Box 3470
Sunnyvale, CA 94088
IBM Corporation
P.O. Box 12195
Research Triangle Park, NC
Lemcon Systems, Inc.
2104 W. Peoria
Phoenix, A2 85029
Burroughs Corporation
Burroughs Place
Detroit, MI 48232
NCR Comten, Inc.
1700 S. Patterson Blvd.
Dayton, OH 45479
CHI Corporation
26055 Emery Road
Cleveland, OH 44128
Computer Conununications, Inc.
2610 Columbia Street
Torrance, CA 90503
Control Data Corporation
8100 34th Avenue, South
Minneapolis, MN 55440
Periphonics
4000 Veterans Memorial Hwy.
Bohemia, NY 11716
Sperry Corporation
Computer Systems Operations
P.O. Box 500
Blue Bell, PA 19424
Honeywell Information
Systems, Inc.
Honeywell Plaza
Minneapolis, MN 55408
SIS DSP
© 1987 Dataquest Incorporated June
27709
Directory of Potential Users
Front-End Processors
(Page intentionally left blank)
© 1987 Dataquest Incorporated June
SIS DSP
Directory of Potential Users
Graphics Systems
COMPAMY DIRECTORY
Adage/ Inc.
One Fortune Drive
Billerica, MA 01821
CIE Terminals, Inc.
2 505 McCabe Way
Irvine, CA 97214
Digital Equipment Corporation
146 Main Street
Maynard, MA 017 54
Hewlett-Packard Company
3000 Hanover Street
Palo Alto, CA 94304
Intecolor Corporation
225 Technology Park/Atlanta
Norcross, GA 30092
ITT Qume Corporation
2350 Qume Drive
San Jose, CA 95131
Tectronix, Inc.
P. 0. Box 1700
Beaverton, OR 97075
SIS DSP
© 1987 Dataguest Incorporated June
Directory of Potential Users
Graphics Systems
(Page intentionally left blank)
1987 Dataquest Incorporated June
SIS DSP
Directory of Potential Users
Integrated Voice/Data Workstations
COMPAMY DIRECTORY
AMBI Corporation
1033 Washington Blvd.
Stamford, CT 06901
Mitel Corporation
5400 Broken Sound Blvd.
Boca Raton, FL 33431
AT&T Information Systems
1766 on The Green
Room 48-5A38
Morristown, NJ 07960
Northern Telecom, Inc.
9705 Data Park
Minnetonka, MN 55440
Basic Telecommunications Corp.
4414 East Harmony Road
Fort Collins, CO 80525
ROLM Corporation
2420 Ridgeport Drive
Austin, TX 78754
Sydis, Inc.
4340 Stevens Creek Blvd.
San Jose, CA 95129
Cygnet Technologies
12 Lawrence Station Road
Sunnyvale, CA 94089
Telerad Telecommunications and
Electronic Industries, Ltd.
P.O. Box 50
Lod, Israel 71100
DAVOX Corporation
4 Federal Street
Billerica, MA 01821
GEC Information Systems, Ltd.
P.O. Box 6
Coventry CVl 5PU
England
Thomson CSF
2 Gannett Drive
White Plains, NY
Zaisan
13910 Champion Forest Drive
Houston, TX 77069
Liberty Electronics
625 Third Street
San Francisco, CA 94107
Matra Communications, Inc.
1202 Charleston Road
Mountain View, CA 94043
SIS DSP
10604
© 1987 Dataquest Incorporated June
Directoiy of Potential Users
Integrated Voice/Data Workstations
(Page intentionally left blank)
© 1987 Dataquest Incorporated June
SIS DSP
Directory of Potential Users
Key Telephone Systems
COMPAMY DIRECTORY
ADAX
6961 Peachtree Industrial Blvd.
Norcross, GA 30071
Ericsson
P.O. Box 3110
Greenwich, CT
AEL Microtel, Ltd.
100 Strowger Blvd.
Brockville, Ontario K6V 5WB
Canada
EXTROM Communications
137 Express Street
Plainview, NY 11803
06832
AT&T Technologies, Inc.
222 Broadway
New York, NY 10017
Fujitsu Business Communications,
Inc.
3190 Mira Loma Avenue
Anaheim, CA 92806
Astrocom Division, QCT, Ltd.
1702 South Del Mar Ave.
San Gabriel, CA 91776
GAI-Tronics
P.O. Box 31
Reading, PA 19603
ASUZI, Inc.
310 Frontage Road
Greer, SC 29651
GTE Business Communications
Systems, Inc.
8301 Greensboro Drive
McLean, VA 22102
Code-a-Phone
P.O. Box 5656
Portland, OR 97228
Goldstar Telecommunications, Inc.
One Madison Street
East Rutherford, NJ 07073
Comdial Telephone Systems, Inc.
1180 Seminole Trail
Charlottesville, VA 22906
Crest Industries, Inc.
6922 North Meridian
Puyallup, WA 98371
Harris/Lanier
1700 Chantilly Drive N.E.
Atlanta, GA 30324
Hitachi America
2696 Peachtree Square
Doraville, GA 30360
DBA Communication Systems, Inc.
339 West 2nd Street
N. Vancouver, B.C. V7M 1E2
Canada
ITT Telecom
3100 Highwoods Blvd.
Raleigh, NC 27604
DEKA, Inc.
16555 Shannon Road
Los Gatos, CA 95030
In Electronic Inc.
667 Madison Ave., Suite 800
New York, NY 10021
Eagle Telephonies, Inc.
375 Oser Avenue
Hauppauge, NY 11788
Information Dynamics
1251 Exchange Drive
Richardson, TX 75081
Emulex Corporation
3545 Harbor Blvd.
Costa Mesa, CA 92626
Integrated Telecomputing Systems
1153 Bordeaux Dr., Ste. 107
Sunnyvale, CA 94086
SIS DSP
© 1987 Dataquest Incorporated June
Directory of Potential Users
Key Telephone Systems
COMPANY DIRECTORY (Continued)
Inter-Tel Equipment, Inc.
3232 West Virginia Ave.
Phoenix, AZ 85009
Multitel
505 North Lake Shore Drive
Chicago, IL 60611
Interconnect Planning Corp.
One Lafayette Place
Greenwich, CT 06830
NEC Telephones, Inc.
8 Old Sod Farm Road
Melville, NY 11747
Isoetec
6 Thorndal Circle
Darien, CT 06970
Nicho
8660 Troy Twp. Rd. #4 RR9
Mansfield, OH 44904
Iwatsu America, Inc.
430 Commerce Blvd.
Carlstadt, NJ 07072
Northeom
600 Industrial Parkway
Industrial Airport, KS
Jackson Associates
505 North Lake Shore Drive
Chicago, IL 60611
Northern
Business
2916 5th
Calgary,
Canada
Kanda Telecom, Inc.
11130 Metric Blvd.
Austin, TX 78758
Kanda Tsushin Kogyo
611 West 6th Street
Los Angeles, CA 90017
Ma Best Telephone Products
P.O. Box 4522
North Hollywood, CA 91607
Melco Labs
14408 N.E. 20th
Bellevue, WA 98007
Micro-Z
900 South Magnolia Avenue
Monrovia, CA 91016
Mitel
5400 Broken Sound Blvd.
Boca Raton, FL 33431
Model Rectifier Corp.
2500 Woodbridge Ave.
Edison, NJ 08817
66031
Telecom, Ltd.
Products Division
Ave., N.E.
Alberta T2A 6K4
Oki Electronics of America
4031 N.E. 12th Terrace
Fort Lauderdale, FL 33308
PKS Communications
46 Quirk Road
Milford, CT 06460
Panasonic
One Panasonic Way, Box 1503
Secaucus, NJ 07094
Philips Communications
89 Marcus Blvd.
Hauppauge, NY 11788
Philips KBX Systems
85 McKee Road
Mahwah, NJ 07430
Plant Equipment Inc.
28075 Diaz Rd.
Temecula, CA 92390
© 1987 Dataquest Incorporated June
SIS DSP
Directory of Potential Users
Key Telephone Systems
COMPANY DIRECTORY (Continued)
Plessey Conununications
235 Yorkland Blvd.
Willowdale, Ontario
Canada
Tecom
8134 Zionville Road
Indianapolis, IN 46268
Proctor & Associates
15050 Northeast 36th Street
Redmond, WA 98052
Telefony and Electronics
of America, Inc.
8675 N.W. 56th St.
Miami, FL 33166
R-TEC Systems
2100 Reliance Parkway
Bedford, TX 76021
Tel-Path Industries
3361 Melrose Ave.
Roanoke, VA 24017
SAN/BAR Corp.
Telephone Systems Division
2405 South Shiloh Road
Garland, TX 75041
TelRad
42-15 Crescent Street
Long Island City, NY 11101
TELEDEX
4051 Burton Drive
Santa Clara, CA 95050
Sanyo Business Systems
51 Joseph Street
Moonachie, NJ 07074
Telephonic Equipment
17401 Armstrong Avenue
Irvine, CA 92714
Scott Technologies Corp.
Foot of Broad Street
Stratford, CT 06497
Solid State Systems
1990 Delk Industrial Blvd.
Marietta, GA 30067
TIE/Communications, Inc.
5 Research Drive
Shelton, CT 06484
Teletec Systems
1380 Old Freeport Road
Pittsburgh, PA 15238
Teltone Corp.
10801-120th Avenue, N.E.
Kirkland, WA 98033
Teltrend Inc.
P.O. Box 400, Dept. lOlA
West Chicago, IL 60185
TT Systems Corp.
9 East 37th Street
New York, NY 10016
Tadiran Electronic Industries
10901 Endeavour Way, Ste. A
Largo, FL 33543
Technicom International
23 Old Kings Highway South
Darien, CT 06820
Thomson-CSF
Two Gannet Drive
White Plains, NY
10604
Tone Commander Systems
4320 150th Avenue N.E.
Redmond, WA 98052
Toshiba America
2441 Michelle Drive
Tustin, CA 92680
SIS DSP
© 1987 Dataquest Incorporated June
Directory of Potential Users
Key Telephone Systems
COMPANY DIRECTORY (Continued)
Touch Com
253 North Grand Avenue
Poughkeepsie, NY 12603
Trillium Telephone Systems
603 March Road
Kanata
Canada K2K 1X3
Turret Equipment Corp. (TEC)
880 3rd Avenue
New York, NY 10022
Tymetek
770 Church Road
Elmshurst, IL 60126
V Band Systems
345 Hudson Street
New York, NiT 10014
Valcom
1111 Production Street
Roanoke, VA 24013
Viking Electronics
P.O. Box 4522
North Hollywood, CA
91607
Vodavi
8300 E. Raintree Drive
Scottsdale, AZ 85260
Walker Communications Corp.
200 Oser Avenue
Hauppauge, NY 11788
WEBCOR Electronics
28 South Terminal Drive
Plainview, NY 11803
Wren Company c/o Nicho
880 Reynard Street
Cincinnati, OH 45231
XTEL
1301 Cornell Parkway
Oklahoma City, OK 73108
© 1987 Dataguest Incorporated June
SIS DSP
Directory of Potential Users
Local Area Network Equipment
COMPAMY DIRECTORY
Acorn Computers
Fulbourn Road
Cherry Hinton
Cambridge CBl 4JN
England
Astra Communications, Inc.
329 North Bernardo
Mountain View, CA 94043
Advanced Computer Communications
720 Santa Barbara Street
Santa Barbara, CA 93101
Allen-Bradley Co., Inc.
555 Briarwood Circle
Ann Arbor, MI 48104
Allied Data Communications Group
5375 Oakbrook Parkway
Norcross, GA 30093
American Photonics
71 Commerce Drive
Brookfield Center, CT
06805
AMP Incorporated
P.O. Box 3608
Harrisburg, PA 17105
AT&T Information Systems
Crawford Corner Road
Holmdel, NJ 07737
Avatar Technologies, Inc.
99 South Street
Hopkinton, MA 01748
Banyan Systems, Inc.
135 Flanders Road
Westboro, MA 01581
BICC Data Networks
1800 West Park Drive
Westborough, MA 01581
Bridge Communications, Inc.
2081 Stierlin Road
Mountain View, CA 94043
Apollo Computer, Inc.
330 Billerica Road
Chelmsford, MA 01824
Cabletron
2514 Seaboard Avenue
San Jose, CA 95131
Apple Computer, Inc.
20525 Mariani Avenue
Cupertino, CA 95014
Applied Knowledge Groups
1095 E. Duane Street, Suite 203
Sunnyvale, CA 94086
Applitek Corporation
107 Audubon Road
Wakefield, MA 01880
Artel Communications Corporation
P.O. Box 100, West Side Station
Worchester, MA 01602
SIS DSP
AST Research
2121 Alton Avenue
Irvine, CA 92714
Chipcom Corporation
195 Bear Hill Road
Waltham, MA 02154
Codenoll Technology
1086 N. Broadway
Yonkers, NY 10701
Codex (a subdivision of Motorola)
20 Cabot Boulevarcl
Mansfield, MA 02048
© 1987 Dataquest Incorporated June
Directory of Potential Users
Local Area Network Equipment
COMPANY DIRECTORY (Continued)
ComDesign, Inc.
751 South Kellog Avenue
Goleta, CA 93117
DAVID Systems, Inc.
701 East Evelyn Avenue
Sunnyvale, CA 94086
Communication Machinery Corporation
1421 State Street
Santa Barbara, CA 93101
The Destek Group
830 East Evelyn Avenue
Sunnyvale, CA 94086
Complex^ Systems, Inc.
4930 Research Drive
Huntsville, AL 35805
Digital Communications Assoc.
1000 Alderman Drive
Alpharetta, GA 30201
Computer Pathways
19102 North Creek Parkway
Bothell, WA 98011
Digital Equipment Corporation
1925 Andover Street
Tewksbury, MA 01876
Concord Communications, Inc.
397 Williams Street
Marlborough, MA 01752
Digital Products, Inc.
108 Water Street
Watertown, MA 02172
Contel Business Networks
4330 East-West Highway
Bethesda, MD 20814
Excelan, Inc.
2180 Fortune Drive
San Jose, CA 95131
Control Data Corporation
8100 34th Avenue South
Minneapolis, MN 55440
Fairchild Data Corporation
350 N. Hayden Road
Scottsdale, AZ 85257
Convergent Technologies, Inc.
2700 N. First Street
San Jose, CA 95150
Fast Feedback Technologies
1505 Aviation Boulevard
Redondo Beach, CA 90278
Corvus Systems, Inc.
2100 Corvus Drive
San Jose, CA 95124
FiberCom, Inc.
3353 Orange Avenue, N.E.
Roanoke, VA 24012
Data General Corporation
4400 Computer Drive
Westboro, MA 01580
FiberLAN
P.O. Box 12726
Research Triangle Park, NC
Datapoint
9725 Datapoint Drive
San Antonio, TX 78284
Fibronics
325 Stevens Street
Hyannis, MA 02601
1987 Dataquest Incorporated June
27709
SIS DSP
Directory of Potential Users
Local Area Network Equipment
COMPAMY DIRECTORY (Continued)
Fox Research
7005 Corporate Way
Dayton, OH 45459
ITT Information Systems
2350 Qume Drive
San Jose, CA 95131
Gandalf Data, Inc.
1019 South Noel Avenue
Wheeling, IL 60090
Kee Incorporated
10727 Tucker Street
Beltsville, MD 20705
Gateway Communications, Inc.
2941 Alton Avenue
Irvine, CA 92714
Kimtron
1709 Junction Court, Bldg. 380
San Jose, CA 95112-1090
Hewlett-Packard Company
5725 West Las Positas
Pleasanton, CA 94566
KMW-Auscom
8307 Highway 71 West
Austin, TX 78735
Honeywell, Inc.
P.O. Box 8000
Phoenix, AZ 85066
Lancore Technologies, Inc.
31324 Via Colinas, #110
Westlake Village, CA 91362
IBM Corporation
Research Traingle Park, NC
27709
IBM Entry Systems Division
1000 NW 51st Street
Boca Raton, FL 33432
IdeAssociates, Inc.
29 Dunham Road
Billercia, MA 01821
Metapath
222 Lincoln Center Drive
Foster City, CA 94404
Micom/Interlan
155 Swanson Road
Boxborough, MA 01719
Infotron Systems
Cherry Hill Industrial Center
Cherry Hill, NJ 08003
Intel Corporation
3065 Bowers Avenue
Santa Clara, CA 95051
Microsoft
16011 N.E. 36th Way
Redmond, WA 98073
Motorola/Four-Phase Systems
10700 N. De Anza Boulevard
Cupertino, CA 95014
Intel Corporation
5000 West Chandler Boulevard
Chandler, AZ 85226
SIS DSP
LanTel Corporation
3100 Northwoods Place, Suite A
Norcross, GA 30071
National Semiconductor
2900 Semiconductor Drive
Santa Clara, CA 95051
© 1987 Dataquest Incorporated June
Directory of Potential Users
Local Area Network Equipment
COMPAMY DIRECTORY (Continued)
NCR Corporation
1700 So. Patterson Boulevard
Dayton, OH 45479
Prime Computer Inc.
Prime Park
Natick, MA 01760
Nestar Systems, Inc.
1345 Shorebird Way
Mountain View, CA 94043
Proteon
Two Technology Drive
Westborough, MA 01581-5008
Netlink
3214 Spring Forest Road
Raleigh, NC 27604
Quadram Corporation
4355 International Boulevard
Norcross, GA 30093
Netronix
1372 North McDowell Boulevard
Petaluma, CA 94952
Racal-Milgo
8600 N.W. 41st Street
Miami, FL 33166
Network General Corporation
1296B Lawrence Station Road
Sunnyvale, CA 94089
Retix
1547 Ninth Street
Santa Monica, CA 90401
Network Research Corporation
2380 North Rose Avenue
Oxnard, CA 93030
Santa Cruz Operation
400 Encinal Street
Santa Cruz, CA 95061
Network Systems Corporation
7600 Boone Avenue North
Brooklyn Park, MN 55428
Server Technology Inc.
1095 E. Duane Street, Suite 103
Sunnyvale, CA 94086
Novell, Inc.
748 North 1340 West
Orem, UT 84057
Siecor Corporation
P.O. Box 13625
Research Triangle Park, NC
Orchid Technology, Inc.
47790 Westinghouse Drive
Fremont, CA 94539
Siemens Information Systems Group
5500 Broken Sound Boulevard
Boca Raton, FL 33431
Paradyne Corporation
8550 Ulmerton Road
Largo, FL 33541
The Software Link, Inc.
8601 Dunwoody Place, Suite 632
Atlanta, GA 30338
Phoenix Digital
2315 North 35th Avenue
Phoenix, AZ 85009
Standard Microsystems Corporation
35 Marcus Boulevard
Hauppauge, NY 11788
27709
Phoenix Technology Inc.
2803 Bunker Hill Drive
Santa Clara, CA 95054
1987 Dataquest Incorporated June
SIS DSP
Directory of Potential Users
Local Area Network Equipment
CroCAMY DIRECTORY (Continued)
Sterling Conununications Corporation
10320 Little Patuxent Parkway
Suite 808
Columbia, MD 21044
Vitalink
1350 Charleston Road
Mountain View, CA 94043
Wang Laboratories, Inc.
One Industrial Way
Lowell, MA 01851
Sun Microsystems
2550 Garcia Avenue
Mountain View, CA 94043
Sytek
1225 Charleston Road
Mountain View, CA 94043
Waterloo Microsystems, Inc.
175 Columbia Street West
Waterloo, Ontario
Canada
Tandem Computers, Inc.
19333 Vallco Parkway
Cupertino, CA 95014
Western Digital Corporation
2445 McCabe Way
Irvine, CA 92714
Texas Instruments, Inc.
Data Systems Group
Dallas, TX 75380
The Wollongong Group
1129 San Antonio Road
Palo Alto, CA 94303
Tiara Computer Systems, Inc.
2685 Marine Way
Mountain View, CA 94043
Xerox Corporation
Xerox Square 006
Rochester, NY 14644
Torus
495 Seaport Court, Suite 105
Redwood City, CA 94063
Xyplex, Inc.
100 Domino Drive
Concord, MA 01742
TRW Information Networks Division
23800 Hawthorne Boulevard
Torrance, CA 90505
Zenith Electronics Corporation
699 Wheeling Road
Mt. Prospect, IL 60056
Ungermann-Bass, Inc.
3990 Freedom Circle
Santa Clara, CA 95052
Zeta Laboratories
3265 Scott Boulevard
Santa Clara, CA 95054
Univation, Inc.
1231 California Circle
Milpitas, CA 94035
SIS DSP
© 1987 Dataquest Incorporated June
Directory of Potential Users
Local Area Network Equipment
(Page intentionally left blank)
© 1987 Dataquest Incorporated June
SIS DSP
Directory of Potential Users
Medical NMR Equipment
COMPANY DIRECTORY
Diasonics, Inc.
Milpitas, CA
Picker International
Highland Heights, OH
Fonar Corporation
Melville, NY
Siemens
Iselin, NJ
General Electric Medical Systems
Milwaukee, WI
Technicare Corporation
Solon, OH
Philips Medical Systems, Inc.
She1ton, CT
SIS DSP
© 1987 Dataquest Incorporated June
Directory of Potential Users
Medical NMR Equipment
(Page intentionally left blank)
© 1987 Dataguest Incorporated June
SIS DSP
Directory of Potential Users
Microwave Radio Equipment
COMPAMY DIRECTORY
AT&T Technologies
222 Broadway
New York, NY 10038
Hughes Microwave
P.O. Box 92426
Los Angeles, CA 90009
Avantek Inc.
481 Cottonwood Dr.
Milpitas, CA 95035
ITT
2912 Wake Forest Rd.
Raleigh, NC 27611
Aydin Microwave
75 East Trimble Dr.
San Jose, CA 95131
Loral TerraCom
9020 Balboa Ave.
San Diego, CA 92123
California Microwave
990 Almanor Ave.
Sunnyvale, CA 94086
M/A Com, Inc.
7 New England Road
Burlingame, MA 01803
Cardion Electronics
Long Island Expressway
Woodbury, NY 11797
Magnxim Microwave Corp.
365 Ravendale Dr.
Mountain View, CA 94043
Digital Microwave Corp.
2363 Calle del Mundo
Santa Clara, CA 95054
Motorola
3103 E. Algonquin Rd.
Schaumburg, IL 60190
Ericsson Information Systems
301 Route 17 North
Rutherford, NJ 07070
NEC America
2990 Telstar Ct.
Falls Church, VA
Fujitsu America
680 Fifth Ave.
New York, NY 10019
Northern Telecom
2300 Park Lake Dr.
Atlanta, GA 30348
General Electric Co.
316 E. Ninth St.
Owensboro, KY 42301
Oki Electronics of America, Inc.
4031 N.E. 12th Terrace
Fort Lauderdale, FL 33308
GTE Communications Company
2500 West Utopia Rd.
Phoenix, AZ 85027
Raytheon Co.
1415 Boston-Providence Turnpike
Northwood, MA 02062
Granger Associates
3101 Scott Blvd.
Santa Clara, CA 95051
Rockwell/Collins
1200 N. Alma Rd.
Richardson, TX 75081
Harris/Farinon
1691 Bayport Ave.
San Carlos, CA 94070
San/Bar
17422 Pullman St.
Santa Ana, CA 92711
SIS DSP
© 1987 Dataquest Incorporated June
22042
Directory of Potential Users
Microwave Radio Equipment
(Page intentionally left blank)
© 1987 Dataquest Incorporated June
SIS DSP
Directory of Potential Users
Modems
COMPAiaY DIRECTORY
Anderson Jacobson, Inc.
521 Charcot Avenue
San Jose, CA 95131
General DataComm Industries, Inc.
Route 63
Middlebury, CT 06762
Astrocom Corporation
120 West Plato Boulevard
St. Paul, MN 55107
Hayes Microcomputer Products, Inc.
5835 Peachtree Corners East
Norcross, GA 30092
AT&T Technologies Inc.
222 Broadway
New York, NY 10035
IBM Data Processing Division
1133 Westchester Avenue
White Plains, NY 10604
Avanti Communications Corp.
Aquidneck Industrial Park
Newport, RI 02840
Intertel, Incorporated
6 Shattuck Road
Andover, MA 01810
Bizcomp Corporation
532 Weddell Drive
Sunnyvale, CA 94089
Micom Systems Inc.
4150 Los Angeles Avenue
Simi Valley, CA 93063
CASE/Rixon
2120 Industrial Parkway
Silver Spring, MD 20904
Microcom, Inc.
1400A Providence Highway
Norwood, MA 02062
Cermetek Microelectronics Inc.
1308 Borregas Avenue
Sunnyvale, CA 94086
Multi-Tech Systems, Incorporated
82 Second Avenue S.E.
New Brighton, MN 55112
Codex Corporation
20 Cabot Boulevard
Mansfield, MA 02048
Novation
18664 Oxnard Street
Tarzana, CA 91356
Concord Data Systems, Inc.
442 Marrett Road
Lexington, MA 02173
Omnitech Data Corporation
2405 South 20th Street
Phoenix, AZ 85034
Data Race, Inc.
5839 Sebastian Drive
San Antonio, TX 78249
Paradyne Corporation
8550 Ulmerton Road
Largo, FL 33541
Digital Communications
Association
303 Research Drive
Atlanta, GA 30092
Penril Corp. Data Comm. Div.
3204 Monroe Street
Rockville, MD 20852
Gandalf Data Incorporated
1019 South Noel
Wheeling, IL 60090
SIS DSP
Prentice Corporation
266 Caspian Drive
Sunnyvale, CA 94086
© 1987 Dataquest Incorporated June
Directory of Potential Users
Modems
COMPANY DIRECTORY (Continued)
Racal-Milgo, Inc.
8600 N.W. 41st Street
Miami, FL 33166
Telebit Corporation
10440 Bubb Road
Cupertino, CA 95014
Racal-Vadic, Incorporated
222 Caspian Drive
Sunnyvale, CA 94086
Universal Data Systems (UDS)
5000 Bradford Drive
Huntsville, AL 35805
Rixon, Incorporated
2120 Industrial Parkway
Silver Spring, MD 20904
Ven-Tel, Incorporated
2342 Walsh Avenue
Santa Clara, CA 95051
Rockwell International
4311 Jamboree Road
Newport Beach, CA 92660
1987 Dataquest Incorporated June
SIS DSP
Directory of Potential Users
Multiplex Equipment
COMPANY DIRECTORY
AT&T Technologies
222 Broadway
New York, NY 10038
Karkar
245 11th Street
San Francisco, CA
Fujitsu America
680 Fifth Avenue
New York, NY 10019
L. M. Ericsson Telecommunications
7465 Lampson
Garden Grove, CA 92641
GTE/Lenkurt
250 West Utopia Road
Phoenix, AZ 85027
Motorola
3103 E. Algonquin Road
Schaumburg, IL 60190
Granger Associates
3101 Scott Boulevard
Santa Clara, CA 95051
NEC America
2990 Telestar Court
Falls Church, VA 22042
Harris/Farinon
1691 Bayport Avenue
San Carlos, CA 94080
Northern Telecom, Inc.
1555 Roadhaven Drive
Stone Mountain, GA 30083
ITT Telecommunications
2912 Wake Forest Road
Raleigh NC 27611
Rockwell International
1200 N. Alma Road
Richardson, TX 75080
SIS DSP
© 1987 Dataquest Incorporated June
94103
Directory of Potential Users
Multiplex Equipment
(Page intentionally left blank)
1987 Dataquest Incorporated June
SIS DSP
Directory of Potential Users
PBX and Centrex Telephone Systems
COMPANY DIRECTORY
AT&T Information Systems
100 Southgate Parkway
Morristown, NJ 07960
IPC Communications
One Lafayette Place
Greenwich, CT 06830
Anderson Jacobson, Inc.
521 Charcot Avenue
San Jose, CA 95131
ITT Telecom
3100 Highwoods Blvd.
Raleigh, NC 27604
CXC Corporation
2852 Alton Avenue
Irvine, CA 92714
Information Dynamics
1251 Exchange Drive
Richardson, TX 75081
Candella Electronics
550 Del Rey Avenue
Sunnyvale, CA 94088
InteCom Corporation
601 InteCom Drive
Allen, TX 75002
Cyber Digital Inc.
175 Commerce Drive
Hauppauge, NY 11788-3901
Jistel
76 Ferry Blvd.
Stratford, CT 06497
DTI
315 Eisenhower Lane South
Lombard, IL 60463
(markets the Rockwell Wescom PBX line)
Matsushita
1072 East Meadow Circle
Palo Alto, CA 94303
MelCO Labs
14408 Northeast 20th Street
Bellevue, WA 98007
Executone Ltd.
Two Jericho Plaza
Jericho, NY 11753
Mitel Corporation
5400 Broken Sound Blvd. N.W.
Boca Raton, FL 33431
GTE Communications Systems
2500 West Utopia Road
Phoenix, AZ 85027
Harris Digital Telephone Systems
One Digital Drive
Novate, CA 94947
Hitachi America
2990 Gateway Drive
Norcross, GA 30071
Honeywell Inc.
Honeywell Plaza
Minneapolis, MN 55408
SIS DSP
NEC Telephones
532 Broad Hollow Road
Melville, NY 11747
Northern Telecom
1001 East Arapaho Road
Richardson, TX 75081
Oki Electronics of America
Palisades Area
5901-B Peachtree Dunwoody Road
Atlanta, GA 30328
1987 Dataguest Incorporated June
Directory of Potential Users
PBX and Centrex Telephone Systems
COMPANY DIRECTORY (Continued)
Philips DVS
85 McKee Drive
Mahwah, NJ 07430
Tele/Resources
Northway 10, Executive Park
Bellston Lake, NY 12019
Redcom Labs
750 Fairport Park
Fairport, NY 14450
Telrad Telecommunications
Electronic Industrial Ltd.
510 Broad Hollow Road
Melville, NY 11747
ROLM Corporation
4900 Old Ironsides Drive
Santa Clara, CA 95050
SRX
15926 Midway Rd.
Dallas, TX 75234
Siemens Corporation
5500 Broken Sound Blvd.
Boca Raton, FL 33431
Solid State Systems
1990 Delk Industrial Blvd.
Marietta, GA 30067
Stromberg-Carlson ECS
2301 Maitland Center Parkway
Maitland, FL 32751
TIE/Communications
5 Research Drive
Shelton, CT 06484
Telex Computer Products
Telexecom Division
31829 La Tienda Dr.
Westlake Village, CA 95362
Teltone Corporation
P.O. Box 657
Kirkland, WA 98033
Thomson CSF Communications
2 Garnett Drive
White Plains, NY 10604
Tone Commander Systems
4320 150th Avenue N.E.
Redmond, WA 98052
Toshiba Telecom
2441 Michelle Drive
Tustin, CA 92680
Z-Tel
181 Ballardale Street
Wilmington, MA 01887
Tadiran
10801 Endeavor Way
Largo, FL 33543
Telenova
102-B Cooper Court
Los Gatos, CA 95030
Tele-Path Industries
3361 Melrose Avenue N.W.
Roanoke, VA 24017
(markets the hotel/motel PBX
Model 87 to the lodging industry)
© 1987 Dataguest Incorporated June
SIS DSP
Directory of Potential Users
Private Packet Data Networks
COMPANY DIRECTORY
PADS
Amdahl Communications Systems
2200 N. Greenville
Richardson, TX 75081
Digital Communications Assoc.
1000 Alderman Drive
Alvaretta, GA 30201
Amnet, Inc.
101 Morse Street
Watertown, MA 02172
Dynatech Packet Technology
6464 General Green Way
Alexandria, VA 22312
Atlantic Research
5390 Cherokee Avenue
Alexandria, VA 22312
Ericsson
1810 N. Glenville, Suite 116
Richardson, TX 75081
AT&T Network Systems
P.O. Box 1278R
Morristovm, NJ 07960
Gandalf Data
1019 S. Noel
Wheeling, IL
Auscom
2007 Kramer Lane
Austin, TX 78758
General Datacom
Middlebury, CT 06762-1299
BBN Communications
70 Fawcett Street
Cambridge, MA 02238
Cableshare, Inc.
20 Enterprise Drive
London, Ontario N6A4L6
Canada
Case Communications
7200 Riverwood Drive
Columbia, MD 21046-1199
ComDesign
751 S. Kellogg
Goleta, CA 93117
Databit, Inc.
110 Ricefield Lane
Hauppauge, NY 11788
Datagram Corp.
11 Main Street
E. Greenich, RI
SIS DSP
60090
GTE Telenet
12490 Sunrise Valley Drive
Reston, VA 22096
Hewlett-Packard Company
19055 Pruneridge
Cupertino, CA 95014
IBC
80 Oser Avenue
Hauppauge, NY 11787
I COT
830 Maude Avenue
Mountain View, CA
94039
Infotron Systems Corporation
N. Olney Avenue
Cherry Hill, NJ 08003
ITT
320 Park Avenue
New York, NY 10022
02818
1987 Dataquest Incorporated June
Directory of Potential Users
Private Packet Data Networks
COMPANY DIRECTORY (Continued)
PADS (Continued)
M/A-COM DCC
11717 Exploration Lane
Germantown, MD 20874
Siemens (Databit)
110 Ricefield Lane
Hauppauge, NY 117 88
Memotec
3320 Holcomb Bridge Road
Norcross, GA 30092
Telematics
1415 NW 62nd Street
Fort Lauderdale, FL
Mi com
20151 Nordhoff Street
Chatsworth, CA 91311
Timeplex, Inc.
400 Chestnut Ridge Road
Woodcliff Lake, NJ 07675
NEC America
1525 Walnut Hill Lane
Irving, TX 75062
Tymnet
2710 Orchard Parkway
San Jose, CA 95131
Northern Telecom
1001 E. Arapaho Road
Richardson, TX 75081
Uninet
10957 Lakeview Avenue
Lenexa, KS 66219
Paradyne
8550 Ulmerton Road
Largo, FL 33540
WoIfdata
187 Billerica Road
Chelmsford, MA 01824
Protocol Computers, Inc.
26630 Agoura Road
Calabasas, CA 91302-1988
XMIT AG
Widen Switzerland CH 8967
33309
Protocom Devices
190 Willow Avenue
Bronx, NY 10454
i2
© 1987 Dataquest Incorporated June
SIS DSP
Directory of Potential Users
Private Packet Data Networks
COMPANY DIRECTORY (Continued)
Nodes/Engines
Amdahl Communications Systems
2200 N. Greenville
Richardson, TX 75081
M/A-COM DCC
11717 Exploration Lane
Germantown, MD 20874
AT&T Network Systems
P.O. Box 1278R
Morristown, NJ 07960
NEC America
1525 Walnut Hill Lane
Irving, TX 75062
Amnet, Inc.
101 Morse Street
Watertown, MA 02172
Northern Telecom
1001 E. Arapaho Road
Richardson, TX 75081
BBN Communications
70 Fawcett Street
Cambridge, MA 02238
Paradyne (SESA)
8550 Ulmeiton Road
Largo, FL 33540
Dynatech (Dynapac) Packet
Technology
6464 General Green Way
Alexandria, VA 22312
Siemens (Databit)
110 Ricefield Lane
Hauppauge, NY 11788
Ericsson
1810 N. Glenville, Suite 116
Richardson, TX 75081
GTE Telenet
12490 Sunrise Valley
Reston, VA 22096
Harris Corporation
1025 W. Nasa Boulevard
Melbourne, FL 32419
Telematics
1415 N.W. 62nd Street
Fort Lauderdale, FL 33309
Tymnet
2710 Orchard Parkway
San Jose, CA 95131
Uninet
10957 Lakeview Avenue
Lenexa, KS 60219
ITT
320 Park Avenue
New York, NY 10022
SIS DSP
© 1987 Dataquest Incorporated June
\ J
Directory of Potential Users
Private Packet Data Netvj^orks
COMPANY DIRECTORY (Continued)
Switch/Concentrators
Amdahl Communications Systems
2200 N. Greenville
Richardson, TX 75081
Mi com
20151 Nordhoff Street
Chatsworth, CA 91311
Dynatech Packet Technology
6464 General Green Way
Alexandria, VA 22312
Northern Telecom
1001 E. Arapaho Road
Richardson, TX 75081
Memotech
3320 Halcomb Bridge Road
Norcross, GA 30092
Protocom Devices
190 Willow Road
Bronx, NY 10454
1987 Dataguest Incorporated June
SIS DSP
Directory of Potential Users
Satellite Earth Station Equipment
COMPANY DIRECTORY
Satellite Carriers awri ^Y''*'e'"s Operates
International Maritime Satellite
Organization-INMARSAT
40 Melton Street
London NWl ZEO
England
Alascom, Inc.
210 E. Bluff Road, Box 6607
Anchorage, AK 99502
American Satellite Company
1801 Research Blvd.
Rockville, MD 20850
AT&T Communications Satellite
Systems
Route 202/206, Room 2A235
Bedminster, NJ 07921
Communications Satellite Corporation
950 L'Enfant Plaza SW
Washington, DC 20024
COMSAT General Corporation
950 L'Enfant Plaza SW
Washington, DC 20024
International Satellite, Inc.
1331 Pennsylvania Ave. NW
Washington, DC 20004
International Telecommunications
Satellite Organization-INTESAT
5400 Inernational Drive
Washington, DC 20028
MCI-Space Resources Division
8283 Greensboro Drive
McLean, VA 22102
Orion Satellite Corporation
1835 K Street NW
Suite 201
Washington, DC 20036
Ford Aerospace Satellite
Services Corporation
1140 Connecticut Ave. NW
Suite 708
Washington, DC 20036
Pan American Satellite Corporation
460 W. 42nd Street
New York, NY 10036
Geostar Corporation
101 Carnegie Center
Suite 302
Princeton, NJ 08540
Rainbow Satellite Communications
P.O. Box 395
Leesburg, FL 32749
GTE Satellite Corporation
One Stamford Forum
Stamford, CT 06904
RCA American Communications, Inc.
Four Research Way
Princeton, NJ 08540
GTE Spacenet Corporation
1700 Old Meadow Road
McLean, VA 22102
Telesat Canada
333 River Road
Ottawa, Ontario KIL 8B9
Canada
Hughes Communications Galaxy, Inc.
Worldway Postal Center
P.O. Box 92424
Los Angeles, CA 90009
SIS DSP
Western Union Telegraph Company
One Lake Street
Upper Saddle River, NJ 07458
1987 Dataquest Incorporated June
Directory of Potential Users
Satellite Earth Station Equipment
COMPANY DIRECTORY (Continued)
Earth Station Equipment Manufacturers
Multipoint Communication
Corporation
1284 Geneva Drive
Sunnyvale, CA 94089
Avantek, Inc.
48761 Kato Road
Fremont, CA 94538
Aydin Systems Division
30 Great Oaks Blvd.
San Jose, CA 95119
COMSAT Telesystems
27521 Prosperity Avenue
Fairfax, VA 22031
COMTEL
2811 Airpark Drive
Santa Maria, CA 93455
Equatorial Communications
189 N. Bernardo Avenue
Mountain View, CA 94046
Fairchild Communications and
Electronics Company
20301 Century Blvd.
Germantown, MD 20874-1182
Harris Corporation
Satellite Communication Division
P.O. Box 1700
Melourne, FL 32901
Satellite Transmission
Systems, Inc.
Subsidiary of California
Microwave, Inc.
125 Kennedy Drive
Hauppauge, NY 11788
Scientific Atlanta
One Technology Parkway
P.O. Box 105600
Atlanta, GA 30348
Telecom General Corporation
2730 Junction Avenue
San Jose, CA 95134
Vitalink Communications
Corporation
1350 Charleston Road
Mountain View, CA 94043
Varian Associates
Microwave Equipment Division
3200 Patrick Henry Drive
Santa Clara, CA 95054
Hughes Aircraft Company
P.O. Box 9219
Los Angeles, CA 90009
M/A-COM DCC, Inc.
11717 Exploration Lane
Cermantown, MD 20874
© 1987 Dataquest Incorporated June
SIS DSP
Directory of Potential Users
Statistical Multiplexers
COMPANY DIRECTORY
AT&T Information Systems
One Speedwell Avenue
Morristown, NJ 07960
Complexx Systems
4939 Research Dr.
Huntsville, AL 35805
Anderson Jacobson
521 Charcot Avenue
San Jose, CA 95131
Datagram Corporation
11 Main Street
East Greenwich, RI 02818
Astrocom Corp.
120 W. Plato
St. Paul, MN 55107
Datatel, Inc.
Pin Oak Avenue
Cherry Hill, NJ
Backus Data Systems
1440 Knoll Circle
San Jose, CA 95112
Develcon Electronics
744 Nina Way
Warminister, PA 18974
Bit 3 Computer Corporation
8210 Penn Ave.
Minneapolis, MN 55431
Digital Communications
Association
303 Technology Park
Norcross, GA 30092
Black Box
P.O. Box 12800
Pittsburgh, PA
15421
Canoga Data Systems
21218 Vanowen Street
Canoga Park, CA 91303
CASE Communications
2120 Industrial Parkway
Silver Spring, MD 20904
Codex Corporation
20 Cabot Blvd.
Mansfield, MA 02048
Coherent Communications
60 Commerce Dr.
Hauppauge, NY 11788
Digital Equipment Corporation
146 Main Street
Maynard, MA 01754
Doelz Networks, Inc.
18581 Teller Avenue
Irvine, CA 92715
Emulex Corporation
3545 Harbor Blvd.
Costa Mesa, CA 92676
Fiberonics International, Inc.
218 West Main Street
Hyannis, MA 02601
Fujitsu America, Inc.
1945 Gallows Road
Vienna, VA 22180
ComDesign, Inc.
751 South Kellog Avenue
Goleta, CA 93117
SIS DSP
08003
1987 Dataquest Incorporated June
Directory of Potential Users
Statistical Multiplexers
COMPANY DIRECTORY (Continued)
Gandalf Data, Inc.
1019 South Noel Avenue
Wheeling, IL 60090
Paradyne Corporation
8550 Ulmerton Road
Largo, FL 33540
General Datacom
One Kennedy Avenue
Danbury, CT 06810
Penril
207 Perry Parkway
Gaithersburg, MD 20877
Halcyon Communications
2121 Zanker Road
San Jose, CA 95131
Prentice Corporation
266 Caspian Drive
Sunnyvale, CA 94088
Infinet, Inc.
6 Shattuck Rd.
Andover, MA 01810
Racal-Milgo
8600 NW 41st Street
Miami, FL 33166
Infotron Systems Corporation
9 North Olny Avenue
Cherry Hill, NJ 08033
Racal-Vadic
1525 McCarthy Blvd.
Milpitas, CA 95035
Interactive Systems/3M
3M Center
Saint Paul, MN 55144
Scitec Corporation
850 Aquidneck
Middletown, RI 02840
Honeywell
830 E. Arapaho Rd.
Richardson, TX 75081
Sequal Data Comm
P.O. Box 4069
Carey, NC 27511
M/A-COM DCC
11717 Exploration Lane
Germantown, MD 20874
Solana Electronics
249 S. Hwy. 101
Solana Beach, CA 92075
Micom Systems, Inc.
20151 Nordhoff Street
Chatsworth, CA 91311
Symplex Communications Corp.
2002 Hogback Road
Ann Arbor, MI 48104
Minntronics Corporation
2599 White Bear Ave.
St. Paul, MN 55109
Teleprocessing Products, Inc.
4565 E. Industrial St.
Simi Valley, CA 93063
Network Products, Inc.
Progress Center
Research Triangle Park, NC
Tellabs, Inc.
4951 Indiana Avenue
Lisle, IL 60532
Nixdorf Computer
300 Third Ave.
Waltham, MA 02154
27709
Teltone Corporation
10801 120th Avenue, N.E.
Kirkland, WA 98033
1987 Dataquest Incorporated June
SIS DSP
Directory of Potential Users
Statistical Multiplexers
COMPANY DIRECTORY (Continued)
Timeplex, Inc.
400 Chestnut Ridge Road
Woodcliff Lake, NJ 07675
Western Datacom
5083 Market Street
Youngstown, OH 44512
Tri-Data
505 East Middlefield Road
Mountain View, CA 94039
SIS DSP
© 1987 Dataquest Incorporated June
Directory of Potential Users
Statistical Multiplexers
(Page intentionally left blank)
1987 Dataguest Incorporated June
SIS DSP
Directory of Potential Users
T-1 Multiplexers
COMPAMY DIRECTORY
Amdahl Corporation
2500 Walnut Avenue
Marina Del Rey, CA
DATATEL, Inc.
Cherry Ind. Ctr.
Cherry Hill, NJ 08003
90291
AT&T Information Systems
One Speedwell Avenue
Morristown, NJ 07960
Digi-Voice Corp.
565 Fifth Avenue
New York, NY 10017
Avanti Communications Corp,
Aquidneck Industrial Park
Newport, RI 02840
Ericsson Communications
7465 Campson
Garden Grove, CA 92642
Aydin Microwave
75 E. Trimble Road
San Jose, CA 95131
Fujitsu America
3055 Orchard Drive
San Jose, CA 95134
Aydin Monitor
502 Office Center Drive
Fort Washington, PA 19034
Galdalf Data
1019 S. Noel
Wheeling, IL
Bayly Engineering
167 Hunt Street
Ajax, Ontario, Canada
General Datacomm Industries
One Kennedy Avenue
Danbury, CT 06810
LIS 1P6
60090
CASE Communications
8310 Guilford Road
Columbia, MD 21046
Granger Associates
3101 Scott Blvd.
Santa Clara, CA 95054-3394
Coastcom
2312 Stanwell Drive
Concord, CA 94520
GTE-Lenkurt
501 Sycamore Drive
Milpitas, CA 95035
Codex
20 Cabot Blvd.
Mansfield, MA 02048
Halcyon Communications
2121 Zanker Road
San Jose, CA 95131
Cohesive Network
1680 Dell Avenue
Campbell, CA 95008
IDS
7 Wellington Road
Lincoln, RI 02865
Comtech Communications
6150 Lookout Road
Boulder, CO 80301
Infotron Systems Corporation
9 North Olney
Cherry Hill, NJ 08003
Data Communications Associates
3030 Research Drive
Norcross, GA 30092
Integrated Telecom Corporation
9216 Markville
Dallas, TX 75243
SIS DSP
1987 Dataquest Incorporated June
Directory of Potential Users
T-1 Multiplexers
COMPANY DIRECTORY (Continued)
ITT
3100 Highwoods Blvd.
Relelgh, NC 27604
San/Bar
9999 Muirlands Blvd.
Irvine, CA 92714
Loral-Terracorn
9020 Balboa Avenue
San Diego, CA 92123
Scitec Corporation
811 Aquidneck Avenue
Middletown, RI 02840
Lynch Communications
204 Edison Way
Reno, NV 89520
Seiscor Technologies
5311 S. 122nd E. Avenue
Tulsa, OK 74147
M/A-Comm Linkabit
3033 Science Park Road
San Diego, CA 92121
Stratcom
10341 Bubb Road
Cupertino, CA 95014
Micom Systems
20151 Nordhoff Street
Chatsworth, CA 91311
Tau-Tron
27 Industrial Avenue
Chelmsford, MA 01824
NEC
2741 Prosperity Avenue
Fairfax, VA 22039
Tellabs
4951 Indiana Avenue
Lisle, IL 60532
Network Equipment Technology
400 Penobscot Drive
Redwood City, CA 94063
Teltone Corporation
1080 120th Avenue
Kirkland, WA 98033
Network Switching Systems
3 Dundee Park
Andover, MA 01810
Timplex, Inc.
400 Chestnut Ridge Road
Woodcliff Lake, NJ 07675
Northern Telecom
1555 Roadhaven Drive
Stone Mountain, GA 30083
Wescom, Inc.
8245 S. Lament Road
Downers Grove, IL 60515
Paradyne Corporation
8550 Umberton Road
Largo, FL 33540
Racal-Milgo
8600 NW 41st Street
Miami, FL 33166
© 1987 Dataquest Incorporated June
SIS DSP
Directory of Potential Users
Telephone Equipment
COMPAMY DIRECTORY
American Telecom, Inc.
3190 East Miraloma Avenue
Anaheim, CA 92806
Goldstar Telecommunications, Inc.
One Madison Street
East Rutherford, NJ 07073
American Telephone & Telegraph
P.O. Box 49209
Atlanta, GA 30359
Graybar Electric Company
900 Commerce Drive
Oak Brook, IL 60521
Buckeye Telephone and Supply Co,
1800 ArUngate Lane
Columbus, OH 43228
ITT Business Communications Corp,
300 East Park Drive
Harrisburg, PA 17111
Centel Supply Co.
16215 Marquardt
Cerritos, CA 90701
Iwatsu America, Inc.
230 Lincoln Centre Drive
Foster City, CA 94404
Comdial Telephone Systems
1180 Seminole Trail
Charlottesville, VA 22906
Melco Labs, Inc.
Rte. 1, Box 96X
Union Springs, AL
Crest Industries, Inc.
6922 North Meridian
Puyallup, WA 98371
Motorola Communications and
Electronics, Inc.
2700 Augustine Drive
Santa Clara, CA 95051
Electra
300 East County Line Road
Cumberland, IN 46229
36089
North Supply Company
600 Industrial Parkway
Industrial Airport, KS
66031
Famous Telephone Supply Co,
110 North Union
P.O. Box 2172
Akron, OH 44309
Northern Telecom
1001 E. Arapaho Road
Richardson, TX 75081
General Electric Company
316 East Ninth Street
Owensboro, KY 42301
Telenova, Inc.
102-B Cooper Court
Los Gatos, CA 95030
GTE Business Communications
Systems, Inc.
Consumer Products Department
P.O. Box 4148
Huntsville, AL 35803
Telicom, Inc.
11411 Addison Street
Franklin Park, IL 60131
GEC Telecommunications Limited
P.O. Box 53
Coventry CV3 IHJ
England
SIS DSP
Webcor Electronics
28 South Terminal Drive
Plainview, NY 11788
1987 Dataguest Incorporated June
Directory of Potential Users
Telephone Equipment
(Page intentionally left blank)
1987 Dataquest Incorporated June
SIS DSP
Directory of Potential Users
Ultrasound Equipment
Hewlett-Packard Company
Palo Alto, CA
COMPAMY DIRECTORY
Advanced Tech Labs
Bellevue, WA
Honeywell
Minneapolis, MN
American Hospital Supply Company
Evanston, IL
Andersen Group, Inc.
Bloomfield, CT
Philips Ultrasound
Santa Ana, CA
Siemens
Iselin, NJ
Cooper Labs
Palo Alto, CA
Squibb Corporation
Princeton, NJ
Diasonics, Inc.
Milpitas, CA
Technicare Corporation
Solon, OH
General Electric Medical Systems
Milwaukee, WI
Picker International
Highland Heights, OH
SIS DSP
© 1987 Dataquest Incorporated June
Directory of Potential Users
Ultrasound Equipment
(Page intentionally left blank)
© 1987 Dataguest Incorporated June
SIS DSP
Directory of Potential Users
Video Teleconferencing Systems
ISACOM
1815 Century Boulevard
Suite 500
Atlanta, GA 30345
COMPANY DIRECTORY
American Satellite Corp.
1801 Research Boulevard
Rockville, MD 20850
American Telephone & Telegraph Co.
195 Broadway Ave.
New York, NY 10007
Avalex
310 Bonifant Rd.
Silver Springs, MD
NEC America, Inc.
Radio and Transmission Division
2740 Prosperity Ave.
Fairfax, VA 22031
20904
Colorado Video
P.O. Box 928
Boulder, CO 80306
PicTel
One Intercontinental Way
Peabody, MA 01960
Conferez Corporation
1848 E. Carnegie Ave.
Santa Ana, CA 92705
Pierce-Phelps, Inc.
2000 N. 59th Street
Philadelphia, PA 19131
Compression Labs, Inc.
2305 Bering Drive
San Jose, CA 95131
Robot Research
7591 Convoy Court
San Diego, CA 92111
Concept Industries Co.
1116 Summer Street
Stamford, CT 06905
Satellite Business Systems
8283 Greensboro Drive
McLean, VA 22102
Decisions and Designs, Inc.
8400 Westpark Drive, Suite 600
McLean, VA 22101
GEC McMichael
Sefton Park
Stoke Poges, Slough, SL2 4HD
England
Image Data Corporation
7986 Mainland Dr.
San Antonio, TX 78250
Interand Corp.
3200 West Peterson Ave.
Chicago, IL 60659
SIS DSP
Luma Telecom
3350 Scott Boulevard
Building 49
Santa Clara, CA 95054
Shure Brothers
222 Hartrey Avenue
Evanston, IL 60204
Vidicom Video Communications
Division of L.D. Bevan Co., Inc.
742 Hampshire Road, Suite D
Westlake Village, CA 91361
Vitalink Communications Corp.
1350 Charleston Road
Mountain View, CA 94043
Widcom
1500 Hamilton Ave.
Campbell, CA 95008
© 1987 Dataquest Incorporated June
Directory of Potential Users
Video Teleconferencing Systems
(Page intentionally left blank)
1987 Dataquest Incorporated June
SIS DSP
Directory of Potential Users
Voice Messaging Systems
COMPANY DIRECTORY
AT&E, Inc.
1400 NW Compton Drive
Suite 300
Beaverton, OR 97006-1922
Genesis Electronics Corporation
Lake Forest Technical Center
103 Woodmere Road
Folsom, CA 95630
AT&T Information Systems
One Speedwell Avenue
Headquarters Plaza West Tower
Morristown, NJ 07060
IBM Corporation
1133 Westchester Avenue
White Plains, NY 10604
NEC America, Inc.
532 Broad Hollow Road
Melville, NY 11747
American Telesystems
Six Piedmont Center
Suite 608
Atlanta, GA 30305
Natural Microsystems Corporation
6 Mercer Road
Natick, MA 01760
Applied Voice Technologies
2445 140th Avenue N.E.
Suite 201
Bellevue, WA 98005
Northern Telecom, Inc.
Business Communications Systems
2305 Mission College Blvd.
Santa Clara, CA 95050
BEL Industries, Inc.
P.O. Box 48488
Atlanta, GA 30362
Octel Communications Corporation,
1841 Zanker Road
San Jose, CA 95131
Centigram Corporation
1362 Borregas Avenue
Sunnyvale, CA 94089
ROLM Corporation
4900 Old Ironsides Drive
Santa Clara, CA 95050
Commterm, Inc.
The Third Avenue
Burlingame, MA 01532
Digital Equipment Corporation
10 Forbes Road
Northboro, MA 01532
Digital Pathways
1060 East Meadow Circle
Palo Alto, CA 94303
Digital Sound
2030 Alameda
Padre Serra, CA
SIS DSP
Solid State Systems, Inc.
1900 Delk Industrial Blvd.
Marietta, GA 30067
Sperry Corporation
P.O. Box 500
Blue Bell, PA 19424
Sudbury Systems
31 Union Avenue
Sudbury, MA 01776
93103
1987 Dataquest Incorporated June
Directory of Potential Users
Voice Messaging Systems
COMPANY DIRECTORY (Continued)
VMX, Inc.
1241 Columbia Drive
Richardson, TX 75081
Votrax, Inc.
1394 Rankin
Troy, MI 48083
Voice Computer Technologies
Corporation
5730 Oakbrook Parkway
Suite 175
Norcross, GA 30093-1888
Wang Laboratories
One Industrial Avenue
Lowell, MA 01851
Voicemail International, Inc.
2225 Martin Avenue
Santa Clara, CA 95050
Voicetek Corporation
61 Chapel Street
Newton, MA 02195
Zaiaz Communications, Inc.
207 Lakin Drive
Huntsville, AL 35801
Zymacom
2 Lyberty Way
Westford, CT 01886
Votan
4487 Technology Drive
Fremont, CA 94538
1987 Dataguest Incorporated June
SIS DSP
Worldwide IC Packaging Update
OVERVIEW
The normal state of affairs in the semiconductor industry is to be in a
"state of transition" or to have "reached a milestone." Or, something has
occurred that will "revolutionize" the industry. Packaging of semiconductors
is no exception.
Significant achievements in VLSI fabrication and design technologies have
reached the point where concurrent improvements in die-level interconnection
technologies are necessary for continued system performance. Of all the
packaging and interconnection technology issues discussed, one issue readily
agreed upon is that both users and suppliers of semiconductors are going
through a demanding transitional phase of component packaging decisions—
decisions that will have to be dealt with in the near future, as the industry
approaches submicron geometries.
One very clear trend that we are seeing is that eguipment manufacturers
are using more and more VLSI devices. There is a sweeping desire to reduce
space and cost through more condensed packaging and to automate as much as
possible.
To accomplish this, packaging technology must approach chip
technology.
PACKAGE CONSUMPTION
Figure 1 shows the estimated worldwide integrated circuit (IC) package
consumption for 1986. The estimates are based on Dataguest's worldwide IC
consumption data and therefore show consiimption by all packaged ICs. Japan
captured 40 percent of packaged ICs in 1986, while U.S. market share dropped
to approximately 33 percent and Europe came in at 17 percent. The remaining
10 percent not shown went to ROW.
We expect that the Japanese will maintain their lead in the 1988 market
using 44 percent of packaged ICs, with the United States holding approximately 38 percent, and Europe with 18 percent. By 1991, we anticipate that
Japan will strengthen its lead to 45 percent, by virtue of its majority share
of the consumer business, concerted efforts in the industrial sector, and its
lead in automated assembly. At this point, U.S. market share will drop to
34 percent, and Europe's share will climb to 21 percent. While Europe is
obviously not defeating its American and Asian competitors, we do expect it
to modestly regain market share. At this time, we believe that European
users are changing to surface-mount technology more readily than the American
and Japanese users. Telecommunications and IC smart card applications,
focusing on small-outline (SO) and tape-automated bonding (TAB) will provide
Europe with the biggest growth opportunities for the next 10 years.
SIS DSP
O 1987 Dataguest Incorporated June
Worldwide IC Packaging Update
Figure 1
ESTIMATED 1986 WORLDWIDE PACKAGING TRENDS
(Units)
U.S.
Japan
Europe
Source: Dataquest
May 1987
THE MEMORY ROLE
Over the last few years, memory devices have been on the leading edge of
packaging technology due to density requirements. We have forecast that
approximately 55 million 1-Mbit DRAMs will be shipped worldwide in 1987. As
shown in Figure 2, 70 percent of those units will be shipped in either
plastic or ceramic dual in-line packages (DIPs). By 1988, DIP package usage
for DRAMs will shrink to 65 percent, while zig-zag in-line package (ZIP) and
small-outline J-lead (SOJ) usage will grow. As we move into the 1990s, the
SOJ package is expected to grow to 32 percent.
High-density device
architectures, led by smaller geometries and line widths, coupled with the
desire to bring down costs while maintaining price competitiveness and
building better and faster machines, will require the increased use of
surface-mount technology (SMT).
e 1987 Dataquest Incorporated June
SIS DSP
Worldwide IC Packaging Update
Figure 2
ESTIMATED 1Mb DRAM PACKAGES
1987
P ^ P/C DIP
l^^^jjM Ow«l
[S3 ZIP
1990
1988
Source: Dataquest
May 1987
SMT ISSUES
Despite the many advantages, implementation of surface-mount (SM)
packages into systems manufacturing is taking longer than anticipated.
Surface-mount technology is still immature and as such the manufacturing
infrastructure is not fully developed. Preferring the tested reliability of
through-hole (TH) packages, users continue to mix SM/TH designs. Reliability
of SM devices has not yet been proven and solder joint inspection is
difficult. However, as shown on Tcible 1, concentrated use of SM devices is
occurring in applications where small size and weight are the primary
issues. As shown on Table 2, computers were the leading end-use segment for
SMDs in the United States, in 1986. While cost reduction was the driving
force, reliability continues to play a major role in acceptance of SMT.
Europe led the United States in acceptance and usage of SMT in telecom
applications; and by virtue of its command over the consumer market, Japan
led the market with 10 percent of ICs packaged in SMT. As a comparison,
Japan's Printed Circuit Association (JPCA) estimated that SM consumption in
Japan reached 13 percent for ICs, and that over the next five years, ICs in
SMT will grow to 33.9 percent in Japan.
SIS DSP
e 1987 Dataquest Incorporated June
Worldwide IC Packaging Update
Table 1
SURFACE-MOUNT TECHNOLOGY
Where?
•
Consiuner
•
Automotive
•
Disk storage
•
Avionics, missiles, and space
•
High-density memories
•
Power supplies
Table 2
SUKFACB-MOUNT TECHNOLOGY
END-USE SEGMENTS
1986
Japan
Evrppe
United States
End Use
Consumer
Telecommunications
Computers
Driving Force
Small size
Reliability
Cost reduction/
reliability
Percent of ICs
Consumed Worldwide
40%
17.7%
32.8%
Percent of ICs in SMT
10%
8.0%
4.0%
Dominant SMT Approach
TAB/QUAD/SO
SO
SO/CC/TAB
Source:
O 1987 Dataquest Incorporated June
Dataquest
June 1987
SIS DSP
Worldwide IC Packaging Update
SUMMARY
At the present time, we believe that there is no single solution to
future VLSI packaging problems. For the 1990s and beyond, we expect that
package designs will continue to proliferate. Advanced multichip product
designs will incorporate ASICs, use advanced circuit design techniques, and
use advanced board assembly methods incorporating TAB and other multichip
packages. While plastic packaging has its hermetic limitations, its highvolume, low-cost, high-performance, 40 pin-and-below characteristics will
make it the dominant package by 1990.
Automated assembly will change the way that ICs and other components are
packaged. TAB or some variation of this method of construction is the most
likely packaging style for ICs in the 1990s. Chip-on-board (COB) has also
made its way up the automated assembly ladder in consumer applications. From
early single-chip digital watch applications, it is now being used in multichip applications such as copiers, facsimile, and IC cards.
REGIONAL ANALYSIS
If we use the premise that memory devices have been on the leading edge
of packaging technology due to density requirements, then we caui assume that
Japan has a two-year lead on the industry and will gain overall leadership in
packaging technology before the 1990s.
With its vertically integrated
structure, Japan can maintain closer technical and strategic cooperation
among members of its packaging chain. Their command over the consumer market
and surface-mount approach has given them a lead in packaging technology.
There are already major efforts among equipment suppliers in Japan to develop
automated assembly processes.
Despite major engineering efforts dedicated to designs, substrate and
component materials, and assembly equipment, cooperation lags among members
of the packaging chain in the United States. At times, cooperation seems
better between U.S./Japanese partners than between U.S./U.S. alliances. The
strong financial/technical megacorporate links of Japan are nonexistent in
the United States. Outside of Texas Instruments and a few systems groups,
everyone else has transported assembly offshore. Unlike Europe and Japan,
there is very little academic research and cooperation. There is some hope
in U.S. research consortiums, but cooperative efforts in packaging are weak.
Finally, except for a few systems houses, the fear of capital investments in
automated assembly technology has paralyzed many companies from making the
decision to automate, a decision that could prevent them from staying on the
competitive edge.
SIS DSP
O 1987 Dataquest Incorporated June
Surface-Mount Technology Overview
SURFACE-MOOHT TECHNOLOGY
What Is It?
Surf ace-moiuit technology (SMT) was first used in the United States in the
early 1960s by the military because it met their requirements for space
savings and high reliability. SMT became available for commercial use in the
United States in the 1970s. Today, SMT is used most often in the automotive,
computer, and consumer electronics industries as well as aerospace.
Dataguest defines SMT CAD as the laying out of printed circuit boards
(PCBs) with chips mounted to the surface of the board. The differences
between through-hole technology and SMT that affect CAD systems include
device footprints, packaging, and access to internal layers of the board.
How Is It Different?
When designers lay out printed circuit boards (PCBs) with through-hole
technology (THT) devices, footprints (the graphic shape of a device) for
components consist of round pads that go through all layers of the board. In
SMT, footprints are made up of rectangular pads that reside on the external
layers of the board only (please refer to Figure 1).
Looking further into the differences among THT and SMT devices, there are
standards for through-hole device packages (i.e., the same-shaped device is
available from a variety of vendors), while there are no such standards for
SMT devices.
Although there are several organizational efforts to
standardize SMT device packages, today's users contend with the confusion
caused by the same technology or device being available in too many packages.
The variation in packages causes designers confusion because they must
create the footprints and physical library for each SMT device. Before they
begin the physical layout, users need to know which manufacturer's components
will be used so that the footprint graphic will match the actual device.
Another difference between these technologies is that through-hole
devices use vias that go through all layers of the board, whereas in surface
mounting, designers use blind and/or buried vias to access the internal
layers of the board (see Figure 2 ) .
SIS DSP
e 1987 Dataquest Incorporated June
Surface-Mount Technology Overview
Figure 1
DEVICE FOOTPRINTS
Source: Nugrafix Group
Design Guideline Book
Figure 2
VIAS
Blind Via<
Buried V\a\
•
1
\
,
t
2
II
1"
3
J
Blind Via
Through Via
/
Source: Nugrailx Group
Oeiign Guideline Hook
G 1987 Dataguest Incorporated June
SIS DSP
Surface-Mount Technology Overview
THE SMT SURVEY
Demographics
Dataguest recently completed a survey focusing specifically on the needs
of end users implementing SMT on CAD systems. The survey sample consisted of
100 PCB CAD end users. Our selection criterion was based on whether the
users were using their CAD systems for PCB layout rather than whether they
were using SMT or not.
Among the responses, 19 percent came from service bureaus and another
18 percent came from computer companies.
Please refer to Figure 3 for
further details on the industries of the respondents.
One-third (33 percent) of the respondents indicated that they have been
using SMT for one year or less. However, the bulk of the respondents
(47 percent) responded that they have been using surface-mount technology for
two to three years. Approximately 20 percent have been using SMT for more
than three years. Less than 1 percent indicated that they do not use SMT and
have no plans to do so in 1987.
' Figure 3
RESPONDENTS B7 END-USER INDUSTRY
Semiconductors
10.6%
Text and .
Measurement
10.6%
Communications
7.1%
Military/
Defense
7.1%
Consumer
Electronics
4.7%
Source: Dataquett
Junel987
SIS DSP
e 1987 Dataquest Incorporated June
Surface-Mount Technology Overview
Most responding organizations had a total design engineering staff of
10 people or less. Furthermore, 61 percent of these organizations responded
that their engineers are doing surface-mount design. Similarly, total layout
designers numbered 10 or less for the majority of respondents, with
68 percent indicating that 1 to 10 designers are designing with surface-mount
technology (please see Figure 4).
Figure 4
NUMBER OF ENGINEERS AND DESIGNERS
Percent of Respondents
10010 or Less
90
11 to 25
60
26 to 50
70
I
60-
H H
I 51 to 75
Greater than 75
6040JOr
3^
10
0
§
"X
%>
X^
g? >ii
I
^^
MUZL
Total Design
EnQlneers
Total Layout
Designers
SMT Design
Engineers
SMT Layout
Designers
Source: Dataqueit
June 1987
How Is SMT Being Used?
In examining the prevalence of SMT design in proportion to traditional
through-hole technology (THT), we looked at this issue from several angles:
Total annual design starts (see Figure 5), which of those use SMT (see
Figure 6), how SMT is implemented (see Figure 7), the number of components,
and the number of layers per design (see Figures 8 and 9).
Results indicate that implementation of surface-mount technology on CAD
systems is still relatively new and definitely not widespread. Why are users
converting to SMT? Figure 10 shows the top five reasons respondents have
chosen to use surface-mount technology over through-hole design methods.
In spite of the sparsity of SMT usage, there is a perception,
particularly among responding service bureaus, that users must support SMT to
stay in business because their customers demand it and their competition
supports it.
O 1987 Dataguest Incorporated June
SIS DSP
Surface-Mount Technology Overview
Figure 11 shows the layout phase of the design cycle as a percent of the
total design time, comparing through-hole technology to SMT. In our focus
research/ end users indicated that one of the benefits of using SMT was that
they could get their products designed and manufactured faster.
Yet,
the
results of our survey show that there is little or no time saved by choosing
surface-mount technology instead of THT. In researching this issue further,
we learned that the time savings involved in SMT comes from the manufacturing
process/ where SMDs are more suited to automated manufacturing processes.
Figure 5
TOTAL ANNUAL PCB DESIGN STARTS
Percent of Respondents
35-
7.e%
5.1%
5.1%
^
15 or lesi
20 to 35
50 to 75
80 to 100
120 to 150
200 to 250
300 to 400
Mor* than 500
Number of Designs
Source: Dataquest
June 1987
SIS DSP
e 1987 Dataquest Incorporated June
Surface-Mount Technology Overview
Figure 6
DESIGNS WITH SMT
Percent of Respondents
50
10% or less
15 to 25
30 to SO
70 to 80
Mors than 90
Percent of Designs
Source: Dataquest
Junel987
O 1987 Dataquest Incorporated June
SIS DSP
Surface-Mount Technology Overview
Figure 7
DISTRIBUTION OF DESIGNS
BY TECHNOLOGY IMPLEMENTATION
Percent of Respondents
50
45%
Surface Mount
Technology
(One Side CJtily)
40
3022%
20
16%
10
Don't Do
10% or Less
10%
ml
^
vSXW'
15 - 25%
26 - 50%
51 - 75%
6%
2%
k'WVV'l
76 - 100%
Percent of Total Boards
Percent of Respondents
50
Mixed Technology
Through-Hole and
Surface Mount
(Single Side)
Don't Do
10% or Less
15 - 25%
26 - 5 0 %
51 - 75%
76 - 100%
Percent of Total Boards
Continued
SIS DSP
O 1987 Dataquest Incorporated June
Surface-Mount Technology Overview
Figure 7 (Continued)
DISTRIBUTION OF DESIGNS
BY TECHNOLOGY IMPLEMENTATION
Percent of Respondents
40
Mixed Technology
Through-Hole and
Surface Mount
(Both Sides)
17%
Don't Do
10% or Less
15-25%
26-50%
76 - 100%
Percent of Total Boards
Percent of Respondents
40
Surface Mount
Technology
(Both Sides)
Don't Do
10% or Less
15 - 25%
26 - 50%
VV\.VSJ
#
51 - 75%
76 - 100%
Percent of Total Boards
Source: Dataquett
June 1987
O 1987 Dataguest Incorporated June
SIS DSP
Surface-Mount Technology Overview
Figure 8
AVERAGE NUMBER OF COMPONENTS
BY BOARD TYPE
Percent of Respondents
50
Don't Do
<100
100-299
300-599
600-699
900 or more
Number of Components
Source: Dauquett
"ine 1987
19
June
Figure 9
AVERAGE NOMBER OF LAYERS
BY BOARD TYPE
Percent of Respondents
70
K"\.'1 Surface Mount, 1 Side, Avg. # Layers
Mixed Technology, 1 Side, Avg. # Layers
60-
l^^^^iJ Surface Mount, 2 Sides, Avg. # Layers
I
I Mixed Technology, 1 Side, Avg. # Layers
50
40
30
20
10
0
1 to 2
Layers
4 to 6
Layers
8 to 10
Layers
11 to 20
Layers
Source: Dauqueit
- e l1987
9—
June
S I S DSP
e 1987 Dataquest Incorporated June
Surface-Mount Technology Overview
Figure 10
SEASONS FOR USING SMT
Density
32,2%
Board Size
Reduction
29.5%
\
\
)
\ \
\ ^ \
\
\
\
/Reduced
X
Cost
X^1.9%
Other /
3,7% /
\>^
Better
_,if^
Performance
Customer
7.8%
Choice
8.8%
Percent of Respondents
Source: Dataquett
June 1987
Figure 11
LAYOUT AS A PERCENT OF TOTAL DESIGN TIME
Percent
60•^
50-
THT
SMT
48%
4030-
27%
2013%
15%
r ^ 1 5 %
10-
1 -24
25 - 49
50 - 74
7 5 - 100
Percent of Design Time
Source: Dataqueit
Junel987
10
e 1987 Dataquest Incorporated June
SIS DSP
Surface-Mount Technology Overview
SMT and CAD
What Do End Users Think?
The attitude of most end users is less than optimistic regarding the
currently available CAD tools. They feel that although CAD vendors have made
a good start, they have a long way to go in terms of adequately supporting
SMT. Other end users stated that they had been sold tricks and workarounds as
true design solutions. Overall, less than 10 percent of the respondents were
satisfied with the way their CAD systems support SMT.
As Figure 12 shows, the likes and dislikes of the end users closely
parallel each other. Closer inspection of the data revealed no technological
reasons, thus leaving us with one conclusion: The likes and dislikes cited
are vendor-specific.
Ranking specific SMT support functions in order of importance, definition
of pad geometries topped the list, followed by multilayer routing, off-grid
design, and auto-routing of two-sided boards. End users are saying that, in
order to support these important SMT functions, PCB CAD systems must be
flexible and interactive enough to accommodate SMT as well as THT features.
Figure 12
END USER LIKES AND DISLIKES
Percent of Respondents
50-,
40-
E ^
UkBS
• •
DIstlkas
30-
20
10-
SMT
Routsr
Flexibility
Source: Datiqueit
June 1987
SIS DSP
e 1987 Dataquest Incorporated June
11
Surface-Mount Technology Overview
Users are also saying that they view SMT support as a PCB CAD system
feature that must be capable of integration into users' particular design
environments.
Because SMT is highly process dependent, users need to
interface easily with manufacturing to ensure the manufacturability of the
design.
Flexibility and integration are the two most important buying criteria
for future SMT CAD purchases cited by respondents.
What Are the Challenges of SMT?
The following characteristics
through-hole technology:
of SMT affect CAD systems that support
•
Footprints
•
The lack of standards
•
Access to internal layers of the board
It is the shape of SMT device footprints as well as the fact that they
reside only on the surface of the board that affects PCB CAD systems. Most
systems are set up to acknowledge the footprints of through-hole devices and
are not surface intelligent.
The lack of standard geometries for the same device functionality was
cited by respondents as the major drawback in converting designs to SMT. The
lack of standards for SMD has created a need for a high degree of flexibility
and interactivity in PCB CAD systems.
Because most PCB CAD systems are not surface intelligent, users have to
trick the system into believing that it is routing a through-hole device. To
accomplish this, designers place stringers (round pads) at the end of each
rectangular pad so that the system thinks it is routing a component whose
leads run through all layers of the board. Although this workaround does
route the board, it does not provide a long-term design solution.
Size of the ShCC Market
End users are budgeting for design solutions that support SMT. Nearly
60 percent of the respondents replied that they have budgeted up to $100,000
for SMT CAD tools in 1987.
The highest amount budgeted for SMT CAD
expenditures in 1987 was just slightly more than $1 million, cited by nearly
10 percent of the respondents.
12
e 1987 Dataquest Incorporated June
SIS DSP
Surface-Mount Technology Overview
To quantify and qualify the SMT opportunity, we based the forecast in
Figure 13 on a combination of factors:
•
Dataquest forecast data base, which consists
research on more than 140 companies
of
four
years
of
•
Data from another ECAD end-user survey, indicating number of designs
and engineers
•
The Dataquest Semiconductor Industry Service's forecast for SMDs
that approximately 16 billion units will be shipped in 1990
•
End users' forecast for 1987, where more than 61 percent indicated
that they will use SMT in 15 to 25 percent of their designs
As Figure 13 shows, Dataquest estimated that in 1985, there were more
than 17,000 workstations in the installed base for PCB CAD. Of those, we
estimate that approximately 2,000 supported SMT. We forecast that by 1990
the installed base of workstations used for PCB applications will be more
than 90,000 units, more than half of which will support SMT.
In compiling our forecast, we considered SMT to be a function of a PCB
CAD system, not a turnkey product offering. Therefore, we believe that a
number of software licenses may be sold as repeat business to a vendor's
installed base as well as to new customers.
Figure 13
PCB HOSKSTATIOH INSTALLED BASE
WITH SMT CAPABILITY
Thousands of installed Workstations
1981
tasa
18B$
19S7
1989
Source: Dataqucit
Junel987
SIS DSP
e 1987 Dataquest Incorporated June
13
Surface-Mount Technology Overview
SUMMARY
We believe that SMT is here to stay because end users need reduced board
size with increased density, more reliable end products, and faster and less
expensive manufacturing processes to keep up with their worldwide competition.
Dataguest believes that SMT support tools will be marketed as features
of a CAD system, not standalone turnkey products. As such, meeting the
technological differences with true design solutions is the challenge for CAD
vendors in this market.
To recapitulate, the likes and dislikes of
vendor specific. We believe that in addition to
vendors have to overcome the negative attitudes of
that they understand the nature of their customers'
responding end users are
watching the competition,
end users by demonstrating
problems.
Dataquest believes that the successful CAD vendors will be those who work
closely with their customers to provide the needed solutions. To take
advantage of the opportunity offered by SMT, vendors need to project that
their products are what the end user wants: A means to an end—a quick
turnaround on a manufacturable design.
14
© 1987 Dataquest Incorporated June
SIS DSP
\
'%:
h,
f^'^^i^
u
-M
W:^-' •'''i •' '' -•"+'-•'•' '''^i
DataQuest
J i B I ThcDjmKWadsbrtlConioralKin
Research Newsletter
SIS Code:
Newsletters 1989
DSP
0003271
DSP AND THE WAVE OF NEW RISC PRODUCTS
INTRODUCTION
The major news topic of the last six months in the semiconductor field has been the
reduced-instruction-set computer (RISC) processor. Every major semiconductor and
computer company now has a stated position on RISC in its product lines. Although
much of the news is market posturing, there can be little question that the new RISC
microprocessors represent a significant next generation for general-purpose computing.
Independent of instruction set size, they embody the latest thinking from both computer
science and market demand in today's VLSI technologies. Such opportunities for a
completely fresh start on a processor architecture occur only rarely. Some suppliers are
taking better advantage of the chance than others.
Any change in the general-purpose microprocessor market is important to digital
signal processing (DSP) because it is consistently estimated that half of the volume of
the integrated circuits used in DSP are conventional microprocessors. Their low price,
wide familiarity, and variety of support tools always make them an attractive
alternative to the higher-performance digital signal processor (DSMPU) solution. In
addition, many of the mips and mflops performance figures of the new RISC processors
are close to those expected only from digital signal processors. Even a casual glance
shows Rise's architectural similarities to DSMPUs, such as deep data pipelining, the
Harvard-style separation of instruction and data memories, and multiple execution
units. RISC processor manufacturers are even talking about the same embedded
controller markets that have been the domain of high-performance DSMPUs and about
things like real-time operating systems.
What does this mean for the suppliers and users of single-chip DSMPUs in the
future? We will explore that question in this newsletter. In addition, we will review
some of the basic performance requirements for digital signal processors and see how
these are met by the major new RISC processors. Next, we will look at the latest
generation of high-performance DSMPUs and see how both are moving to solve some
common new systems requirements. Comparing the two types of processors leads to
some strategies for DSMPU makers to protect and expand their markets in the face of
this potentially strong competition.
© 1989 Dataquest Incorporated March—Reproduction Prohibited
The content of this report represents our interpretation and analysis qfinprmatitm generally available to the public or released by responsible individuals in the subject corrqHtnies, but
is not guaranteed as to accuracy or completeness. It does not contain material provided to us in a)nfidence by our clients. Individual corr^xmies reported on and analyzed by Dataquest
may be clients cfthis and/or other Dataquest services. This ir^rmation is notfiimished in c(mnection with a sale or offer to sell securities or in connection with the solicitation of an
o^r to buy securities. This firm and its parent ard/or their officers, stockholders, or members of their pmiUes may, fnrni time to time, have a long or short position in the securities
mer^ioned and may sell or buy sudi securities.
Dataquest Incorporated, 1290 Ridder Park Drive, San Jose, CA 95131-2398 / (408) 437-8000 / Telex 171973 / Fax (408) 437-0292
DSP REQUIREMENTS AND THE NEW RISC PRODUCTS
Historically, signal processors have been distinct from
microprocessors because of the following three major requirements:
•
Higher precision
•
Higher speed
•
Special functions operating on large amounts of data
general-purpose
Table 1 lists these requirements with some of the architectural or implementation
techniques used to meet them. The first two columns indicate which of these techniques
has generally been used in current-generation CISC microprocessors and in the first- or
second-generation DSMPUs. The next two columns show the RISC processors and the
latest (or third generation) of high-performance DSMPUs. Following the historical
requirements are the additional DSP requirements considered to be important today as
the result of larger, more complex systems.
Just looking at the relative predominance of the Xs over the Os in Table 1 confirms
the general trends: New CISC microprocessors have few attributes other than precision
to make them suited for DSP, whereas even the first-generation DSMPUs are a
significant improvement. New higher-performance DSMPUs are complete in their use of
such techniques, while RISC, for its own performance needs, has used more than even the
early DSMPUs. The Xs and Os represent only rough averages across a number of
products. Table 2 shows some of the specific features and performance parameters for
four representative RISC products. Three high-performance DSMPUs are included for
comparison on the same basis.
© 1989 Dataquest Incorporated March
SIS Newsletter
Table 1
Product Overlap in Meeting DSP Requirements
DSP
Sequirements
Implementation
Technique
CISC
Products
DSP-1&2 RISC DSP-3
Historical DSP
Requirements
High precision
8-bit
16-bit
32-bit
Floating-point
High speed
Pipelined data path
Parallel operation
Data memories
Instruction memory
I/O controller
Address generators
Fixed and floating po int
Loop counters
Full processors paral leled
Memory speed-size hierar chy
Instruction caching
I/O buffering
Special processing
Large amounts of data
Complex address generation
Vector
2-D
Arithmetic
Multiply-accumulate
Saturation
Large memory space
High-speed I/O
Real-time control
X
X
X
0
0
X
0
0
X
X
X
X
X
X
X
X
0
0
0
0
0
0
0
X
X
X
X
0
X
0
X
X
0
0
X
0
X
X
X
X
X
X
X
X
0
0
0
0
X
X
X
X
0
0
X
0
0
0
X
X
0
0
X
X
X
0
X
X
0
0
0
0
0
X
X
0
X
X
X
X
X
X
X
X
Hew DSP Requirements
High-level languages
Operating systems
Industry standard
functions
X
X
Source:
SIS Newsletter
© 1989 Dataquest Incorporated March
Dataquest
March 1989
Table 2
Major New RISC and DSMPU Feature and Performance Summary
RISC
SPARC
AMD
7C601
29000
8-32
32, 64
8-32
Ext
50
iMfl
TI
320C30
DSP
AT&T
32C
Motorola
9^002
8-32
Ext
8-32
32, 64
16, 32
32, 40
8-24
32, 40
32, 64
32, 96
40
30
30
60
80
75
1
2
1
1
N/A
1/2
1
N/A
1
1
2
1
2
S
1
1
1
2
2
S
2
3
Y
3/2
H
2
y
4
Y
3
Y
3
N
3
Y
32
S
Ext
Ext
192
N/A
Ext
Ext
136
N/A
Ext
Ext
32
32
y
Y
16
S
N
Y
22
4
N
N
10
S
N
Y
N
Y
N
N/A
N
N/A
N
Y
Y
S
Y
Y
Y
Y
30
40
S
66
Y
Y
32
S
132
y
Y
2x10
12
24 (S)
132
Y
Y
2x9
10
24 (S)
50
Y
N
32
S
32
106
Y
Y
Motorola
88100
Precision
Integer
F-P
Speed
Cycle time (ns)
Execution unit
data pipelines
Integer
F-P
I/O
Concurrent data
pipelines
Parallel processors
Memory hierarchy
Integer RF
F-P RF
Data cache
Inst, cache
Special processing
Address generation
Multiplieraccumulator
Integer
F-P
Intel
N
Address space (bits)
Data 1
Data 2
Instruction
I/O bandwidth (MB/sec)
Interrupts
Context switch
30
80
Y
Y
32
S
50
Y
Y
High-level language
1
Y
Y .
Y
Y
Y
Y
R-T operating system
Y
Y
y
N
Y
N
N
Ext
N/A
Y =
N =
S =
= External
= Not Applicable
Yes
No
Shared
Source:
© 1989 Dataquest Incorporated March
Dataquest
March 1989
SIS Newsletter
Precision and Speed
Precisions are at rough parity now. DSMPUs tend to preserve more bits for
accumulation, but RISC processors often have greater word length flexibility, which can
be useful for DSP image data. RISC meets the precision need. Basic data pipeline cycle
times are shorter for RISC processors, and the difference is real for small vectors.
Nevertheless, address generation times in the RISC integer ALU and data memory
bottlenecks reduce performance for most signal processing operations below that of the
DSMPUs. As the number of separate pipelines in the execution units and the
concurrency figures show, however, the differences may not be large. Note that the
64-bit data busses of the i860 give it higher large-vector performance than the DSMPUs
due to concurrency. Thus, RISC can meet many DSP speed requirements now. So-called
vector processors are being considered by RISC suppliers now to provide muUiport
address generation for large real data memories to increase DSP and vector
performance, so the gap may narrow in the future. NEC has even announced such a
vector processor for its V-series product line, which employs traditional complexinstruction-set computer (CISC) architectures. RISC processors for the moment seem to
lead DSPs in providing for paralleling of complete processors.
Memory Hierarchy
RISC processors have large data register files that, for most functions, equate to
the much larger separate data memories on the DSMPUs. Concurrent load/store I/O
operations on the RISC processor can reduce this size difference; however, speed may
degrade quickly due to I/O bottlenecks. The large number of registers or accumulators
on the DSMPUs reflect the desire to support high-level language compilers. Caching of
instruction memory is used in both RISC processors and DSMPUs, although the modes of
operation are much different. A low-cost, low-complexity solution is possible with a
RISC processor that is sufficient to meet signal processing needs. Data caching is
handled overtly with partitioned memories and programmed control in DSMPUs rather
than "automatically" as in RISC processors. The large on-chip data cache on the i860
with its 128-bit bus is a real performance booster for signal processing operations.
Overall the RISC memory hierarchy may seem ill suited for signal processing, but it can
be scaled down and be cost and performance effective for large DSP systems.
Special Processing
The addition of vector processors to RISC processors may more nearly even the
score, but now DSMPUs clearly excel at the concurrent and complex address generation
needed in large data spaces for signal processing. This extends to I/O with DMA
controllers as well as for on-chip memory. The concurrent multiply-accumulate
arithmetic function so central to DSP is not common in RISC except in the
floating-point execution units. This directly affects DSP speed performance on the RISC
processors.
SIS Newsletter
© 1989 Dataquest Incorporated March
Large Amounts of Real-Time Data
The important address space change for RISC is to separate data and instruction
spaces for higher performance. DSMPUs have increased the size of both spaces in order
to handle the larger programs from high-level languages and the graphics and image data
bases. DSMPU memory spaces have become more linear (like RISC) as they have gone
off-chip. Thus, RISC processors can meet the separate and large memory space
requirements of current signal processing systems. DSMPU I/O bandwidths remain
higher than RISC processors and generally can be more fully utilized, but RISC I/O rates
exceed many early DSMPUs and can be sufficient in many DSP systems.
RISC processors have interrupts, stacks, and other context-switching hardware
assists, but they often lack the deterministic response times necessary for real-time
DSP. Cypress Semiconductor is moving to improve this in its implementation of the
SPARC architecture, and it seems likely that others will also. RISC processors, likewise,
have the more complete high-level language support but not in a real-time operating
system environment.
TODAY'SfflGH-PERFORMANCESYSTEMS AND THEIR MARKETS
This growing similarity between digital signal processors and general-purpose RISC
microprocessors results from manufacturers of these products recognizing the needs of
an increasingly common high-performance system. Figure 1 is a block representation of
such systems. It represents functional blocks of the typical new high-performance
systems and their varied CPU processing and software requirements. Tj^jically, some
physical process that generates a large amount of data is analyzed or controlled by
computations on the data. The computations are altered by operator controls, often
interactively, from results that are presented on a display. The display itself often
involves much processing, as does the final output result on some peripheral device.
For economic reasons, and because not all processing is simultaneous, a single CPU
is desired. Speed is important because of the large amount of data, the fact that the
system is interactive, and the fact that it often must be real time in the strictest sense
for closed-loop control purposes. The speed must be in I/O as well as arithmetic
functions to support displays and the data collection.
Large amounts of high-level language applications code are used, often running
under UNIX. This user- and third-party-supplied software accommodates industry
standards processing and standard I/O peripherals, drivers, and formats. The high-level
language improves maintainability, but often it is used initially because it allows the
function to be transported in to get the system operating in a minimum amount of time.
Critical time to market is improved.
Typical applications that use these systems are listed in Figure 2. Frequently, they
are referred to collectively as high-performance embedded controller systems. Note
that high-performance workstations in this context are a subset with less demanding
real-time I/O.
$
© 1989 Dataquest Incorporated March
SIS Newsletter
Figure 1
New High-Performance Systems and Their Varied
CPU Processing and Software Requirements
Display
CPU
Transducers
Compute Core
-*"
Control
Display Generation""'
\t/
Controls
In Data Reduce
i'Dut Data Format.
Physical Process
1/0
Perripherals
0003271-1
Source: Dataquest
March 1989
Figure 2
Important High-Performance Markets for RISCs and DSMPUs
Medical Tomography Imaging
Ultrasound Imaging
Communications Instrumentation
Vibration Testing
Electrical, Chemical, and Mechanical Design, Simulation,
and Analysis Workstations
Image Scanning and Electronic Publishing
0003S71-2
SIS Newsletter
Source: Dataquest
March 1989
© 1989 Dataquest Incorporated March
THE COMPETITIVE THREAT AND NEW DSP STRATEGIES
Few significant quantity shipments of RISC processors occur today, except for
workstation shipments, and it will be two years before the important product families
and markets can be confirmed. However, the prudent DSP product strategist cannot
wait for market erosion to react.
Dataquest believes that certain DSP performance issues are important ones for
DSMPU suppliers trying to maintain their markets. DSMPU suppliers must continued to
do the following:
•
Accommodate the real-time nature of DSP operations—The first requirement
is to continue to accommodate the real-time nature of the processing while
adapting to the need for operating system and high-level language benefits.
This can be done through integrated hardware assists and real-time software
function libraries that support industry standards and device independence yet
do not get in the way of the other real-time processing required. Developing a
standardized library of real-time functions and a suite of DSP performance
measures, like the recent SPEC benchmarks, would help.
•
Support greater memory flexibility
Even with the larger data bases and programs of DSP systems today, the
memory hierarchy needed always will be different from the more
general-purpose data processing system. The need for large, multiported
nonvirtual memory always will exceed the RISC on-chip register file.
Continued attention to this memory distinction will protect DSP markets.
Vector processors that provide concurrent address generation for arrays
are expected to be added to both CISC and RISC microprocessors, but
DSMPUs always should be able to exceed the performance achieved in a
linear memory, particularly for 2-D functions and transforms like the
FFT.
•
Develop workable multiprocessor languages and interprocessor protocols—
Paralleling complete processors to increase computing power is everyone's
candidate for the next major leap in performance, yet progress has been very
slow in systems that can be used today. Because DSP is so amenable to
partitioning between parallel processors, it can take the lead in simple,
workable languages and interprocessor communications conventions.
•
Emphasize high-bandwidth, real-time I/O—A final area of emphasis for DSP
should be input/output (I/O). Graphics and imaging have made I/O dataflow an
issue for all processors; however, the serial telecommunications interfaces,
complex multiplexing/demultiplexing, and high real-time bandwidths should
allow important product distinctions.
© 1989 Dataquest Incorporated March
SIS Newsletter
DATAQUEST ANALYSIS
Dataquest believes that if DSP suppliers are successful in providing this special DSP
performance, their growth will continue and they will remain an important portion of the
semiconductor processor market. The discussion here has centered only on the
high-performance, higher-cost devices, but they represent a major growth area now and
the dominant products of the future. Failure to act could bring a repeat of the
generation-earlier contest between DSP array processors and general-purpose
mini supercomputers. In spite of FORTRAN library support and parallel processors, the
array processors lost vital market share to the more general-purpose
minisupercomputers when they had the same floating-point multiprocessor parity. The
near demise of Floating Point Systems, the leading array processor company, at the
hands of Alliant and Convex closed out the first significant generation of DSP
high-performance systems. The parallel between those rival minicomputer systems and
today's rival microprocessors bears careful attention by suppliers of DSP integrated
circuits.
Robert E. Owen
SIS Newsletter
© 1989 Dataquest Incorporated March
Dataoyest
M^k0^
llDTbcDunKBiadstrcetCorporation
Research Newsletter
SIS Code: Newsletters 1989
DSP
0003826
THE INVISIBLE DSP IC MARKET: GATE ARRAY, CELL-BASED,
CUSTOM, AND SILICON COMPILER DESIGNS
SUMMARY
Application-specific digital signal processors (ASDSPs) constitute a large and
rapidly growing segment of the general application-specific integrated circuit (ASIC)
market. The major suppliers are broad-based ASIC firms that provide little DSP support,
rather than the traditional DSP IC companies. By supporting the DSP designer better,
DSP-focused suppliers can secure some of this market, which is nearly equal in size to
the DSP microprocessor (DSMPU) market.
INTRODUCTION
DSMPUs, building blocks, and special-function DSP chips (SFDSPs) constitute a very
visible market because of the large marketing promotions for these devices. Suppliers
and users alike advertise the successful incorporation of these ICs into end products.
Almost totally invisible are the custom ASDSPs developed by product manufacturers for
their special DSP needs.
ASDSPs are a portion of the broad ASIC market and include all of the same
techniques in their design: gate array; cell based—standard cell, as well as extensions to
a microprocessor core; full-custom; or silicon compilation. However, they are distinct
within DSP IC markets from the SFDSPs such as modems and FFT chips, which are
designed and marketed broadly for specific functions rather than specific "applications."
These invisible ASDSP chip sales are very substantial, estimated at $131 million in
1988, or roughly the same as the $158 million for the highly visible DSMPUs. For many
domestic ASIC suppliers, 20 to 30 percent of their output is DSP related, with some
companies approaching 50 percent. The invisibility comes from the proprietary nature of
the business, not its lack of market importance. This newsletter looks at what is
happening in this market with the thought that its invisibility may be hiding DSP business
opportunities and important trends. We first .review the major general ASIC suppliers
and their marketing positions toward DSP. Then we examine the major DSP suppliers
and their involvement with ASDSPs.
© 1989 Dataquest Incorporated May—Reproduction Prohibited
The content cfthis report represents our interpretation and analysis (rf information generally available to the public or released by responsible individuals in the subject companies, but
is not guaranteed as to accuracy or completeness It does not contain material provided U) us in confidence by our clients. Individual companies reported on and analyzed by Dataquest
may be clients of this and/or other Dataquest services. This information is notfitmished in connecticm with a sale or offer to sell securities or in connection with the solicitation of an
c^rto Imy securities This firm and its parent and/or their officers, stockholders, or members of their families may, from time to time, have a long or short position in the securities
meraioned and may sell or buy such securities.
Dataquest Incorporated, 1290 Ridder Park Drive, San Jose, CA 95131-2398 / (408) 437-8000 / Telex 171973 / Fax (408) 437-0292
PRODUCTS AND DSP MARKETING POSITIONS OF MAJOR ASIC SUPPLIERS
The general ASIC market was $7.4 billion in 1988, nearly 20 percent of the total IC
market, with a compound annual growth rate (CAGR) of 16 percent. Figure 1 shows the
worldwide sales for the major suppliers in all technologies (MOS, bipolar, and BiCMOS)
and design types (gate array, cell-based, etc.). Estimated North American ASIC
consumption by application market is shown in Figure 2. Note the prominence of the
communications and military areas, major DSP markets.
Figure 1
Estimated 1988 Worldwide ASIC Ranking
Suppliers
Fujitsu
NEC
LSI Logic
Toshiba
<x-^^^^^^:s^^>^^x^^
483
455
--^^^^^^^^^^K^<^^
367
V<X:N:-^<^^<>^^
s^^^^':>:^^'>i^-::^<S^<<^':^^^^:^^:S:^
^
360
AIVID
350
AT&T
Ti
Hitachi
National
Motorola
217
msssss^m^
^-^^^^^^^i
Cx'C^VCvvxi
ea
0003826-1
120
203
Gate Array
CBIC
173
PLD
140
ISO
240
300
Millions of Dollars
© 1989 Dataquest Incorporated May
360
420
480
540
Source: Dataquest
May 1989
SIS Newsletter
^
Figure 2
Estimated North American ASIC Consumption
by Application Market—^Total
(1989 and 1994)
KXM
Data Processing
Communications
Industrial
Consumer
IVIilitary
Transportation
2.3%
1994
$7,431 Million'
'Excludes Full-Custom ICs
Source: Dataqueit
May 1989
0003286-2
SIS Newsletter
© 1989 Dataquest Incorpxjrated May
Gate Arrays
The largest ASIC segment is gate array designs, at $2.9 billion in 1988. The ranking
five suppliers are the top four overall ASIC companies—Fujitsu, NEC, LSI Logic, and
Toshiba—plus Hitachi. None, with the exception of LSI Logic, have significant design
aids for DSP users. The closest thing is macrofunctions of the AMD Am2900 series
building blocks, which are popular in DSP. LSI Logic, the top CMOS gate array supplier
with an estimated 25 percent DSP business, has the MACGEN compiler for generating
multiplier-accumulators of varying precision and arithmetic formats. Although backed
up with a full arithmetic and functional simulator, it still lacks specific DSP features
like overflow saturation and coverage of the often complex address generation and
microprogramming functions needed for a full processor.
Cell-Based Designs
The smallest but fastest growing segment of the ASIC market is the so-called
standard cells segment. In 1988, revenue was $1.3 billion, with AT&T, Texas
Instruments, Toshiba, NCR, and VLSI Technology as the top-ranked suppliers. Growth
was 43 percent last year. Here again, DSP support has been limited mostly to Am2900
series building blocks.
Full-Custom and Silicon Compilers
The second largest ($2.5 billion) portion of the ASIC market in 1988 was still the
full-custom segment, but it is declining at a 3 percent annual rate. Silicon compilation,
however, counters the overall figure with strong growth. DSP accounts for nearly half of
all silicon compiler applications because of its acceptance by large communications and
military systems companies. DSP support is a natural fit for silicon compilation, with its
emphasis on high-level functional design, but even leader Silicon Compiler Systems, Inc.,
provides no specific DSP support.
The motivation for a full-custom design is often proprietary design protection and
cost, but it also can be the high performance that DSP requires. The largest custom
suppliers are NEC, Matsushita, Sharp, and Toshiba. Although much of their output is for
consumer products (e.g., ultrasonic autofocus controllers for cameras), the companies
are often solving DSP problems. That trend should continue as consumer products
become smarter. Philips, the large European consumer products firm, estimates that
half of its custom silicon output is for DSP functions.
MAJOR DSP SUPPLIERS' ROLES WITH ASDSPs
Dominant DSP supplier Texas Instruments has surely leveraged its position with
application-specific designs, but these designs have been mostly full-custom done with
internal design resources. One that became visible is the TMS 320C20, now a standard
product, which grew from specific speech processing requirements at ITT. But Texas
Instruments does not actively encourage ASDSPs, particularly those that involve users in
any active role in their design. The new TMS 320C30 has a modular layout and a future
as a processing core, but it is not a major thrust at this time.
© 1989 Dataquest Incorporated May
SIS Newsletter
Number two DSP supplier NEC has produced a myriad of DSP designs in most DSP
applications, such as speech recognition, signal encoding and image processing, as well as
in more experimental areas like data-flow processors. Most have been cell-based
designs to keep development costs low and design times short. However, these devices
have been mostly for internal telecommunications requirements, with no public attempt
to secure ASDSP business using the cell libraries.
Similarly, Fujitsu has supplied a large internal telecommunications need with
cell-based designs. It has had less commercial success with standard products. Perhaps
because of this, Fujitsu has now made available its processing cores and cell-based
peripherals and memory configurations in the MB8220/232 product line.
AT&T, a major internal ASIC and DSP supplier, has not used its limited commercial
success with standard parts to expand its ASDSP business. Motorola and Analog Devices
have no ASIC programs in DSP, even though Motorola's 56200 was a silicon compiler
design that could presumably have been the start of an application-specific filter
business.
TRW LSI Products is understood to have replaced much of its loss of merchant
market share with custom DSP designs using its own design teams. There are no tools
for public use. AMD's lack of participation in the general ASIC market has kept the
company from capitalizing on the Am2900 series building blocks.
DATAQUEST CONCLUSIONS
The distinction between DSP and general-purpose data processing is becoming
blurred, but clearly a large portion of the fastest growing segment of the IC business,
ASICs, is DSP related. Dataquest expects ASDSP to be a $181 million market in 1989
(see Table 1). The major participants in this business are the traditional ASIC suppliers
rather than the DSP IC firms. Business is being secured in spite of not having device
libraries or support tailored to DSP designer needs. At this time, users are limited to
sophisticated users who do not require much support.
The major DSP suppliers, although they do high-volume,
full-custom,
application-specific designs, have not pursued this business either. Because their own
standard products have usually been custom designed, they have not internally developed
the libraries or tools that would assist them in the public ASDSP market. They also
might view an aggressive ASDSP program as eroding the programmable solutions with
their standard products in which they have made such an extended investment. This
explains their cautious approach of expanding from a programmable core processor for
ASDSPs. Within large IC companies, DSP and ASIC are often separate divisions, with
many organizational forces working to impede cooperation on a workable strategy. Even
in a narrowly focused company like LSI Logic, the DSP effort has been an attempt to
establish a viable standard product line (something new for the company) rather than to
strengthen its position in the ASDSP market.
SIS Newsletter
© 1989 Dataquest Incorporated May
i
Table 1
Application-Specific DSP (ASDSP) Market
(Millions of Dollars)
Estimated
1990
1987
1988
$68
$98
$131
CAGR
1986-1992
37.6%
1989
Forecast
1990
1991
1992
$181
$250
$461
$340
Source:
Dataquest
May 1989
The ASDSP market is understandably undersupported at this time. Because both
general ASIC suppliers and DSP firms have been growing rapidly, they have had other
more important tasks. Each type of company would have to master a new set of skills to
solidify a position, but as competition increases, some company will likely move to claim
ASDSPs as its own. ASIC houses would seem to have a head start, but DSP
manufacturers may have the strongest motivation.
As DSP increasingly becomes possible on general-purpose, particularly RISC,
processors, a quick-response, application-specific approach to the remaining diversified
DSP market will be necessary. Cell-based designs seem the best design approach today,
besides being a good basis for any long-term plan for DSMPUs, or special-function or
building block DSP standard products.
Robert E. Owen
© 1989 Dataquest Incorporated May
SIS Newsletter
Dataoyest
aoHnpanyof
TlicDun&Biadslicet Corporation
jij^fii^s&r'
i^'^ih
* T *^s-
m
Research Newsletter
SIS Code:
Newsletters 1989
DSP
0003555
NEW DSP PRODUCTS AND TRENDS AT ISSCC '89 AND CICC '88
SUMMARY
Significant new products were shown in all digital signal processing (DSP) segments
at the major semiconductor conferences during the last year. Most new DSP products
were in the video and image special-function segment. Recent product, architecture,
and technology trends continued with programmed, functional block, and
application-specific solutions coexisting. Reduced-instruction-set computing (RISC)
processors and their floating-point coprocessors invited comparison with highperformance signal processors.
INTRODUCTION
The first public awareness of significant new integrated circuit products usually
comes through papers presented at the International Solid State Circuits Conference
(ISSCC) in February or the Custom Integrated Circuits Conference (CICC) in May. This
certainly has been true for DSP, where whole sessions are usually devoted to the topic.
An important part of Dataquest's DSP research is coverage and interpretation of new
products and related technologies that are described at these two conferences. We also
will be reporting on signal processing technology advances and products introduced at the
International Conference on Acoustics, Speech, and Signal Processing (ICASSP) in April.
Of
obvious
interest
are
the new DSP microprocessors
(DSMPUs),
microprogrammable building blocks (MPDSPs), special-function DSPs (SFDSPs) for video
and imaging, and application-specific circuits for DSPs (ASDSPs) in fields such as
telecommunications. But related products like analog/digital converters, highperformance microprocessors, and coprocessors also affect DSP markets. Both
conferences provide a guide to the latest design methodologies that can be important in
fitting DSP techniques to new application needs rapidly. New semiconductor
technologies described often impact DSP products early because of their need for the
highest speeds and high density. Our purpose here is to collect those items that seem
important for a thorough understanding of DSP product and technology directions.
Completeness of coverage is considered more important than details and comparisons of
specific products, due to the preliminary nature of this early information.
© 1989 Dataquest Incorporated April—Reproduction Prohibited
The contera c^this report represents our interpretation and analysis of information generally avaihble to the public or released by responsible individuals in the subject companies, but
is not guananleed as to accuracy or cxjmpletertess h does not (xmtain material pravidai to us in amfidence by our clients. Individual amipanies reported on and analyze by Dataquest
may be clieras cfthis and/or other Dataquest services This information is notjumishai in connection with a sale or q^r (o sell securities or in connecti/m with the solicitation cfan
offer to buy securities This firm and its parera and/or their officers, stoddiolders, or members of their families may, from time to time, have a long or short position in the securities
mentioned and may sell or buy such securitiesk
Dataquest Incorporated, 1290 Ridder Park Drive, San Jose, CA 95131-2398 / (408) 437-8000 / Telex 171973 / Fax (408) 437-0292
ISSCC '89
DSP Microprocessors
Only one truly general-purpose, single-chip DSMPU was described at either
conference; it was an upward-compatible, third-generation enhancement from
Mitsubishi. However, Mitsubishi's new product's speed at 40ns per instruction cycle for
24-bit floating-point is noteworthy. The 24-bit (16E8) precision is an increase over the
initial 18-bit (12E6). Other differences are a shared on-chip data and instruction cache
and shared external memory space, in addition to the normal external instruction-only
and data-only memories. Caching is increasingly being seen in DSPs, and here it includes
a novel clock-scaling circuit to allow it to be easily loaded from slower external memory.
Although described as video processors, a 24-bit integer unit also from Mitsubishi
and a 16-bit integer processor from NEC are really full-function programmable
DSMPUs. Both are impressively fast, and as Table 1 shows, they have all the functions
of the general-purpose Mitsubishi device except the serial telecommunication interface.
The chief distinction between the video processors and the DSMPU is the richness of
address generation capability, a welcome addition for most applications on a DSMPU,
along with the faster speed. Prices will be higher for these two larger-size chips.
Special-Function DSPs
Video Processors
The special function receiving the most attention this year, as it has for the past
several years, was video (see Table 2). This follows from the increased interest in HDTV
and the establishment of ISDN video compression standards. The most flexible product is
the 24-bit integer processor from Mitsubishi (also shown in Table 1). The three data
execution units supported by two large dual-ported memories with three address
generators driven by the 48-bit instructions provide very high performance. The data
precision is high enough for transforms, yet the instruction set also makes it data byte
efficient. Address generation is two-dimensional both on- and off-chip. Performance
figures were given for a large number of video functions, but full attention was given to
the video codec requirements for transforms, vector quantization, and motion
compensation.
© 1989 Dataquest Incorporated April
SIS Newsletter
Table 1
Significant New DSP Micrc^rocessors (DSMPUs)
at ISSCC '89
Mitsubishi
Mitsubishi
Precision
24-bit
FP (16E8)
SI integer
24-bit
integer
16-bit
integer
Instruction Cycle
40ns
50ns
25ns
Arithmetic Element Stages
ALU
MPY
ACC
1
1
0
1
1
1
1
1
1
4Kx32
60Kx32
512x48
16Kx48
512x32
-
Fegityireg
Address Generators
Memories
Instruction
Internal ROM
Internal RAM
External
Data & Instruction
Internal RAM
External
Data
Internal
Internal
External
External
64x32
4Kx32
—
—
512x24 (DP)
60Kx24
—
512x24 (DP)
512x24 (DP)
64Kx24
-
128x16
128x16
lMxl6
lMxl6
Input/Output
Serial
Parallel
1-bit
24-bit
24-bit
16-bit
Technology
Process
Feature Size
Transistor Count
Pin Count
Die Size (mm)
CMOS
lu
300K
135
7.0 X 8.6
CMOS
lu
538K
177
13.8 X 15.,5
CMOS
1.2u
220K
176
14.0 X 13.4
Notes:
RAM 1
RAM 2
1
2
FP = Floating-Point, DP = Dual-Port
Source: Dataquest
April 1989
SIS Newsletter
© 1989 Dataquest Incorporated April
f
Table 2
Significant New Special-Function DSPs (SFDSPs)
at ISSCC '89 and CICC '88
Function
Company
Video
Mitsubishi
Instruction
Cycle Time
Precision
Transistors
Broad/
microprograiranable
Broad/
microprogrammable
40ns
24-bit
538K
25ns
16-bit
220K
2-D FFT
2-D DCT
Codec
Color correction
100ns
70ns
41ns
70ns
11-bit
12-bit
8-bit
14-bit
152K
73K
288K
g4K
FIR filter
Template match
Rank value
Delay line
MAC
Correlation
16 MAC array/
selectable
MAC (BiCMOS)
50ns
50ns
50ns
50ns
40ns
40ns
50ns
8-bit
1-bit
12-bit
8-bit
18/32-bit
32-bit (FP)
12-bit
240K
94K
140K
llOK
340K
90K
124K
5ns
16-bit
20K
Fast Fourier
Transform (FFT)
Plessey
FFT, weighting
25ns
16-bit
500K
Filter
Fujitsu
Adaptive filter
100ns
16-bit
42K
NEC
MicroElectronics
Center
Bellcore
Toshiba
Kodak
Image
LSI Logic
Siemens
Sony
NEC
Notes:
DCT = Discrete Cosine Transform/ FP = Floating-Point
Source:
© 1989 Dataquest Incorporated April
Dataquest
April 1989
SIS Newsletter
The NEC chip is a 16-bit integer processor (also shown in Table 1) and is
correspondingly faster than Mitsubishi's at 25ns. It compares with a simpler three-stage
pipeline, 70ns version presented two years ago. The seven stages of the newest chip are
in a sequence that is useful for most video tasks. NEC has provided for multiprocessor
synchronization, because even with these data rates, more than one processor may be
necessary for real-time performance.
These two jwogrammable processors, with their writable control stores for
instructions, are contrasted with the video processors from Toshiba and Kodak, which are
fixed in function. Toshiba's video codec uses the time-compressed integration format for
HDTV conversion in either direction. The link may be either analog, 400-Mbps digital,
or video disk. For an established format, it offers functional flexibility in one large but
potentially high-volume part. Kodak's processor serves a more narrow function—it
provides the digital processing from color filter CCD arrays to create correct
three-color video signals. The slower speed is acceptable because of the smaller number
of pixels in current CCD arrays.
Image Processors
Video processing generally is seen as a subset of image processing that
accommodates in some manner the specific real-time bandwidth requirements of video
signals and makes use of the sequential raster nature of data and its storage. For more
basic functions like multiply-accumulate, the distinction may not be meaningful. Sony's
processor containing 16 multiplier-accumulators (MACs) and adders and NEC's BiCMOS
MAC are described as video building blocks, while Siemens' products are aimed at
"real-time images." The Siemens two-chip set and the Sony processor are programmed
processors but fit together like building blocks. The input MAC is integer for 6-bit
neighborhood data, while the arithmetic processing unit is 32-bit floating-point for full
image 2-D operations like correlation. There are three concurrent floating-point data
execution units and concurrent I/O. Sony's processor's limited functions, all based on the
inner product operation, can be combined in systolic arrays to boost the 1.04 Gigaoperations per second (GOPS) for a single chip. NEC, in one-fourth the chip area and
with 65 percent of the power consumption, achieves 0.2 GOPS or one-fifth the
performance—more of a BiCMOS test vehicle than part of any video or imaging product
strategy.
Fast Fourier Transform (FFT) Products
With the flurry of new FFT chips recently introduced by Austek, Honeywell, TRW,
UTC, and Zoran, interest was high in the new PDSP 16510 from Plessey. Its most
notable feature is that it will do a full 1,024-point complex transform using only internal
memory and in less than lOOus. Most other products use large, fast, and expensive
external memories and/or multiple processors to do transforms this large. Plessey's
version is also very flexible in transform size, real or complex data type, use of data
buffers, and selection of weighting functions. The 13.1 x 13.3mm area will make it not
be cost competitive with DSMPUs for small transforms; however, in the many
applications requiring midsize transforms, this product will be very attractive in price,
size, and complexity of design. Plessey's FFT was the most significant pure DSP product
introduced at ISSCC '89.
SIS Newsletter
© 1989 Dataquest Incorporated April
General-Purpose Microprocessors
Coprocessors
The major public product announcement at ISSCC was of course Intel's i860, and it
was of not just a little interest to the DSP community. The hype was about RISC, UNIX,
and standalone operation, but competitors' concerns were about a high-performance
$750 coprocessor, and that is how we have compared it in Table 3. The significant trends
among the coprocessors are toward multiple, fast pipeline stages; multiple execution
units including I/O; and wider-than-32-bit data paths. All are moves toward DSP-like
performance in a standard microprocessor system. This was particularly evidenced by
the concurrent multiply-add and the vector address generation capability of NEC's new
coprocessor for the V60, 70, and 80 series.
RISC Processors
RISC processors are the engines that drive coprocessors with high clock rates and
wide 64-bit busses. A Digital Equipment Corporation spokesperson said in his session 7.1
presentation that a vector processor was a future part of this (non-VAX) family. The
Matsushita chip, with its four parallel execution units, clearly rivals the i860 in power,
and it, too, has features for multiprocessing.
Building Blocks
Multipliers are the traditional regular-function, medium-complexity test vehicles
for any new technology, so it is not surprising to see some appear now in gallium arsenide
(GaAs). Seldom are they significant in commercial DSP markets. However, Honeywell's
GaAs multiplier introduced at ISSCC clearly was designed for DSP because it does a full
16-bit complex multiplication (four multiplies and two additions) and does it in only 8ns.
Mitsubishi's 32-bit floating-point building block is highly pipelined with five 10ns stages
and is self-timed. This technique, where data is passed from stage to stage only when
the processing is complete, is bound to see wider use as multiple variable-length data
pipelines become common. Mitsubishi does not use this technique in its purest form, but
the company has a head start at learning about its use. This product was the most
significant DSP circuit/architecture innovation presented at ISSCC '89.
© 1989 Dataquest Incorporated April
SIS Newsletter
Table 3
Significant New General-Purpose Microprocessors
and Building Blocks (MPDSPs) at ISSCC '89
Functional
Unit
Coprocessors
Intel i860*
Pipeline
Delay
Pipeline
Depth
Precision
Integer ALU
FP ALU
FP MPY
I/O
30ns
30ns
30ns
60ns
1
3
3
1-3
32-bit
32/64-bit
32/64-bit
64-bit
NEC
FP ALU
FP MPY
50ns
50ns
8
9/11/12
32/64/80-bit
32/64/80-bit
GE
32-bit Integer or
FP ALU or MPY or MAC
I/O
25ns
25ns
3
1
32/64-bit
32-bit
FP ALU or MPY
20ns
4/5
32/64-bit
Integer ALU
20ns
32/64-bit
Matsushita
Integer ALU
FP MPY
FP MPY
I/O
50ns
50ns
50ns
50ns
24-bit
64-bit
64-bit
64-bit
Digital (7. 3)
Integer ALU
I/O
28ns
28ns
32-bit
64-bit
HP
Integer ALU
33ns
32-bit
Digital (7. 5)
Integer ALU
I/O
40ns
40ns
32-bit
64-bit
Integer MPY
(cotnpl(3z)
2ns
3
16-bit
Integer & FP ALU
Integer & FP MAC
10ns
10ns
5
5
32-bit
32-bit
Digital
RISC Processors
Digital (7. 1)
Building Blocks
Honeywell
(GaAs)
Mitsubishi
Notes:
ALU = Arithmetic Logic Unit, FP = Floating-Point, MAC = MultiplierAccumulator, MPY = Multiplier
*Intel's i860 processor is considered a standalone RISC processor with
integrated floating-point and 3-D graphics. For purposes of this
comparison, we will focus only on the floating-point portion of the i860.
Source:
SIS Newsletter
© 1989 Dataquest Incorporated April
Dataguest
April 1989
Applications
It is always interesting to check on the migration of DSP techniques into chips for
complete applications. There is a concern that all of the high-volume applications will
be met this way and that there will be no lasting market for the general-purpose
programmable solutions. Oki's modem chip showed that this high-volume application's
requirements can be met with a DSMPU core and specialized on-chip A/D and D/A
peripheral circuits.
Panel Discussions
The regular "Future of DSP" panel session was less partisan and more penetrating
than usual. A major topic was questioning the need of the latest large, full-featured
DSMPUs. Is there a place for other than "lean and mean" (simple and fast) processors?
As part of that topic, the usual floating-point versus integer discussion ensued, but this
time it ended differently than usual. AT&T virtually acknowledged that floating-point
had been put in its DSP-32 because it was the next step and not because it was clearly
justified. And sure enough, AT&T said, the most successful application had been graphics
transformations, not a traditional DSP function at all.
CICC '88
Products
At last year's Custom Integrated Circuits Conference, the DSP products were all
special-function products, as one might expect, and there was more information about
DSP-specific design methodologies. Video and imaging dominated again. For example,
Bellcore's discrete cosine transform chip does 16 x 16 pixels in real-time for video
coding. It is a multiplierless design using only additions and ROM lookup tables for
compactness.
The MicroElectronics Center (MEC) 2-D FFT array of chips processes 256 x 256
pixels in real-time at a 30Hz rate. It uses the long-forgotten shift register method of
FFT data sequencing to advantage.
Two papers by LSI Logic introduced a family of 20-MHz image-processing chips that
has since grown to six and increased in speed. They were the most significant DSP
papers presented at CICC '88 because in one short span of time LSI has made available a
complete processor and memory building block set for imaging. This came from a
thorough product line plan, fast turnaround design tools, and a commitment to provide
what was needed even if it resulted in large chips. The FIR filter, for example, has
64 MACs and is 1.4cm on a side. Rarely is there such a complete product thrust into a
market that is in an early development stage. The commercial battle line is clearly
drawn now between programmable processors and functional blocks in this DSP
application area.
Fujitsu described an adaptive filter that offers improved I/O and multichannel
processing over a DSMPU. It is unusually well supported with design software and
documentation.
© 1989 Dataquest Incorporated April
SIS Newsletter
Design Methods
CICC papers have more discussion of design tools than circuit design details, and the
silicon compiler session always includes DSP because DSP data processing is more
regular and amenable to description and synthesis. The University of California at
Berkeley and IMEC in Belgium are both centers of much of this compiler work, and their
progress is steady, with effects being seen in commercial products at Philips and LSI
Logic. No company is visible yet that has made silicon compilation a cornerstone of its
DSP product strategy.
DATAQUEST CONCLUSIONS
What did these two important conferences say about the major DSP issues being
watched today? They confirmed first that general-purpose microprocessors in the form
of RISC and their coprocessors are getting function and performance levels close to
high-end DSMPUs. Secondly, in the growing area of video/image processing, a
special-function DSP segment, there are still new product examples of programmable,
functional block, and application-specific implementations. These three forms may
continue to coexist with no clear indication of a single product "best" strategy.
The product trends for each of the segments were as follows:
•
DSMPU—Little new; the third generation is still being digested.
•
SFDSP—Video/image is the function getting the most attention recently.
•
MPDSP—No major products except at new technology frontiers.
•
ASDSP—Core DSP processors are expanding into this area more
compilers/design tools are adapting to DSP.
than
The architectural trend is to separate highly pipelined data execution units for
higher performance and more run-time assists to control them. One micron is the new
technology norm, and BiCMOS' impact, if any, is yet to be felt in DSP.
Generally, DSP integrated circuit progress is strong and healthy, but one feels a
little cautious when all of the excitement is about general-purpose processors with
speeds and precisions that were only so recently the sole domain of DSP.
Robert E. Owen
SIS Newsletter
© 1989 Dataquest Incorporated April
Dataqyest
acon^nyof
The Dun & Bradstrect Corporation
Research Newsletter
SIS Code:
Newsletters 1989
DSP
0003555
NEW DSP PRODUCTS AND TRENDS AT ISSCC '89 AND CICC '88
SUMMARY
Significant new products were shown in all digital signal processing (DSP) segments
at the major semiconductor conferences during the last year. Most new DSP products
were in the video and image special-function segment. Recent product, architecture,
and technology trends continued with programmed, functional block, and
application-specific solutions coexisting. Reduced-instruction-set computing (RISC)
processors and their floating-point coprocessors invited comparison with highperformance signal processors.
INTRODUCTION
The first, public awareness of significant new integrated circuit products usually
comes through papers presented at the International Solid State Circuits Conference
(ISSCC) in February or the Custom Integrated Circuits Conference (CICC) in May. This
certainly has been true for DSP, where whole sessions are usually devoted to the topic.
An important part of Dataquest's DSP research is coverage and interpretation of new
products and related technologies that are described at these two conferences. We also
will be reporting on signal processing technology advances and products introduced at the
International Conference on Acoustics, Speech, and Signal Processing (ICASSP) in April.
Of
obvious
interest
are
the new DSP microprocessors
(DSMPUs),
microprogrammable building blocks (MPDSPs), special-function DSPs (SFDSPs) for video
and imaging, and application-specific circuits for DSPs (ASDSPs) in fields such as
telecommunications. But related products like analog/digital converters, highperformance microprocessors, and coprocessors also affect DSP markets. Both
conferences provide a guide to the latest design methodologies that can be important in
fitting DSP techniques to new application needs rapidly. New semiconductor
technologies described often impact DSP products early because of their need for the
highest speeds and high density. Our purpose here is to collect those items that seem
important for a thorough understanding of DSP product and technology directions.
Completeness of coverage is considered more important than details and comparisons of
specific products, due to the preliminary nature of this early information.
© 1989 Dataquest Incorporated April—Reproduction Prohibited
The content of this report represents our interpretation and analysis of information generally available to the public or released by responsible individuals in the subject amipanies, but
is not guaranteed as to accuracy or awipleteness. It does not contain material provided to us in confidence by our clients. Individual a>mpames reported on and analyzed by Dataquest
may be clients of this and/or other Dataquest services. This informatitm is notfiimished in connection with a sale or offer to sell securities or in connection with the solicitation of an
offer to buy securities. This firm and its parent and/or ^ir officers, stockholders, or members cff their fomilies may, frtmi time to time, have a long or short position in the securities
mentioned and may sell or buy sudi securities.
Dataquest Incorporated, 1290 Ridder Park Drive, San Jose, CA 95131-2398 / (408) 437-8000 / Telex 171973 / Fax (408) 437-0292
ISSCC '89
DSP Microprocessors
Only one truly general-purpose, single-chip DSMPU was described at either
conference; it was an upward-compatible, third-generation enhancement from
Mitsubishi. However, Mitsubishi's new product's speed at 40ns per instruction cycle for
24-bit floating-point is noteworthy. The 24-bit (16E8) precision is an increase over the
initial 18-bit (12E6). Other differences are a shared on-chip data and instruction cache
and shared external memory space, in addition to the normal external instruction-only
and data-only memories. Caching is increasingly being seen in DSPs, and here it includes
a novel clock-scaling circuit to allow it to be easily loaded from slower external memory.
Although described as video processors, a 24-bit integer unit also from Mitsubishi
and a 16-bit integer processor from NEC are really full-function programmable
DSMPUs. Both are impressively fast, and as Table 1 shows, they have all the functions
of the general-purpose Mitsubishi device except the serial telecommunication interface.
The chief distinction between the video processors and the DSMPU is the richness of
address generation capability, a welcome addition for most applications on a DSMPU,
along with the faster speed. Prices will be higher for these two larger-size chips.
Special-Function DSPs
Video Processors
The special function receiving the most attention this year, as it has for the past
several years, was video (see Table 2). This follows from the increased interest in HDTV
and the establishment of ISDN video compression standards. The most flexible product is
the 24-bit integer processor from Mitsubishi (also shown in Table 1). The three data
execution units supported by two large dual-ported memories with three address
generators driven by the 48-bit instructions provide very high performance. The data
precision is high enough for transforms, yet the instruction set also makes it data byte
efficient. Address generation is two-dimensional both on- and off-chip. Performance
figures were given for a large number of video functions, but full attention was given to
the video codec requirements for transforms, vector quantization, and motion
compensation.
© 1989 Dataquest Incorporated April
SIS Newsletter
Table 1
Significant New DSP Microprocessors (DSMPUs)
at ISSCC '89
NEC
Mitsubishi
Mitsubishi
Precision
24-bit
FP (16E8)
& integer
24-bit
integer
16-bit
integer
Instruction Cycle
40ns
50ns
25ns
Arithmetic Element Stages
ALU
MPY
ACC
1
1
0
1
1
1
1
1
1
4Kx32
60Kx32
512x48
16Kx48
512x32
—
Features
Address Generators
Memories
Instruction
Internal ROM
Internal RAM
External
Data & Instruction
Internal RAM
External
64x32
4Kx32
Data
Internal
Internal
External
External
512x24 (DP)
60Kx24
-
512x24 (DP)
512x24 (DP)
64Kx24
—
128x16
128x16
lMxl6
lMxl6
Input/Output
Serial
Parallel
1-bit
24-bit
24-bit
16-bit
Technology
Process
Feature Size
Transistor Count
Pin Count
Die Size (mm)
CMOS
lu
300K
135
7.0 X 8.6
CMOS
lu
538K
177
13.8 X 15..5
CMOS
1.2u
220K
176
14.0 X 13.4
Notes:
RAM 1
RAM 2
1
2
FP = Floating-Point, DP
—
-
Dual-Port
Source:
SIS Newsletter
© 1989 Dataquest Incorporated April
Dataquest
April 1989
Table 2
Significant New Special-Function DSPs (SFDSPs)
at ISSCC '89 and CICC '88
Function
Company
Video
Mitsubishi
NEC
MicroElectronics
Center
Bellcore
Toshiba
Kodak
Image
LSI Logic
Siemens
NEC
Transistors
Precision
Broad,
microprogrammable
Broad,
microprogrammable
40ns
24-bit
538K
25ns
16-bit
220K
2-D FFT
2-D DCT
Codec
Color correction
100ns
70ns
41ns
70ns
11-bit
12-bit
8-bit
14-bit
FIR filter
Template match
Rank value
Delay line
50ns
50ns
50ns
50ns
40ns
40ns
50us
8-bit
1-bit
12-bit
8-bit
18/32-bit
32-bit (FP)
12-bit
140K
llOK
340K
5ns
16-bit
20K
MAC
Correlation
16 MAC array,
selectable
MAC (BiCMOS)
Sony
Instruction
Cycle Time
152K
73K
288K
94K
240K
94K
90K
124K
Fast Fourier
Transform (FFT)
Plessey
FFT, weighting
25ns
16-bit
500K
Filter
Fujitsu
Adaptive filter
100ns
16-bit
42K
Notes:
DCT = Discrete Cosine Transform, FP = Floating-Point
Source:
© 1989 Dataquest Incorporated April
Dataquest
April 1989
SIS Newsletter
The NEC chip is a 16-bit integer processor (also shown in Table 1) and is
correspondingly faster than Mitsubishi's at 25ns. It compares with a simpler three-stage
pipeline, 70ns version presented two years ago. The seven stages of the newest chip are
in a sequence that is useful for most video tasks. NEC has provided for multiprocessor
synchronization, because even with these data rates, more than one processor may be
necessary for real-time performance.
These two programmable processors, with their writable control stores for
instructions, are contrasted with the video processors from Toshiba and Kodak, which are
fixed in function. Toshiba's video codec uses the time-compressed integration format for
HDTV conversion in either direction. The link may be either analog, 400-Mbps digital,
or video disk. For an established format, it offers functional flexibility in one large but
potentially high-volume part. Kodak's processor serves a more narrow function—it
provides the digital processing from color filter CCD arrays to create correct
three-color video signals. The slower speed is acceptable because of the smaller number
of pixels in current CCD arrays.
Image Processc»-s
Video processing generally is seen as a subset of image processing that
accommodates in some manner the specific real-time bandwidth requirements of video
signals and makes use of the sequential raster nature of data and its storage. For more
basic functions like multiply-accumulate, the distinction may not be meaningful. Sony's
processor containing 16 multiplier-accumulators (MACs) and adders and NEC's BiCMOS
MAC are described as video building blocks, while Siemens' products are aimed at
"real-time images." The Siemens two-chip set and the Sony processor are programmed
processors but fit together like building blocks. The input MAC is integer for 6-bit
neighborhood data, while the arithmetic processing unit is 32-bit floating-point for full
image 2-D operations like correlation. There are three concurrent floating-point data
execution units and concurrent I/O. Sony's processor's limited functions, all based on the
inner product operation, can be combined in systolic arrays to boost the 1.04 Gigaoperations per second (GOPS) for a single chip. NEC, in one-fourth the chip area and
with 65 percent of the power consumption, achieves 0.2 GOPS or one-fifth the
performance—more of a BiCMOS test vehicle than part of any video or imaging product
strategy.
Fast Fourier Transform (FFT) Products
With the flurry of new FFT chips recently introduced by Austek, Honeywell, TRW,
UTC, and Zoran, interest was high in the new PDSP 16510 from Plessey. Its most
notable feature is that it will do a full 1,024-point complex transform using only internal
memory and in less than lOOus. Most other products use large, fast, and expensive
external memories and/or multiple processors to do transforms this large. Plessey's
version is also very flexible in transform size, real or complex data type, use of data
buffers, and selection of weighting functions. The 13.1 x 13.3mm area will make it not
be cost competitive with DSMPUs for small transforms; however, in the many
applications requiring midsize transforms, this product will be very attractive in price,
size, and complexity of design. Plessey's FFT was the most significant pure DSP product
introduced at ISSCC '89.
SIS Newsletter
© 1989 Dataquest Incorporated April
General-Purpose Microprocessors
Coprocessors
The major public product announcement at ISSCC was of course Intel's i860, and it
was of not just a little interest to the DSP community. The hype was about RISC, UNIX,
and standalone operation, but competitors' concerns were about a high-performance
$750 coprocessor, and that is how we have compared it in Table 3. The significant trends
among the coprocessors are toward multiple, fast pipeline stages; multiple execution
units including I/O; and wider-than-32-bit data paths. All are moves toward DSP-like
performance in a standard microprocessor system. This was particularly evidenced by
the concurrent multiply-add and the vector address generation capability of NEC's new
coprocessor for the V60, 70, and 80 series.
RISC Processors
RISC processors are the engines that drive coprocessors with high clock rates and
wide 64-bit busses. A Digital Equipment Corporation spokesperson said in his session 7.1
presentation that a vector processor was a future part of this (non-VAX) family. The
Matsushita chip, with its four parallel execution units, clearly rivals the i860 in power,
and it, too, has features for multiprocessing.
Building Blocks
Multipliers are the traditional regular-function, medium-complexity test vehicles
for any new technology, so it is not surprising to see some appear now in gallium arsenide
(GaAs). Seldom are they significant in commercial DSP markets. However, Honeywell's
GaAs multiplier introduced at ISSCC clearly was designed for DSP because it does a full
16-bit complex multiplication (four multiplies and two additions) and does it in only 8ns.
Mitsubishi's 32-bit floating-point building block is highly pipelined with five 10ns stages
and is self-timed. This technique, where data is passed from stage to stage only when
the processing is complete, is bound to see wider use as multiple variable-length data
pipelines become common. Mitsubishi does not use this technique in its purest form, but
the company has a head start at learning about its use. This product was the most
significant DSP circuit/architecture innovation presented at ISSCC '89.
© 1989 Dataquest Incorporated April
SIS Newsletter
Table 3
Significant New General-Purpose Microprocessors
and Building Blocks (MPDSPs) at ISSCC '89
Functional
Unit
Coprocessors
Intel i860*
Pipeline
Delay
Pipeline
Depth
Precision
Integer ALU
FP ALU
FP MPY
I/O
30ns
30ns
30ns
60ns
1
3
3
1-3
32-bit
32/64-bit
32/64-bit
64-bit
FP ALU
FP MPY
50ns
50ns
8
9/11/12
32/64/80-bit
32/64/80-bit
32-bit Integer or
FP ALU or MPY or MAC
I/O
25ns
25ns
3
1
32/64-bit
32-bit
FP ALU or MPY
20ns
4/5
32/64-blt
Integer ALU
20ns
32/64-bit
Matsushita
Integer ALU
FP MPY
FP MPY
I/O
50ns
50ns
50ns
50ns
24-bit
64-bit
64-bit
64-bit
Digital (7.3)
Integer ALU
I/O
28ns
28ns
32-bit
64-bit
HP
Integer ALU
33ns
32-bit
Digital (7.5)
Integer ALU
I/O
40ns
40ns
32-bit
64-bit
Integer MPY
(complex)
2 ns
3
16-bit
Integer S. FP ALU
Integer & FP MAC
10ns
10ns
5
5
32-bit
32-bit
NEC
GE
Digital
RISC Processors
Digital (7.1)
Building Blocks
Honeywell
(GaAs)
Mitsubishi
Notes:
ALU = Arithmetic Logic Unit, FP = Floating-Point, MAC = MultiplierAccumulator, MPY = Multiplier
^Intel's i860 processor is considered a standalone RISC processor with
integrated floating-point and 3-D graphics. For purposes of this
comparison, we will focus only on the floating-point portion of the i860.
Source:
SIS Newsletter
© 1989 Dataquest Incorporated April
Dataguest
April 1989
Applications
It is always interesting to check on the migration of DSP techniques into chips for
complete applications. There is a concern that all of the high-volume applications will
be met this way and that there will be no lasting market for the general-purpose
programmable solutions. Oki's modem chip showed that this high-volume application's
requirements can be met with a DSMPU core and specialized on-chip A/D and D/A
peripheral circuits.
Panel Discussions
The regular "Future of DSP" panel session was less partisan and more penetrating
than usual. A major topic was questioning the need of the latest large, full-featured
DSMPUs. Is there a place for other than "lean and mean" (simple and fast) processors?
As part of that topic, the usual floating-point versus integer discussion ensued, but this
time it ended differently than usual. AT&T virtually acknowledged that floating-point
had been put in its DSP-32 because it was the next step and not because it was clearly
justified. And sure enough, AT&T said, the most successful application had been graphics
transformations, not a traditional DSP function at all.
CICC '88
Products
At last year's Custom Integrated Circuits Conference, the DSP products were all
special-function products, as one might expect, and there was more information about
DSP-specific design methodologies. Video and imaging dominated again. For example,
Bellcore's discrete cosine transform chip does 16 x 16 pixels in real-time for video
coding. It is a multiplierless design using only additions and ROM lookup tables for
compactness.
The MicroElectronics Center (MEC) 2-D FFT array of chips processes 256 x 256
pixels in real-time at a 30Hz rate. It uses the long-forgotten shift register method of
FFT data sequencing to advantage.
Two papers by LSI Logic introduced a family of 20-MHz image-processing chips that
has since grown to six and increased in speed. They were the most significant DSP
papers presented at CICC '88 because in one short span of time LSI has made available a
complete processor and memory building block set for imaging. This came from a
thorough product line plan, fast turnaround design tools, and a commitment to provide
what was needed even if it resulted in large chips. The FIR filter, for example, has
64 MACs and is 1.4cm on a side. Rarely is there such a complete product thrust into a
market that is in an early development stage. The commercial battle line is clearly
drawn now between programmable processors and functional blocks in this DSP
application area.
Fujitsu described an adaptive filter that offers improved I/O and multichannel
processing over a DSMPU. It is unusually well supported with design software and
documentation.
© 1989 Dataquest Incorporated April
SIS Newsletter
Design Methods
CICC papers have more discussion of design tools than circuit design details, and the
silicon compiler session always includes DSP because DSP data processing is more
regular and amenable to description and synthesis. The University of California at
Berkeley and IMEC in Belgium are both centers of much of this compiler work, and their
progress is steady, with effects being seen in commercial products at Philips and LSI
Logic. No company is visible yet that has made silicon compilation a cornerstone of its
DSP product strategy.
DATAQUEST CONCLUSIONS
What did these two important conferences say about the major DSP issues being
watched today? They confirmed first that general-purpose microprocessors in the form
of RISC and their coprocessors are getting function and performance levels close to
high-end DSMPUs. Secondly, in the growing area of video/image processing, a
special-function DSP segment, there are still new product examples of programmable,
functional block, and application-specific implementations. These three forms may
continue to coexist with no clear indication of a single product "best" strategy.
The product trends for each of the segments were as follows:
•
DSMPU—Little new; the third generation is still being digested.
•
SFDSP—Video/image is the function getting the most attention recently.
•
MPDSP—No major products except at new technology frontiers.
•
ASDSP—Core DSP processors are expanding into this area more
compilers/design tools are adapting to DSP.
than
The architectural trend is to separate highly pipelined data execution units for
higher performance and more run-time assists to control them. One micron is the new
technology norm, and BiCMOS' impact, if any, is yet to be felt in DSP.
Generally, DSP integrated circuit progress is strong and healthy, but one feels a
little cautious when all of the excitement is about general-purpose processors with
speeds and precisions that were only so recently the sole domain of DSP.
Robert E. Owen
SIS Newsletter
© 1989 Dataquest Incorporated April
$
Dataquest
a company of
11ieDan& Bradsticet<Corporation
Research Newsletter
SIS Code: Newsletters 1989
DSP
0003826
THE INVISIBLE DSP IC MARKET: GATE ARRAY, CELL-BASED.
CUSTOM, AND SILICON COMPILER DESIGNS
SUMMARY
Application-specific digital signal processors (ASDSPs) constitute a large and
rapidly growing segment of the general application-specific integrated circuit (ASIC)
market. The major suppliers are broad-based ASIC firms that provide little DSP support,
rather than the traditional DSP IC companies. By supporting the DSP designer better,
DSP-focused suppliers can secure some of this market, which is nearly equal in size to
the DSP microprocessor (DSMPU) market.
INTRODUCTION
DSMPUs, building blocks, and special-function DSP chips (SFDSPs) constitute a very
visible market because of the large marketing promotions for these devices. Suppliers
and users alike advertise the successful incorporation of these ICs into end products.
Almost totally invisible are the custom ASDSPs developed by product manufacturers for
their special DSP needs.
ASDSPs are a portion of the broad ASIC market and include all of the same
techniques in their design: gate array; cell based—standard cell, as well as extensions to
a microprocessor core; full-custom; or silicon compilation. However, they are distinct
within DSP IC markets from the SFDSPs such as modems and FFT chips, which are
designed and marketed broadly for specific functions rather than specific "applications."
These invisible ASDSP chip sales are very substantial, estimated at $131 million in
1988, or roughly the same as the $158 million for the highly visible DSMPUs. For many
domestic ASIC suppliers, 20 to 30 percent of their output is DSP related, with some
companies approaching 50 percent. The invisibility comes from the proprietary nature of
the business, not its lack of market importance. This newsletter looks at what is
happening in this market with the thought that its invisibility may be hiding DSP business
opportunities and important trends. We first review the major general ASIC suppliers
and their marketing positions toward DSP. Then we examine the major DSP suppliers
and their involvement with ASDSPs.
© 1989 Dataquest Incorporated May—Reproduction Prohibited
The content of this report represents our interpretation and analysis cf information generally available to the public or released by responsible individuals in the subject companies, but
is nor guaranteed as to accuracy or completeness. It does not amtain material provided to us in confidence by our clients. Individual companies reported on and analyzed by Dataquest
may be clients of this andlor other DMaquest services. This informati<m is notfitmished in connection with a sale or offer to sell securities or in connection with the solicitation cfan
qf^r to buy securities. This firm and its parent and/or their officers, stodiholders, or members of their families rnay.fivm time to time, have a long or ^rt position in the securities
mentioned and may sell or buy such securities
Dataquest Incorporated, 1290 Ridder Park Drive, San Jose, CA 95131-2398 / (408) 437-8000 / Telex 171973 / Fax (408) 437-0292
PRODUCTS AND DSP MARKETING POSITIONS OF MAJOR ASIC SUPPLIERS
The general ASIC market was $7.4 billion in 1988, nearly 20 percent of the total IC
market, with a compound annual growth rate (CAGR) of 16 percent. Figure 1 shows the
worldwide sales for the major suppliers in all technologies (MOS, bipolar, and BiCMOS)
and design types (gate array, cell-based, etc.). Estimated North American ASIC
consumption by application market is shown in Figure 2. Note the prominence of the
communications and military areas, major DSP markets.
Figure 1
Estimated 1988 Worldwide ASIC Ranking
Suppliers
Fujitsu
48S
^v-^^^<^^^^^^^
NEC ^ ^ ^ - ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ c ^ ^ ^
455
LSI Logic
Toshiba
:ss::<^-s::m^^^^^
360
AMD
350
AT&T
Tl
Hitachi
Natlonai
Motorola
Gate Array
m^^^^?^^^
<V^X<>^<^v60
0003826-1
CBIC
120
PLD
140
laO
240
300
Millions of Ooliars
© 1989 Dataquest Incorporated May
360
420
480
540
Source; Dataquest
May 1989
SIS Newsletter
Figure 2
Estimated North American ASIC Consumption
by Application Market—^Total
(1989 and 1994)
2.1%
\-''"-:>:\ Data Processing
Communications
Industrial
Consumer
Military
Transportation
2.3%
1994
$7,431 Million*
* Excludes Full-Custom ICs
Source: Dataquest
May \ 989
0003286-2
SIS Newsletter
© 1989 Dataquest Incorporated May
Gate Arrays
The largest ASIC segment is gate array designs, at $2.9 billion in 1988. The ranking
five suppliers are the top four overall ASIC companies—Fujitsu, NEC, LSI Logic, and
Toshiba—plus Hitachi. None, with the exception of LSI Logic, have significant design
aids for DSP users. The closest thing is macrofunctions of the AMD Am2900 series
building blocks, which are popular in DSP. LSI Logic, the top CMOS gate array supplier
with an estimated 25 percent DSP business, has the MACGEN compiler for generating
multiplier-accumulators of varying precision and arithmetic formats. Although backed
up with a full arithmetic and functional simulator, it still lacks specific DSP features
like overflow saturation and coverage of the often complex address generation and
microprogramming functions needed for a full processor.
Cell-Based Designs
The smallest but fastest growing segment of the ASIC market is the so-called
standard cells segment. In 1988, revenue was $1.3 billion, with AT&T, Texas
Instruments, Toshiba, NCR, and VLSI Technology as the top-ranked suppliers. Growth
was 43 percent last year. Here again, DSP support has been limited mostly to Am2900
series building blocks.
Full-Custom and Silicon Compilers
The second largest ($2.5 billion) portion of the ASIC market in 1988 was still the
full-custom segment, but it is declining at a 3 percent annual rate. Silicon compilation,
however, counters the overall figure with strong growth. DSP accounts for nearly half of
all silicon compiler applications because of its acceptance by large communications and
military systems companies. DSP support is a natural fit for silicon compilation, with its
emphasis on high-level functional design, but even leader Silicon Compiler Systems, Inc.,
provides no specific DSP support.
The motivation for a full-custom design is often proprietary design protection and
cost, but it also can be the high performance that DSP requires. The largest custom
suppliers are NEC, Matsushita, Sharp, and Toshiba. Although much of their output is for
consumer products (e.g., ultrasonic autofocus controllers for cameras), the companies
are often solving DSP problems. That trend should continue as consumer products
become smarter. Philips, the large European consumer products firm, estimates that
half of its custom silicon output is for DSP functions.
MAJOR DSP SUPPLIERS' ROLES WITH ASDSPs
Dominant DSP supplier Texas Instruments has surely leveraged its position with
application-specific designs, but these designs have been mostly full-custom done with
internal design resources. One that became visible is the TMS 320C20, now a standard
product, which grew from specific speech processing requirements at ITT. But Texas
Instruments does not actively encourage ASDSPs, particularly those that involve users in
any active role in their design. The new TMS 320C30 has a modular layout and a future
as a processing core, but it is not a major thrust at this time.
© 1989 Dataquest Incorporated May
SIS Newsletter
Number two DSP supplier NEC has produced a myriad of DSP designs in most DSP
applications, such as speech recognition, signal encoding and image processing, as well as
in more experimental areas like data-flow processors. Most have been cell-based
designs to keep development costs low and design times short. However, these devices
have been mostly for internal telecommunications requirements, with no public attempt
to secure ASDSP business using the cell libraries.
Similarly, Fujitsu has supplied a large internal telecommunications need with
cell-based designs. It has had less commercial success with standard products. Perhaps
because of this, Fujitsu has now made available its processing cores and cell-based
peripherals and memory configurations in the MB8220/232 product line.
AT&T, a major internal ASIC and DSP supplier, has not used its limited commercial
success with standard parts to expand its ASDSP business. Motorola and Analog Devices
have no ASIC programs in DSP, even though Motorola's 56200 was a silicon compiler
design that could presumably have been the start of an application-specific filter
business.
TRW LSI Products is understood to have replaced much of its loss of merchant
market share with custom DSP designs using its own design teams. There are no tools
for public use. AMD's lack of participation in the general ASIC market has kept the
company from capitalizing on the Am2900 series building blocks.
DATAQUEST CONCLUSIONS
The distinction between DSP and general-purpose data processing is becoming
blurred, but clearly a large portion of the fastest growing segment of the IC business,
ASICs, is DSP related. Dataquest expects ASDSP to be a $181 million market in 1989
(see Table 1). The major participants in this business are the traditional ASIC suppliers
rather than the DSP IC firms. Business is being secured in spite of not having device
libraries or support tailored to DSP designer needs. At this time, users are limited to
sophisticated users who do not require much support.
The major DSP suppliers, although they do high-volume,
full-custom,
application-specific designs, have not pursued this business either. Because their own
standard products have usually been custom designed, they have not internally developed
the libraries or tools that would assist them in the public ASDSP market. They also
might view an aggressive ASDSP program as eroding the programmable solutions with
their standard products in which they have made such an extended investment. This
explains their cautious approach of expanding from a programmable core processor for
ASDSPs. Within large IC companies, DSP and ASIC are often separate divisions, with
many organizational forces working to impede cooperation on a workable strategy. Even
in a narrowly focused company like LSI Logic, the DSP effort has been an attempt to
establish a viable standard product line (something new for the company) rather than to
strengthen its position in the ASDSP market.
SIS Newsletter
© 1989 Dataquest Incorporated May
Table 1
Application-Specific DSP (ASDSP) Market
(Millions of Dollars)
198$
$68
Estimated
1997
$98
1988
$131
CAGR
1986-1992
37.6'%
1989
Forecast
1991
1990
1992
$181
$250
$340
$461
Source:
Dataquest
May 1989
The ASDSP market is understandably undersupported at this time. Because both
general ASIC suppliers and DSP firms have been growing rapidly, they have had other
more important tasks. Each type of company would have to master a new set of skills to
solidify a position, but as competition increases, some company will likely move to claim
ASDSPs as its own. ASIC houses would seem to have a head start, but DSP
manufacturers may have the strongest motivation.
As DSP increasingly becomes j)ossible on general-purpose, particularly RISC,
processors, a quick-response, application-specific approach to the remaining diversified
DSP market will be necessary. Cell-based designs seem the best design approach today,
besides being a good basis for any long-term plan for DSMPUs, or special-function or
building block DSP standard products.
Robert E. Owen
© 1989 Dataquest Incorporated May
SIS Newsletter
Dataoyest
aoMnpanyof
The Dttn& Bradsticct Cwporabcm
Research Newsletter
SIS Code:
Newsletters 1989
DSP
0003271
DSP AND THE WAVE OF NEW RISC PRODUCTS
INTRODUCTION
The major news topic of the last six months in the semiconductor field has been the
reduced-instruction-set computer (RISC) processor. Every major semiconductor and
computer company now has a stated position on RISC in its product lines. Although
much of the news is market posturing, there can be little question that the new RISC
microprocessors represent a significant next generation for general-purpose computing.
Independent of instruction set size, they embody the latest thinking from both computer
science and market demand in today's VLSI technologies. Such opportunities for a
completely fresh start on a processor architecture occur only rarely. Some suppliers are
taking better advantage of the chance than others.
Any change in the general-purpose microprocessor market is important to digital
signal processing (DSP) because it is consistently estimated that half of the volume of
the integrated circuits used in DSP are conventional microprocessors. Their low price,
wide familiarity, and variety of support tools always make them an attractive
alternative to the higher-performance digital signal processor (DSMPU) solution. In
addition, many of the mips and mflops performance figures of the new RISC processors
are close to those expected only from digital signal processors. Even a casual glance
shows Rise's architectural similarities to DSMPUs, such as deep data pipelining, the
Harvard-style separation of instruction and data memories, and multiple execution
units. RISC processor manufacturers are even talking about the same embedded
controller markets that have been the domain of high-performance DSMPUs and about
things like real-time operating systems.
What does this mean for the suppliers and users of single-chip DSMPUs in the
future? We will explore that question in this newsletter. In addition, we will review
some of the basic performance requirements for digital signal processors and see how
these are met by the major new RISC processors. Next, we will look at the latest
generation of high-performance DSMPUs and see how both are moving to solve some
common new systems requirements. Comparing the two types of processors leads to
some strategies for DSMPU makers to protect and expand their markets in the face of
this potentially strong competition.
© 1989 Dataquest Incorporated March—Reproduction Prohibited
The cotaent of this report represents our interpretation mid analysis ofir^muuion ^neraUy available to the public or released by respcmsible individuals in the subject companies, but
is not guaranteed as to accuracy or amipleteness. b does not cxmtain material provided to us in atnfidence by our clients. Individual conqmnies reported on and analyzed by Dataquest
may be cliaus of this and/or other Dataquest services. This infijrmation is not furnished in connection with a sale or qffir to sell securities or in connection with //w solicitation cfan
o^r to buy securities This firm and its parent and/or their cheers, stockholders, or members of ^ir families may, from time to time, have a long or short position in the securities
mentioned and may sell or buy suck securities.
Dataquest Incorporated, 1290 Ridder Park Drive, San Jose, CA 95131-2398 / (408) 437-8000 / Telex 171973 / Fax (408) 437-0292
DSP REQUIREMENTS AND THE NEW RISC PRODUCTS
Historically, signal processors have been distinct from
microprocessors because of the following three major requirements:
•
Higher precision
•
Higher speed
•
Special functions operating on large amounts of data
general-purpose
Table 1 lists these requirements with some of the architectural or implementation
techniques used to meet them. The first two columns indicate which of these techniques
has generally been used in current-generation CISC microprocessors and in the first- or
second-generation DSMPUs. The next two columns show the RISC processors and the
latest (or third generation) of high-performance DSMPUs. Following the historical
requirements are the additional DSP requirements considered to be important today as
the result of larger, more complex systems.
Just looking at the relative predominance of the Xs over the Os in Table 1 confirms
the general trends: New CISC microprocessors have few attributes other than precision
to make them suited for DSP, whereas even the first-generation DSMPUs are a
significant improvement. New higher-performance DSMPUs are complete in their use of
such techniques, while RISC, for its own performance needs, has used more than even the
early DSMPUs. The Xs and Os represent only rough averages across a number of
products. Table 2 shows some of the specific features and performance parameters for
four representative RISC products. Three high-performance DSMPUs are included for
comparison on the same basis.
© 1989 Dataquest Incorporated March
SIS Newsletter
Table 1
Product Overlap in Meeting DSP Requirements
DSP
Requirements
Implementation
Technique
CISC
Products
DSP-1S2 RISC
PSP-3
Historical DSP
Requirements
High precision
8-bit
16-bit
32-bit
Floating-point
High speed
Pipelined data path
Parallel operation
Data memories
Instruction memory
X/0 controller
Address generators
Fixed and floating point
Loop counters
Full processors paralleled
Memory speed-size hierarchy
Instruction caching
I/O buffering
Special processing
Large amounts of data
X
X
X
0
0
X
0
0
X
X
X
X
X
X
X
X
0
0
0
0
0
0
0
X
X
X
X
0
X
0
X
X
0
0
X
0
X
X
X
X
X
X
X
X
0
0
0
0
X
X
X
X
Complex address generation
Vector
2-D
Arithmetic
Multiply-accumulate
Saturation
0
0
X
0
0
0
X
X
0
0
X
X
X
0
X
X
Large memory space
High-speed I/O
Real-time control
0
0
0
0
0
X
X
0
X
X
X
X
X
X
0
0
X
X
X
X
New DSP Requirements
High-level languages
Operating systems
Industry standard
functions
Source:
SIS Newsletter
© 1989 Dataquest Incorporated March
Dataquest
March 1989
Table 2
Major New RISC and DSMPU Feature and Performance Summary
i960
TI
320C30
DSP
AT&T
32C
8-32
Ext
8-32
32, 64
16, 32
32, 40
8-24
32, 40
32, 64
32, 96
40
30
30
60
80
75
1
2
1
1
N/A
1/2
1
N/A
1
1
2
1
2
S
1
1
1
2
2
S
2
3
Y
3/2
H
2
Y
4
Y
3
Y
3
N
3
Y
32
S
Ext
Ext
192
N/A
Ext
Ext
136
N/A
Ext
Ext
32
32
Y
Y
16
S
N
Y
22
4
N
N
10
S
N
Y
M
Y
N
N/A
N
N/A
H
Y
X
X
i
•t
Y
Y
30
40
S
66
Y
Y
32
S
132
Y
Y
2x10
12
24 (S)
132
Y
Y
2x9
10
24 (S)
50
Y
N
32
S
32
106
Y
Y
RTSC
Precision
Integer
F-P
Speed
Cycle time (ns)
Execution unit
data pipelines
Integer
F-P
I/O
Concurrent data
pipelines
Parallel processors
Memory hierarchy
Integer RF
F-P RF
Data cache
Inst, cache
Special processing
Address generation
Multiplieraccumulator
Integer
F-P
Motorola
AMD
SPARC
Intel
88100
29Q00
7C601
8-32
32, 64
8-32
Ext
50
Motorola
96002
H
Address space (bits)
Data 1
Data 2
Instruction
I/O bandwidth (MB/sec)
Interrupts
Context switch
30
SO
Y
Y
32
S
50
Y
Y
High-level language
Y
t
T_
Y
Y
Y
X
R-T operating system
St
Y
Y
N
Y
H
H
Ext
N/A
Y =
N =
S =
= External
= Not Applicable
Yes
No
Shared
Source:
© 1989 Dataquest Incorporated March
Dataquest
March 1989
SIS Newsletter
Precision and Speed
Precisions are at rough parity now. DSMPUs tend to preserve more bits for
accumulation, but RISC processors often have greater word length flexibility, which can
be useful for DSP image data. RISC meets the precision need. Basic data pipeline cycle
times are shorter for RISC processors, and the difference is real for small vectors.
Nevertheless, address generation times in the RISC integer ALU and data memory
bottlenecks reduce performance for most signal processing operations below that of the
DSMPUs. As the number of separate pipelines in the execution units and the
concurrency figures show, however, the differences may not be large. Note that the
64-bit data busses of the i860 give it higher large-vector performance than the DSMPUs
due to concurrency. Thus, RISC can meet many DSP speed requirements now. So-called
vector processors are being considered by RISC suppliers now to provide multiport
address generation for large real data memories to increase DSP and vector
performance, so the gap may narrow in the future. NEC has even announced such a
vector processor for its V-series product line, which employs traditional complexinstruction-set computer (CISC) architectures. RISC processors for the moment seem to
lead DSPs in providing for paralleling of complete processors.
Memory Hierarchy
RISC processors have large data register files that, for most functions, equate to
the much larger separate data memories on the DSMPUs. Concurrent load/store I/O
operations on the RISC processor can reduce this size difference; however, speed may
degrade quickly due to I/O bottlenecks. The large number of registers or accumulators
on the DSMPUs reflect the desire to support high-level language compilers. Caching of
instruction memory is used in both RISC processors and DSMPUs, although the modes of
operation are much different. A low-cost, low-complexity solution is possible with a
RISC processor that is sufficient to meet signal processing needs. Data caching is
handled overtly with partitioned memories and programmed control in DSMPUs rather
than "automatically" as in RISC processors. The large on-chip data cache on the i860
with its 128-bit bus is a real performance booster for signal processing operations.
Overall the RISC memory hierarchy may seem ill suited for signal processing, but it can
be scaled down and be cost and performance effective for large DSP systems.
Special Processing
The addition of vector processors to RISC processors may more nearly even the
score, but now DSMPUs clearly excel at the concurrent and complex address generation
needed in large data spaces for signal processing. This extends to I/O with DMA
controllers as well as for on-chip memory. The concurrent multiply-accumulate
arithmetic function so central to DSP is not common in RISC except in the
floating-point execution units. This directly affects DSP speed performance on the RISC
processors.
SIS Newsletter
© 1989 Dataquest Incorporated March
Large Amounts of Real-Time Data
The important address space change for RISC is to separate data and instruction
spaces for higher performance. DSMPUs have increased the size of both spaces in order
to handle the larger programs from high-level languages and the graphics and image data
bases. DSMPU memory spaces have become more linear (like RISC) as they have gone
off-chip. Thus, RISC processors can meet the separate and large memory space
requirements of current signal processing systems. DSMPU I/O bandwidths remain
higher than RISC processors and generally can be more fully utilized, but RISC I/O rates
exceed many early DSMPUs and can be sufficient in many DSP systems.
RISC processors have interrupts, stacks, and other context-switching hardware
assists, but they often lack the deterministic response times necessary for real-time
DSP. Cypress Semiconductor is moving to improve this in its implementation of the
SPARC architecture, and it seems likely that others will also. RISC processors, likewise,
have the more complete high-level language support but not in a real-time operating
system environment.
TODAY'SfflGH-PERFORMANCESYSTEMS AND THEIR MARKETS
This growing similarity between digital signal processors and general-purpose RISC
microprocessors results from manufacturers of these products recognizing the needs of
an increasingly common high-performance system. Figure 1 is a block representation of
such systems. It represents functional blocks of the typical new high-performance
systems and their varied CPU processing and software requirements. Typically, some
physical process that generates a large amount of data is analyzed or controlled by
computations on the data. The computations are altered by operator controls, often
interactively, from results that are presented on a display. The display itself often
involves much processing, as does the final output result on some peripheral device.
For economic reasons, and because not all processing is simultaneous, a single CPU
is desired. Speed is important because of the large amount of data, the fact that the
system is interactive, and the fact that it often must be real time in the strictest sense
for closed-loop control purposes. The speed must be in I/O as well as arithmetic
functions to support displays and the data collection.
Large amounts of high-level language applications code are used, often running
under UNIX. This user- and third-party-supplied software accommodates industry
standards processing and standard I/O peripherals, drivers, and formats. The high-level
language improves maintainability, but often'it is used initially because it allows the
function to be transported in to get the system operating in a minimum amount of time.
Critical time to market is improved.
Typical applications that use these systems are listed in Figure 2. Frequently, they
are referred to collectively as high-performance embedded controller systems. Note
that high-performance workstations in this context are a subset with less demanding
real-time I/O.
© 1989 Dataquest Incorporated March
SIS Newsletter
Figure 1
New Highr-Performance Systems and Tlieir Varied
CPU Proce^ing and Software Requirements
c
Display
t
]
CPU
Computa Cora
Control
^ t f
\ • /
Controis
Display Generation
•-)h Data Reduce
Out Data Format .-::*:;
Physical Process
jk
I/O
Perripherals
0003271-1
Source: Dataguest
March 1989
Figure 2
Important High-Performance Markets for RISCs and DSMPUs
Medical Tomography Imaging
Ultrasound Imaging
Communications Instrumentation
Vibration Testing
Electrical, Chemical, and Mechanical Design, Simulation,
and Analysis Workstations
Image Scanning and Electronic Publishing
"^^^'"^""Rn'mr'^^'^-^'^^'^it^^''^'^^^^
0003271-2
SIS Newsletter
Source: Dataquest
March 1989
© 1989 Dataquest Incorporated March
f
THE COMPETITIVE THREAT AND NEW DSP STRATEGIES
Few significant quantity shipments of RISC processors occur today, except for
workstation shipments, and it will be two years before the important product families
and markets can be confirmed. However, the prudent DSP product strategist cannot
wait for market erosion to react.
Dataquest believes that certain DSP performance issues are important ones for
DSMPU suppliers trying to maintain their markets. DSMPU suppliers must continued to
do the following:
•
Accommodate the real-time nature of DSP operations—The first requirement
is to continue to accommodate the real-time nature of the processing while
adapting to the need for operating system and high-level language benefits.
This can be done through integrated hardware assists and real-time software
function libraries that support industry standards and device independence yet
do not get in the way of the other real-time processing required. Developing a
standardized library of real-time functions and a suite of DSP performance
measures, like the recent SPEC benchmarks, would help.
•
Support greater memory flexibility
Even with the larger data bases and programs of DSP systems today, the
memory hierarchy needed always will be different from the more
general-purpose data processing system. The need for large, multiported
nonvirtual memory always will exceed the RISC on-chip register file.
Continued attention to this memory distinction will protect DSP markets.
-
Vector processors that provide concurrent address generation for arrays
are expected to be added to both CISC and RISC microprocessors, but
DSMPUs always should be able to exceed the performance achieved in a
linear memory, particularly for 2-D functions and transforms like the
FFT.
•
Develop workable multiprocessor languages and interprocessor protocols—
Paralleling complete processors to increase computing power is everyone's
candidate for the next major leap in performance, yet progress has been very
slow in systems that can be used today. Because DSP is so amenable to
partitioning between parallel processors, it can take the lead in simple,
workable languages and interprocessor communications conventions.
•
Emphasize high-bandwidth, real-time I/O—A final area of emphasis for DSP
should be input/output (I/O). Graphics and imaging have made I/O dataflow an
issue for all processors; however, the serial telecommunications interfaces,
complex multiplexing/demultiplexing, and high real-time bandwidths should
allow important product distinctions.
© 1989 Dataquest Incorporated March
SIS Newsletter
s
DATAQUEST ANALYSIS
Dataquest believes that if DSP suppliers are successful in providing this special DSP
performance, their growth will continue and they will remain an important portion of the
semiconductor processor market. The discussion here has centered only on the
high-performance, higher-cost devices, but they represent a major growth area now and
the dominant products of the future. Failure to act could bring a repeat of the
generation-earlier contest between DSP array processors and general-purpose
minisupercomputers. In spite of FORTRAN library support and parallel processors, the
array processors lost vital market share to the more general-purpose
minisupercomputers when they had the same floating-point multiprocessor parity. The
near demise of Floating Point Systems, the leading array processor company, at the
hands of Alliant and Convex closed out the first significant generation of DSP
high-performance systems. The parallel between those rival minicomputer systems and
today's rival microprocessors bears careful attention by suppliers of DSP integrated
circuits.
Robert E. Owen
SIS Newsletter
© 1989 Dataquest Incorporated March
Dataoyest
aconu)any<rf
TIK Dun & wadstreetCorporabon
Research Newsletter
SIS Code: SIS/DSP Newsletters:
1988-1
February
NEW FLOATING-POINT DSP PRODUCTS THREATEN TI'S LEAD
SUMMARY
With the 1983 introduction of its first digital signal processing product, the
TMS32010, Texas Instruments (TI) established a seemingly unshakeable lead in the
emerging digital signal processing (DSP) market. Today, the heavy investments made by
TI and others in educating the technical community to the wonders of DSP are paying
off. DSP product use is becoming more pervasive. In addition, DSP product techjiology
is currently evolving from 16-bit integer products to high-performance 32-bit
floating-point products. As highlighted in Table 1, 1988 will be a banner year for DSP
microprocessor (DSMPU) product introductions.
Table 1
A Sampling of High-Performance DSMPU Products
Company
Product
Expected
Availability
AT&T
DSPS 2
DSP32C
Available now
Q2 1988
32-bit Floating Point
32-bit Floating Point
Fujitsu
MB86232
MB86220
Ql 1988
Q3 1988
32-bit Floating Point
24-bit Floating Point
Motorola
DSP56000
Available now
24-bit Integer
NEC
UPD77230
UPD77220
Available now
Q2 1988
32-bit Floating Point
24-bit Integer
Oki
M6992
M699210
Available now
Ql 1988
22-bit Floating Point
22-bit Floating Point
TI
TMS320C30
Q2 1988
32-bit Floating Point
Zoran
VSP325
Q3 1988
32-bit Floating Point
Description
Source:
Dataquest
February 1988
© 1988 Dataquest Incorporated February—Reproduction Prohibited
The corUenis of this report represent our interpretation and analysis of information generally available to [he public or released by responsible individuals in the subject companies but is
not guaranteed as to accuracy or completeness It does not contain material provided to us in confidence by our clients. Individual companies reported on and analyzed by Dataquest
may be clients of this and/or other Dataxjuest services This information is not furnished in connection with a sale or offer to sell securities or in connection with the solicitation of an
offer to buy securities This firm and its parent and/or their officers, stoddiolders, or members of their fomilies may, from time to time, have a long or short position in the securities
mention^ and may sell or buy such securities
Dataquest Incorporated, 1290 Ridder Park Drive, San Jose, CA 95131-2398, (408) 437-8000, Telex 171973, Fax (408) 437-0292
What does this mean for Texas Instruments? It means that, for the first time, TI
may not be alone with the latest technology in the DSP market. TI's TMS320C30 will
most certainly experience heavy competition from this new crop of products and the
competitors that TI faces are not new to this market, either. These vendors are all on
their second- or third-generation DSP products.
In the last five years, TI has built a strong customer-support structure. If the
TMS320C30 can be delivered close to its targeted introduction date and for its projected
price, TI shouldn't lose much ground at the high end of the DSP market. However, the
new midrange 22- and 24-bit products, which promise 32-bit floating-point performance
at 16-bit integer prices, do pose a threat in such emerging new DSP markets as digital
audio.
MARKET OUTLOOK
A half dozen vendors are trying to get a jump on Texas Instruments and stake out
some high ground in the digital signal processing market. Their products fall into the
following two broad categories:
•
High-performance 32-bit floating-point DSMPUs
•
Midrange 22- and 24-bit floating-point DSMPUs
First, there are 32-bit floating-point superchips that will compete with TI's
announced TMS320C30. These DSMPUs are some of the most powerful silicon yet
produced. They can multiply two numbers as fast as a Weitek arithmetic processor, have
more transistors than Intel's 80386, and will replace some building block systems with a
single chip.
Secondly, there is emerging a class of 22- and 24-bit floating-point DSMPUs that
promise floating-point capability at the price of midrange, 16-bit integer DSP chips.
Also not to be overlooked in this category are the 24-bit integer processors from
Motorola and NEC. These products, while lacking floating-point capability, do offer the
customer another 8 bits of dynamic range to play with.
Floating-point processors can be both easier to use and more powerful than
fixed-point processors—a great combination! All this activity in floating-point products
raises two interesting market-related questions:
•
How fast will customers move to these parts?
•
How will they affect TI's dominance of the DSP market?
In this newsletter, we will examine the products listed in Table 1 with respect to
their market potential and probable effect on Texas Instruments' lead in this market.
Located at the end of this newsletter. Table 2 compares and contrasts the features of
each product reviewed.
© 1988 Dataquest Incorporated February
SIS Newsletter
High-End Market Prognosis
The new 32-bit floating-point products will find their first uses in applications that
currently are implemented by building blocks (i.e., high-performance bit slices and
multipliers). These areas are:
•
High-end graphics
•
Imaging
•
Array processors
•
Military systems
These applications are more driven by performance than by cost. Thus, a $300 or
$400 DSP product may well provide a cost-effective solution. Because performance is
what counts in these areas, the competition is just getting started for the 32-bit
floating-point market. Products with sub-lOO-nanosecond cycle times—such as those
from AT&T, Fujitsu, and Texas Instruments—look like contenders.
While this market may be bounded by building block applications in the short term,
one would be foolish to ejqsect that to last. Forces that will help move these products
into a broad spectrum of applications include:
•
Price reductions
•
Availability of C compilers
•
An increase in customer awareness of product capabilities
The bright spot for TI's comi^titors is that engineers now using building blocks are
perhaps the most sophisticated consumers of DSP products. They will tend to choose
DSP products mainly because of performance, and they will be less influenced by TI's
imposing presence and customer-support structure.
Market Participants
AT&T. AT&T's second-generation 32-bit floating-point DSP product, the DSP32C,
appears competitive from a hardware standpoint; it is twice as fast as its predecessor.
AT&T's earlier product and NEC's initial 32-bit product, the uPD77230, are both too
slow in comparison with the newer products and limited in external memory addressing
capability,
Perhaps more importantly, AT&T is providing serious software support to its new
product. The DSP32C's 32-bit arithmetic unit is limited to floating-point operations,
however, while the Fujitsu and Texas Instruments products can also perform 32-bit
fixed-point adds, subtracts, and logical operations.
SIS Newsletter
© 1988 Dataquest Incorporated February
Fujitsu. Fujitsu's MB86232 will likely be the first of this new generation of
products. This product and the Zoran VSP325 are the only devices that directly handle
the IEEE 754 single-precision floating-point format. Additionally, the MB86232 has
highly parallel memory addressing. A big " i f is whether or not Fujitsu can overcome the
somewhat stereotypical Japanese company's weakness in supporting complicated
processors.
Texas Instruments. Texas Instruments clearly has the lead in the DSMPU market
today. Its latest product, the TMS320C30, should be as fast as the competitors' chips
with more on-chip memory. Additionally, TI's software experience and extensive
customer-support network will work to the company's advantage.
It is probably realistic to assume that the high-end 32-bit market will develop
slowly. And, as stated earlier, if the TMS320C30 can maintain its targeted introduction
date and projected price, TI shouldn't lose much ground in this market. The only
question concerns how imposing TI's lead will be. However, should the introduction date
slip appreciably, TI may find itself sharing more of the market than it had expected or
wanted.
Zoran. Zoran's VSP325, like its 16-bit predecessor, is hardwired to perform DSP
functions. This gives the VSP325 the highest performance available for applications that
it fits, but this also narrows its appeal quite a bit. However, because the 32-bit DSMPU
market is a performance-driven market, Zoran has a chance to do better than it did in
the 16-bit DSMPU market.
High-End Pricing
Initial samples of some of these 32-bit floating-point DSMPUs may cost more than
$1,000 apiece, which is a serious price for an integrated circuit. Prices are expected to
drop below $500 each (in IK quantities) before the end of the year—still a significant
price!
Midrange Market Prognosis
The threat to Texas Instruments' market domination comes not so much from the
flagship 32-bit products as it does from the new 22- and 24-bit floating-point products,
such as those from Fujitsu and Oki. If these products can be delivered for the price of
midrange 16-bit integer DSMPUs, they pose a very attractive option. Even with a
qualitatively more advanced architecture and a superior price/performance ratio,
however, strong customer support and an assertive selling effort will be needed to
capitalize on this opportunity. Figure 1 illustrates the potential DSP opportunities in
consumer electronic products that could fuel the development of a midrange DSP market.
© 1988 Dataquest Incorporated February
SIS Newsletter
Figure 1
DSP Oi^xsrtunities in Consumer Products
Compact Disc Players
Higher-Definition Television*
Digital Audio Tape Rayers
Digital Video Cassette Recorder
Radio Receiver for
PCM"* Broadcasting
Still-Vision Camera
* Refers to Improved-definition (IDTV). extended-definition (EDT\^ and
iiigh-definltion television (HDTV) standards.
** Pulse code modulated.
Source: Dataquest
February 1988
Market Participants
Fujitsu. Fujitsu's 24-bit floating-point product, the MB86220, has the potential to
be very successful. Its projected pricing is extremely competitive. In addition, its word
size and format are excellent for digital audio applications. Digital audio may become
the largest application for midrange DSP products, as design wins in consumer
applications mean big volumes.
Motorola. Motorola's DSP56000 is a 24-bit integer device. Although it doesn't have
the convenience of a floating-point device, its ALU is 8 bits wider than other integer
DSMPUs. This translates roughly to an additional 48db of dynamic range. The DSP56000
also rates as one of the fastest of the midrange products. Another plus for the product is
that people are comfortable buying advanced processors from Motorola. Comfort is a
big factor in the DSP market—a market dominated by Texas Instruments.
NEC. NEC's UPD77230 is a bit schizophrenic. It simply does not have the raw speed
of the newer 32-bit parts. Its relatively low price, however, makes it a viable
alternative for midrange applications. Another big plus for the product is that it is
available now. Die shrinks with resultant increases in clock speed could even make the
UPD77230 more competitive at the high end. Not to be ignored, of course, is NEC's
position as the number two market leader behind Texas Instruments.
NEC will soon spin off a lower-cost 24-bit integer version of this chip. The new
product, the uPD77220, will be upwardly compatible with the uPD77230. Ironically,
although floating-point products should make development less complicated, NEC feels
that customers are still just learning about floating point. This attitude further
illustrates the amount of education and selling required for DSP products.
SIS Newsletter
© 1988 Dataquest Incorporated February
Oki. Oki's 22-bit M699210 is an upgrade and die shrink of its M6992. It is a
1.5-micron CMOS full-custom design. The M6992 was a 2.0-micron standard cell
design. The die-size reduction offered by the full-custom design should move Oki's
pricing further down the learning curve, and Oki is supplying good software tools. The
company believes that engineers who have had to implement products based on the
fixed-point TMS320 family are good candidates for their floating-point devices.
Midrange Pricing
Existing midrange products, such as TI's TMS320C25 and NEC's uPD77230, are
currently priced at slightly more than $100 each in IK quantities. To be competitive in
the midrange market, prices need to drop below $50 (for moderate quantities) by the end
of the year.
DATAQUEST ANALYSIS
The Rationale for Floating Point
Telecommunications has been the single largest market for DSMPUs to date. These
applications are well served by the 16-bit integer products. Additionally, most DSP
applications require interface to the real world through A/D and D/A converters, which
are normally 8 to 16 bits in resolution. So the question is: Do customers really want or
need floating-point digital signal processors?
The answer is "yes" in a surprisingly high percentage of applications. A key point is
that each calculation results in an increased number of bits in the output. The more
calculations, the larger the resultant word. Rounding of the results to fit a 16-bit word
leads to loss of resolution. In highly iterative algorithms, this round-off error can be
quite large.
Seemingly innocuous applications can quickly outgrow a 16-bit word. For example,
suppose you wish to design a digital equalizer to work with a compact disc (CD) player.
Compact disc players use a 16-bit data word. At first glance, a 16-bit integer DSP chip
may seem a good match.
The natural way to implement a digital equalizer is by doing a fast Fourier
transform (FFT), modifying the spectrum the way you want, and then doing an inverse
FFT. An FFT can grow as much as one bit per stage; however, and we have to remember
that for an equalizer, the more bands the better. Thus, for a 256-band equalizer, the
growth could be as much as 9 bits. That means that the equalizer needs at least a 25-bit
integer DSP chip in order to ensure maximum resolution. And, becaxise people are
fanatic about audio quality, they will want those bits.
Applications that require two-dimensional FFTs, such as medical imaging, robotics,
or video data compression, experience bit growth in both the row and the column FFT
results. So an N X N transform will have twice the bit growth of a length N transform.
Therefore, even 8-bit video signals could benefit from using floating-point DSMPUs.
© 1988 Dataquest Incorporated February
SIS Newsletter
Even if resolution is not a problem, fixed-point programs usually require some
software scaling and checking for overflow. This additional code can be quite
substantial. It slows down program execution, and programmers would be quite happy
not to have to do it. Because DSP algorithms are often developed on computers that use
floating-point arithmetic, the need for scaling and overflow checking can be an
unpleasant surprise. Floating-point DSMPUs handle these issues automatically,
shortening algorithm development time.
The Threat to TI's Lead
Texas Instruments enjoys a dominant position in the DSP market—a position in
which the company has invested heavily. However, it is facing more competition than
ever before. Today's competition is seasoned, having encountered TI's immense
third-party software/hardware vendor network and customer-support structure in the
past. While TI is still recognized as the leader in customer support, the level of
customer support offered by other DSP vendors has improved with each succeeding
generation of products.
As software and hardware support lessens as an issue, the technical merits of a
product and the price/performance ratio become more important. At the high end of the
DSP market, TI's TMS320C30 compares very favorably with its competition. It is
expected to be one of the fastest products available, in addition to having more on-chip
memory than any of its competitors (see Table 2). Furthermore, the TMS320C30 is
software compatible with its predecessors in the TMS320 family. The TMS320C30 will
not be the first 32-bit floating-point DSMPU to be available; nevertheless, if its
targeted introduction date doesn't slip appreciably, TI should maintain its impressive
grasp of the high-end DSP market.
TI is vulnerable, however, in the new developing midrange market. Their current
products address the high-end 32-bit floating-point and low-end 16-bit integer markets.
TI doesn't have a midrange product similar in price/performance to those of Fujitsu and
Oki. If a midrange market such as digital audio develops, TI will be out in the cold on
two counts. The first factor is that the company lacks a midrange product. The second
factor is that such a market will be heavily influenced by Japanese consumer product
manufacturers.
Alice K. Leeper
SIS Newsletter
© 1988 Dataquest Incorporated February
Table 2
D S M P U Product Comparison
AT&T
DSP32
Supplier:
Part:
AT&T
DSP32C
Fujitsu
86232
Fujitsu
86220
Word:
Word Format:
32 FP
24E8
32 FP
24E8
32 FP
24E8
24 FP
18E6
Cycle Time:
160ns
80ns
75ns
(2 mac)
80ns
Clock:
25 MHz
50 MHz
40 MHz
25 MHz
Internal RAM:
512 X 32 X 2
512 X 32 X 3
512 X 32
256 X 24 X 2
External
Data RAM:
14K X 32
4M X 32
IM X 32
64K X 24
Internal
Program ROM:
512 X 32
IK X 32
IK X 32
2K X 30
Internal
Data ROM:
Shared
Shared
Shared
Shared
External
Program Memory
Shared
Shared
64K X 32
4K X 30
Yes
Yes
-
Wait States:
DMA:
PIG
PIG
Yes
External
Accumulators:
4 X 40
4 X 40
2 X 40
1 X 24
PIG:
1 X 8
1 X 16
1 X 32
1x8
SIO:
1
1
2
1
IK Complex FFT:
14ms
4ms
N/A
N/A
Process:
1.5 NMOS
0.75 CMOS
1.3 CMOS
1.3 CMOS
Package:
40 DIP
100 PGA
133 PGA
208 PGA
—
135 PGA
80 FPT
$170
$300
$500
$30
Estimated
IK Price:
(Continued)
«
© 1988 Dataquest Incorporated February
SIS Newsletter
Table 2 (Continued)
DSMPU Product Comparison
Motorol a
50QOQ
Supplier;
Part:
NEC
77220
NEC
77?3Q
Oki
§992
Word:
Word Format:
24 Int
N/M
32 FP
24E8
24 Int
N/M
22 FP
16E6
Cycle Time:
75ns
150ns
122ns
100ns
Clock:
26.7 MHz
13.3 MHz
16.4 MHz
40 MHz
Internal RAM:
256 X 24 X 2
512 X 32 X 2
256 X 24 X 2
128 X 22 X 2
External
Data RAM:
64K X 24 X 2
8K X 32
8K X 24
64K X 22
Internal
Program ROM:
2K X 24
2K X 32
2K X 32
IK X 32
Internal
Data ROM:
256 X 24 X 2
IK X 32
IK X 24
Shared
External
Program Memory
64K X 24
4K X 32
4K X 32
64K X 32
Wait States:
Yes
Ho
No
No
DMA:
Ho
No
No
External
Accumulators:
2 X 56
8 X 55
8 X 47
2 X 22
PIG:
1x8
Shared
Shared
Shared
SIC:
2
1
i
-!-:
IK Complex FFT:
2.6ms
12.5ms
10ms
7ms
Process:
1.5 CMOS
1.75 CMOS
1.75 CMOS
2.0 CMOS
Package:
88 PGA
88 SLAM
68 PGA
68 PGA
68 PLCC
132 PGA
-
$120
$115
$70
$165
Estimated
IK Price:
(Continued)
SIS Newsletter
© 1988 Dataquest Incorporated February
Table 2 (Continued)
DSMPU Product Comparison
Zoran
VSP325
TI
320C30
Oki
699210
Supplier:
Part:
Word:
Word Format;
22 FP
16E6
32 FP
24E8
32 FP
24E8
Cycle Time:
100ns
60ns
80ns
Clock:
40 MHz
33 MHz
25 MHz
Internal RAM:
256 X 22 X 2
IK X 32 X 2
64 X 32 X 2
External
Data RAM:
64K X 22
16M X 32
8K X 32
64M X 32
Internal Program ROM:
2K X 32
4K X 32
-
Internal Data ROM:
Shared
Shared
IK X 32 X 2
External Program Memory
Shared
Shared
Shared
Wait States:
Yes
Yes
Yes
DMA:
External
Yes
External
Accumulators:
2 X 22
8 X 40
2 X 32
PIO:
Shared
Shared
Shared
SIO:
-
2
-
IK Complex FFT:
7ms
3ms
1.7ms
Process:
1.5 CMOS
1.0 CMOS
1.5 CMOS
Package:
84 PLCC
100 Flat
$80
180 PGA
100 PGA
$400
84PGA
$395
Estimated IK Price:
H/A = Not Available
N/M = Not Meaningful
Notes:
Fastest parts are shown. FFT benchmarks are not consistent as to
radix 2 or radix 4, including or not including bit reversal, including
or not including data transfer onto and off chip. Price is estimated
only.
Source:
10
© 1988 Dataquest Incorporated February
Dataquest
February 1988
SIS Newsletter
DataQuest
I attHnpanyoT
p ThtDan^BcattdEcclCoqHiratKin
Research Newsletter
SIS Code:
DS3 1988 Newsletters: May
0000073
PERSPECTIVE ON ICASSP '88
INTRODUCTION
The thirteenth annual International Conference on Acoustics, Speech, and Signal
Processing (ICASSP) hosted by the IEEE was held at the New York Hilton from April 11
through April 14, 1988. This year boasted the largest ICASSP attendance ever, with
46 exhibitors, 1,950 attendees, and more than 700 presentations.
The main focus of ICASSP since the first conference in 1976 has been to provide a
forum where academic advancements in digital signal processing (DSP) technology can be
presented and discussed. ICASSP has remained true to its technical charter through the
years, even with the recent inclusion of exhibitor booths in the early 1980s. It remains
the premier industry conference dealing with general topics in digital signal processing.
As the name "digital signal processing" implies, DSP is a technology used for
processing signals digitally. These digital signals are often acquired from analog signals
by using an analog-to-digital converter, as illustrated in Figure 1. The output of DSP
systems is sometimes converted back into an analog form by using a digital-to-analog
converter, also shown in Figure 1. The processing of these digital signals between the
data converters is broadly categorized as digital signal processing.
Figure 1
A Generic Digital Signal Processing System
Analog-to-Digital
Converter
Digital Signal
Processing
Digital-to-Analog
Converter
Input
Source: Dataquest
May 1988
© 1988 Dataquest Incorporated May—Reproduction Prohibited
The content cf^Us Kport r^tresents our aUerpretcaUm and analysis afir^nnation generally available to tiie public or released by responsible individuals in Oie subject axnpanies, but
is not guaranteed as to accuracy or completeness. It does not contain material provided to us in conjidertce by our clients. Individual amtpanies reportal on and analyzed by Dataquest
may be clients of this and/or other Dataquest services This inprmation is rux fitmished in conneaion wUh a sale or qffir to sell securities or in connection with /Ac soUcitation of an
c^r U) buy securities TTusfirm and its parent and/or their cfficers, stoddiolders, ormendjers of their pmilies may, from time to time, have a long (^ ^tort position in /fe securities
meruitmed and may sell or buy sudi securities
Dataquest Incorporated, 1290 Ridder Park Drive, San Jose, CA 95131-2398 / (408) 437-8000 / Tfelex 171973 / Fax (408) 437-0292
The size of the transaction volumes from ICASSP '88 is testimony to the largely
infinite amount of digital processing that can be performed on signals. This year's
transactions consist of nearly 3,000 pages in five volumes that total seven inches in
thickness.
ICASSP EVOLUTION
The IEEE Acoustics, Speech, and Signal Processing (ASSP) society of today evolved
from the old IEEE Audio and Electroacoustics society. The first ICASSP conference was
held in 1976 and consisted of more than 200 papers presented and 600 participants.
Both the attendance and the number of presentations at ICASSP have more than
tripled since the first conference. Additionally, a number of other conferences focus oh
other areas in which DSP is significant, such as applications and VLSI implementations.
(As an example of DSP proliferation at other conferences, please refer to April 1988 SIS
Research Newsletter number GA4/DS2, entitled "DSP Evolution Evident at 35th ISSCC."
It highlights the fact that about 30 percent of the papers presented at the February 1988
ISSCC conference described semiconductor devices oriented toward digital signal
processing.)
In recent years, ICASSP has taken on an additional dimension with the inclusion in
the early 1980s of exhibits by a variety of semiconductor, hardware, software, systems,
and publishing vendors. Products from these companies are targeted at a variety of
people in the DSP community: research engineers, systems and algorithm engineers,
hardware designers, and software designers. The addition of exhibits adds a level of
maturity and credibility to the conference (and the industry) that should continue to
build over the next decade.
THE SESSIONS
Presentation sessions were organized into seven broad categories, as follows:
•
Digital Signal Processing
•
Spectral Estimation and Modeling
•
Speech Processing: Coding, Analysis, Synthesis, and Recognition
•
Multidimensional Signal Processing
•
VLSI for Signal Processing
•
Underwater Acoustic Signal Processing
•
Audio and Electroacoustics
The following paragraphs will examine special topics and applications of interest in
some of these categories from a variety of the sessions. Future newsletters will go into
additional detail on these and other issues.
2
© 1988 Dataquest Incorporated May
SIS Newsletter
Speech Processing
The speech processing category boasted the largest number of sessions (14) and
presentations (178). This is not a new trend, as all aspects of speech processing have
received tremendous amounts of research over the years. In fact, five to eight years
ago, most of the major U.S. semiconductor companies had engineering organizations
designing VLSI products for speech processing, including Advanced Micro Devices,
General Instrument, Motorola, National Semiconductor, and Texas Instruments. In all
cases, the VLSI activities at these companies have been disbanded. It is fair to say that
Texas Instruments still supports speech from an applications standpoint, but the
hardware implementation of speech algorithms is almost exclusively done using
general-purpose digital signal processors such as the TMS320—not using custom VLSI. In
many cases, the speech groups from these companies were redirected into more general
digital signal processing organizations.
Audio and Electroacoustics
The smallest category was Audio and Electroacoustics, with only 3 sessions and
32 presentations. In Dataquest's opinion, however, one of these sessions was one of the
most interesting at the conference. It discussed (in part) hearing and speech aids for the
handicapped.
As an introduction, according to the Ear Research Institute in Southern California,
there are 2 million people in the United States who either are totally deaf or lack the
ability to detect speech without aid. External-aid techniques in the form of either
electrical stimulation of the cochlea by implanted electrodes or tactile cutaneous
stimulation are often used in order to simulate hearing for the deaf. Another 12 million
"hearing-impaired" individuals, who suffer from serious hearing loss, often are treated
using conventional analog hearing aids. A number of different digital signal processing
techniques are being developed to aid both groups of people, providing a level of hearing
quality not achievable today using analog technology. Seven papers were presented at
ICASSP discussing a variety of these techniques.
VLSI
The second smallest presentation category was that of VLSI signal processing
implementations, containing 4 sessions and 54 papers. A related session entitled
"Hardware and Software in Signal Processing" was actually in another category, but it
will also be counted here because it is related to the VLSI category. This session added
another 20 papers.
Papers on VLSI focused largely on hardware architectures designed to solve a
variety of signal processing problems. Of special note were products that either are
available now or will soon be from a variety of different semiconductor manufacturers.
A summary of these follows.
LSI Logic described a set of four chips optimized for high-performance signal
processing applications with data sampling rates greater than 5 MHz, such as image and
radar processing. One of the processors is a 64-tap transversal filter (MFIR) containing
64 8x8 multipliers on one die. This device can operate as a one-dimensional 64-tap filter
or can be reconfigured to operate as an 8x8 two-dimensional filter. A second device in
the set is a 64-tap rank-value filter (RVF) that can also operate in both one and two
SIS Newsletter
© 1988 Dataquest Incorporated May
3
dimensions. Common operations with the RVF are minimum, median, and maximum
value filtering. A third device in the set is a binary filter and template matcher (BFIR).
It is optimized to perform 1-bit filtering, morphology, and template matching. The final
device in the set is a video shift register that will hold 4K 8-bit pixels. The lengths (in
pixels) of the video lines are configurable from as few as 8 pixels per line up to
4,000 pixels per line. This device is configurable also as a two-dimensional array of pixel
buffers to work with the other devices in the family.
Honeywell described a chip set (processor plus controller) with an architecture
optimized for high-throughput fast Fourier transforms (FFTs). The processor is able to
process data at up to 500 million arithmetic operations per second. The worst-case
performance benchmark for a IK complex FFT is 204.8 microseconds. Using different
memory configurations and cascading multiple processors, this execution time can be
reduced to 20.48 microseconds.
AT&T presented the architecture for* its new 25-mflops DSP-32C general-purpose
floating-point digital signal processor. It is an extension to the architecture of the
company's DSP-32, which has been in production for more than a year. Because the
DSP-32C has been previously announced, it will not be described in detail here.
Zoran Corporation presented the architecture of its new floating-p>oint vector signal
processor. This processor differs from many general-purpose digital signal processors in
the way it handles data. Most processors operate on single data samples in each
instruction. The vector processor is optimized to operate on multiple data samples in
each instruction, which is achieved by use of a "high-level" instruction set that is
microcoded into the processor. Its instruction set resembles that of an array processor.
For instance, single-instruction commands exist in the processor to do finite-impulse
response (FIR) filtering as well as fast Fourier transforms (FFTs). This architecture
allows very high performance when the function desired is already microcoded into the
device. The processor tends to be less efficient if the desired function is not already
microcoded into the processor.
A very important issue for designers building systems using digital signal processors
is one of software development support. In many ways, the issues facing software
developers for DSP systems parallel those facing microprocessor system developers.
Early generations of DSP processors required that users program them at the level of
assembly code (analogous to the way early microprocessors were programmed).
Programming at this level is tedious when dealing with sophisticated algorithms. On the
other hand, assembly coding of algorithms also provides the highest level of
performance, important in most DSP applications. For programmers, the ideal solution
to this dichotomy is to have available optimized high-level language (such as C or
FORTRAN) compilers that translate code efficiently to the instruction set of the
DSP microprocessor,
Recent generations of DSP microprocessors have been introduced with C compilers
to help solve these programming problems. A number of papers were presented
describing C compilers optimized for DSP microprocessors, including papers by Texas
Instruments and AT&T for their new processors.
© 1988 Dataquest Incorporated May
SIS Newsletter
Image Compression
Another prominent topic spread through a number of different sessions dealt with
algorithm and architecture implementations of the discrete cosine transform (DCT).
This mathematical transform is usually implemented in two dimensions as part of the
signal processing chain for compressing images. Image compression promises to be an
exciting area in the next few years as we begin handling and transmitting images
(pictures) using our personal computers and the telephone network. Already a number of
companies are selling commercial products that transmit images over a standard
telephone network. AT&T's original introduction of the picture-phone was only 20 years
too early!
For example, a good-quality color photograph stored on a computer hard disk
without compression can take 3 Mbytes or more. The transmission of this image using a
standard 2,400-bps modem would take 10,000 seconds—nearly three hours! Fortunately,
compression techniques such as cosine transforms can help reduce the storage or
transmission time requirements of an image by a factor of 10 to 50, depending on the
resulting quality desired. Thus, nearly three hours of transmission time can be reduced
to about three minutes.
THE EXHIBITORS
This year's ICASSP drew 46 exhibitors, more than have attended the conference at
any time in the past. Most of the major U.S. DSP semiconductor manufacturers were
present, including Analog Devices, AT&T, Microchip Technologies, Motorola, and Texas
Instruments. (Microchip Technologies is the newly named company spun off recently
from General Instrument's microelectronics division.) The two most prominent
exhibitors at the show were Texas Instruments and Motorola, both of whom had large
booths directly in the center of the exhibition area. Both of these companies also had
evening presentation sessions to discuss their recently announced 32-bit floating-point
DSP microprocessors.
In addition, Texas Instruments (TI) also had the die of its floating-point TMS320C30
on display under a microscope. The die contains some 700,000 transistors and is more
than 500 mils on a side in a 1-micron CMOS technology. Observing the die through the
microscope, one can see the impressive functional modularity designed into the die
layout. Each of the functional blocks appears on the die much as one might draw a block
diagram of the architecture on a piece of paper. The obvious implication of this
organization is the ability to customize the product for targeted application areas (or
customers) by adding or deleting unique functional blocks. The architecture supports this
concept with what TI calls its "peripheral bus."
TI announced that the TMS320C30 functional blocks will be available to customers
as standard cells in the future. It also announced that the TMS32010 architecture will be
available in the next year as an ASIC core.
The Motorola booth display was effective also. A number of Motorola's existing and
potential customers were given space in the booth to demonstrate current and future
products using Motorola's DSP processors. The company's third-party developers also
were present, demonstrating hardware and software development tools.
SIS Newsletter
© 1988 Dataquest Incorporated May
A number of independent hardware and software development companies were
present with a plethora of tools to aid the designer. Most of these companies are small,
with a limited ability to market and distribute their products. Some of these companies
are trying to make arrangements with the main DSP semiconductor manufacturers for
distribution of their products. Although considerable opportunities exist for third-party
development tools, many seem to be "me-too" kinds of products. For instance, about six
different small companies were displaying various versions of filter design programs.
Each product had its own small uniqueness relative to the others, but they were all
created to solve the same basic problem, that of designing FIR and infinite-impulse
response (IIR) filters. It seems likely that many of these companies will not be able to
survive the intense competition without finding a unique market position or distribution
channel.
One impressive development tool was displayed by a small company called
MicroWorkshop, located in Bohemia, New York. This company has a Microsoft
Windows-based development tool for the 32010 DSP microprocessor. It combines truly
easy-to-use text editor, assembler, signal editing, debug, and board driver software with
a hardware development board for the processor, all of which runs on a PC. The board
contains a 32-MHz DSP-320C10 built by Microchip Technologies of Chandler, Arizona.
MicroWorkshop promises additional development tools for other processors in the future.
DATAQUEST CONCLUSIONS
A tremendous number of interesting topics were discussed, presented, and further
advanced at this year's ICASSP. The vast majority of ICASSP presentations are generated in universities and research laboratories from around the world, however, and are
oriented toward the academic community; most will never have the ability to directly
affect our lives. To this end, most will never see ultimate implementation in hardware.
Nevertheless, it is through forums of this nature that we obtain insight into future
directions for theoretical work, applications, algorithms, and VLSI architectures that will
ultimately affect all of us. Future newsletters will address many of these topics in
greater detail.
David M. Taylor
© 1988 Dataquest Incorporated May
SIS Newsletter
Dataoyest
a compiniycrf
I 1 K Dun KwadsUcct Oxporabon
Research Newsletter
SIS Code:
DS4 1988 Newsletters: June
0000275
THE FINAL FRONTIER IN VOICEBAND MODEMS
SUMMARY
Fueled by advances in semiconductor digital signal processing (DSP) architectures,
the final frontier in voiceband modems is now beginning to unfold. This frontier provides
transmission rates of 9,600 bits per second (bps) over the standard switched telephone
network. The 9,600-bps transmission rate approaches the theoretical maximum information rate achievable over the limited bandwidth of the telephone network, effectively
prohibiting significantly higher modem data rates. Higher transmission rates eventually
will occur, but are most likely to use future digital networks such as the Integrated
Services Digital Network (ISDN).
Transmission speeds of switched-network modems have evolved rapidly in the last
ten years, moving from the primitive acoustically coupled 300-bps modem boxes that
dominated the market through the 1970s, to the sophisticated direct-connect 9,600-bps
modems now beginning to appear. It is interesting to note that the "bits-per-second"
rating for modem transmission speed has roughly doubled every two years, nearly keeping
pace with the more widely watched indicator representing the increase in capacity of
dynamic random-access memory (DRAM) chips.
A number of incompatible techniques currently exist for 9,600-bps modem
transmission over the switched telephone network. This problem is often encountered
when new, higher-speed modems are introduced. Clearly the leading contender to
ultimately dominate the personal computer marketplace at 9,600 bps is the V.32, which
was defined by the Consultative Committee for International Telephony and Telegraphy
(CCITT).
© 1988 Dataquest Incorporated June—Reproduction Prohibited
The content (^ this leporj npresents cur interpretation and analysis {if ir^rrmition generaUy availabk to the public or released by re^
individuals in the subject companies, but
is not guaranteed as to accuracy or completeness. It does not contain material provided to us in confidence by our clients. Individual companies reported on and analyzed by Dataquest
may be clients of this and/or other Dataquest services Ihis inforrmitivn is notfiimished in connection with a sale or offer to sell securities or in connection with the solicitation t^an
qf^r to buy securities. Thisfirmand its parent and/or their cheers, stockholders, or members of theirfamiliesmay, from time to time, have a long or short position in the securities
mentioned and may sell or buy such securities.
Dataquest Incorporated, 1290 Ridder Park Drive, San Jose, CA 95131-2398 / (408) 437-8000 / Tfelex 171973 / Fax (408) 437-0292
BACKGROUND
Modem technology has evolved swiftly since the Federal Communications
Commission (FCC) ruled in 1978 that any manufacturer could apply for registration to
connect its communications equipment directly to the telephone network. Until that
time, acoustically coupled 300-bps Bell 103 modems dominated the market for more than
20 years. After this important FCC decision, 1,200-bps Bell 212A modems requiring
direct connection began to gain in popularity.
Beginning in 1982, a number of major semiconductor companies (including Advanced
Micro Devices, Motorola, National Semiconductor, and Texas Instruments) began
supplying single-chip 300-bps modems. Most design approaches at that time relied on an
integrated analog design combined with switched capacitor filter technology. AMD was
a leading innovator by building a true digital signal processing architecture that included
both analog-to-digital and digital-to-analog converters on its chip, all optimized to
solve the modem function.
These new devices significantly drove down prices of low-speed box modems, giving
additional life (albeit short) to the low-speed modem market. These same devices also
allowed for the birth of the embedded personal computer modem.
The semiconductor content of 1,200-bps Bell 212A modems evolved a bit
differently. Many of the same semiconductor companies that competed in the 300-bps
marketplace also introduced single-chip Bell 212A modems beginning in 1984. However,
the early 1980s also saw the introduction of the first single-chip digital signal
processors. Texas Instruments ultimately abandoned its single-chip Bell 212A modem in
favor of implementing the modem function on its general-purpose TMS32010 DSP
microprocessor. Thus began the trend that is still important in today's higher-speed
modem technology. Modems that operate at 2,400 bps (and higher speeds) require digital
signal processing approaches to implement sophisticated functions such as adaptive
equalizers (to correct for telephone line distortions) and echo cancelers. In fact, the
modem marketplace is still the single largest consumer of DSP microprocessors.
To have broad appeal, high-speed modems must also support the popular slower
standards: V.22bis at 2,400 bps. Bell 212A at 1,200 bps, and Bell 103 at 300 bps. This is
cost-effectively implemented in firmware on a DSP microprocessor. The interest in
modems that support multiple standards extends both to leased-lines that have dial-up
standards as a backup and to fax modem standards for the electronic office.
At least nine manufacturers currently supply 2,400-bps V.22bis chip sets.
Interestingly enough, only one of these chip sets—the K224 from Silicon Systems,
Inc.—is a single-chip solution combining all of the analog and digital processing on one
device. The K224 is currently being sampled. The other chip sets are built around
ROM-coded, DSP microprocessors. TI's DSP2400 uses a TMS32011, Telebit's T24 uses a
TMS320C10, and VLSI Technology's chip set uses an ASIC version of Intel's 8096. Silicon
Systems also has a chip-set (2404) solution using a ROM-coded version of the NEC 7720.
Competition here is pushing prices down to $25 in quantities of 10,000 for a chip-set
solution. This is within $10 of a 1,200-bps, Bell 212A solution.
© 1988 Dataquest Incorporated June
SIS Newsletter
To complete the modem chip set (except for the SSI K224 product) an additional one
to three parts are necessary to implement the front-end and back-end analog
processing. TI adds a preprogrammed TMS7042 plus two analog front-end chips (a 29C19
codec plus an SC11005 filter) to the TMS32011. Telebit adds a preprogrammed 80C51
and an Oki analog front-end chip (the 6950B). VLSI Technology adds its own analog part
(the VL7C224A). Silicon Systems adds its own 73M214 front end to the 2404,
DATAQUEST ANALYSIS OF 9,600-bps MODEMS
The 9,600-bps marketplace is currently dominated by modems obeying the CCITT
V.29 half-duplex specification. The original V.29 modems were used on leased telephone
lines, but more recently have been used on the switched network as well. The largest
application for this modem is in facsimile machines where the bulk of the data
transmission is unidirectional. Dataquest estimates that the total number of 9,600-bps
modems shipped in 1987 was about 3 million, with approximately 80 percent of these in
fax machines. Only about 1 percent of the total shipments were dial-up modems. The
remaining modems were used in 4-wire and other dedicated applications.
The requirements of personal computer users, though, differ from those of dedicated
fax machine users. Personal computer users require bidirectional data transmission,
although not necessarily simultaneously. This characteristic implies that a simpler
half-duplex modem is really all that PC users require.
However, Dataquest believes that history will again repeat itself as it did in the
early 1980s when the full-duplex Bell 212A modem gained nearly exclusive market
dominance over the less-expensive, half-duplex Bell 202 modems for personal computer
applications. Dataquest expects the more-complicated CCITT V.32 modem specification
to win easily in the 9,600-bps personal computer marketplace over simpler rival modem
specifications. The reasons for this are as follows:
•
The CCITT V.32
internationally.
9,600-bps
modem
specification
has
been
ratified
•
The availability of low-cost V.32 chip-sets can and will impose standards.
•
Full-duplex data transmission protocols between modems are easier to
implement (and standardize) than the line-turnaround required for half-duplex
transmission.
•
Most incompatible V.32 modems will talk only with identical modems from the
same manufacturer.
It is important to understand, though, that alternative 9,600-bps modem standards
(primarily V.29) will not disappear. They will remain important for specific applications
because of their inherently simpler operation and lower cost. However, their future
growth rate is expected to be much lower than that of V.32 modems.
SIS Newsletter
© 1988 Dataquest Incorporated June
CCITT V.32 Modem Specification
Implementation requirements for the CCITT V.32 modem specification are certainly
the most sophisticated yet to be introduced for telephony modems. It allows 9,600-bps
full-duplex transmission over the switched telephone network. The entire telephone
bandwidth is used simultaneously for both the transmit and receive channels. This
differs from all other full-duplex modems, which use a frequency-division technique;
i.e., the transmit and receive channels split the available 3-kHz bandwidth in half, so
there is (in theory) no interference between channels.
In order for V.32 modems to use the entire telephone bandwidth simultaneously for
both transmission and reception, they employ echo cancelers in the receiver to eliminate:
•
The crosstalk caused by the local transmitter
•
The far-end echo caused by impedance mismatches in the telephone network
Also used is a sophisticated data-encoding technique called trellis coding, which embeds
error correction capability in the modulation. Trellis-coding techniques provide an
approximate 3dB signal-to-noise (SNR) performance improvement. It is the incorporation of the echo-canceling and trellis-coding techniques that make building V.32
modems more expensive.
V.32 Chip Sets
V.32 chip sets that are low enough in cost to allow 9,600-bps modems to compete in
the personal computer market will appear beginning in 1988. The most competitive of
these will comprise two $15 to $25 DSP microprocessors, one $6 to $10 analog front end,
a $4 to $15 single-chip microcontroller, and $5 to $8 worth of static RAM. These sets
will cause retail prices of 9,600-bps modems to drop precipitously from their current
range of $1,500 to $1,800, to less than $300 by 1992 for plug-in modem cards—the price
of current 2,400-bps modems. Currently, the major semiconductor companies that have
announced products for the V.32 modem marketplace include Rockwell and
SGS/Thomson. A third company, Phylon Communications, while not a semiconductor
company, has announced a V.32 modem module, which is expected to directly compete
with the chip-set solutions being introduced by semiconductor companies.
Figure 1 illustrates the historical and projected future pricing of modem chip sets
(in volume) versus time for four different transmission rates. In this case, the chip set
refers only to the data pump; e.g., the specific modem function, k o t included are costs
of the telephone line interface nor intelligent-modem features such as auto-dialing.
Note that the dramatic decline in projected prices for 9,600-bps chip sets between 1988
and 1992 follows a trend similar to the one expected for V.22bis modems between 1986
and 1990.
© 1988 Dataquest Incorporated June
SIS Newsletter
Figure 1
Volume Selling Prices of Modem Data Pump Chip Sets
Dollars
130
1986
1987
1988
1989
1990
1991
1992
Source: Dataquest
June 1988
Rockwell
Rockwell is the major high-speed manufacturer of modem chip sets. It was the first
company to offer a 2,400-bps V.22bis chip set: the R2424. Rockwell advertises that it
has shipped over 1 million V.22bis sets. Dataquest estimates that Rockwell controls
nearly 50 percent of this market. Dataquest also estimates that Rockwell chip sets are
being used in approximately 90 percent of the V.29 modems in production.
In 1987, Rockwell introduced the R9696DP, a V.32 module solution. It is the only
V.32 solution currently available to OEMs, although shipments to date have been
relatively small. The R9696DP consists of five processor chips—three proprietary DSP
microprocessors, a DSP echo canceler chip, and an analog front end chip—all on a
12-square-inch board. Initially, Rockwell should do well selling this board to high-end
modem manufacturers, who will add value such as "intelligent" features to the module.
Its current price of $350 in quantities of 1,000 is expected to spur growth for the PC
marketplace. Dataquest expects this price to drop as competition increases.
SGS/Thomson
SGS/Thomson's first V.32 offering will be a six-chip set, to be priced at $150 in
quantities of 10,000. It consists of three of SGS/Thomson's 68930 DSP microprocessors
(one for transmission, one for reception, one for echo canceling), two analog front-end
ICs, and a clock generator. In May, SGS/Thomson released the chip set, which uses
external ROMs. Samples of its internally ROM-coded devices are scheduled for the
fourth quarter of 1988. A CMOS version of its DSP microprocessor is expected to be
available in the third quarter. Porting the modem firmware to the CMOS parts should
follow by October. SGS/Thomson will also sell its echo canceler chip and analog front
ends individually.
SIS Newsletter
© 1988 Dataquest Incorporated June
Phylon Communications
Also of significance is a new product from Phylon Communications, of Santa Clara,
California. The PHY-96 is a V.32 modem module designed as a direct functional,
physical, and electrical replacement for the Rockwell V.32 module. It reportedly goes
beyond Rockwell's capabilities by also including V.22bis, Bell 212A, V.22, Bell 103, and
V.21 modem standards as part of the module. The module uses two DSP microprocessors
plus a two-chip analog front end to implement all of the modem signal processing
operations. Phylon expects to begin production shipments in the third quarter of 1988.
Other 9,600-bps Standards: The Electronic Tower of Babel
A number of box modem manufacturers are promoting their own 9,600 bps (and
higher) standards, most of which are incompatible with the others. Some of them may
even be technically superior to the V.32 although this point is arguable. However, once
V.32 chip sets are in production, Dataquest expects that volumes will grow, prices will
fall, and alternative solutions will diminish. One can recall the failure of Racal-Vadic's
proprietary pre-Bell-212A 1,200-bps modem, which was acknowledged by the technical
community to be superior in performance to the Bell 212A. It failed due in large part to
the creation of a de facto standard by the major communications company in the
world—AT&T—which preempted Racal-Vadic's superior implementation.
Telebit
The Telebit Trailblazer is currently the most technically advanced voiceband
modem. It transmits and receives data in the frequency domain using fast Fourier
transform (FFT) techniques in order to take full advantage of the bandwidth and
signal-to-noise characteristics present in a telephone connection. It adapts to the
telephone line it is using and packetizes the information it has to transmit.
In Telebit's zealousness to keep its competitive edge, the company kept its
admittedly unique technology too proprietary just as 9,600-bps modem standards were
evolving. It is the timing on just that kind of decision which can spell success or disaster
for a young company. In the case of modems, where standards are ever-important,
Telebit was not quick enough to license its technology to others. Had it done so, Telebit
might have had more influence in the evolution of the 9,600-bps modem standard.
One caveat remains for the Telebit implementation: A current proposal in front of
CCITT Study Group XVII includes the Telebit multicarrier approach as an alternative to
V.32. However, the chances of this approach gaining wide acceptance in the marketplace over the V.32 are slim indeed. Telebit's current installed base is approximately
20,000 units, whereas it is reasonable to anticipate the U.S. market to be using
200,000 V.32 chip sets annually by 1990.
Pseudo-V.32 Implementations
The concept behind pseudo-V.32 implementations is to use the V.32 modulation at
half duplex, not supporting true full-duplex transmission. Hence, when data are to go in
the other direction, the line must be "turned around." In order to simulate full-duplex
transmission, buffering of data occurs on the side that is not currently transmitting. The
S
© 1988 Dataquest Incorporated June
'
SIS Newsletter
advantage to this approach is the elimination of the echo canceler required in the
receive path of true full-duplex V.32 modems. The echo canceler is undoubtedly the
most sophisticated part of the signal processing chain. Unfortunately, most pseudo-V.32
implementations are incompatible between different manufacturers.
Hayes is certainly one of the champions of pseudo-V.32 modems. It claims that its
line turnaround is very fast and transparent to the user. U.S. Robotics was the first
company to introduce a pseudo-V.32 modem. Its method, which is incompatible with
that of Hayes, uses an asymmetrical technique that combines a 9,600-bps forward
channel and a 300-bps back channel. Still other 9,600-bps approaches are also competing
in the marketplace. For example, Microcom and Racal-Vadic use half-duplex V.29
modulation.
However, Concord Data Systems recently performed detailed tests on all of the
existing 9,600-bps approaches and presented the results to CCITT Study Group XVII. The
results indicate that under all test conditions, V.32-compliant modems perform better
than any of the other V.29 or multicarrier half-duplex modulation techniques.
Dataquest expects many of the modem companies now providing modems that are
incompatible with V.32 to begin offering V.32-compliant solutions combined with their
existing incompatible solution. In other words, a company might package its existing
incompatible modem in the same box with a new entry into the V.32 market. This keeps
these companies from alienating their existing installed bases while also addressing the
mainstream V.32 market.
High-Speed Networks Should Encourage Growth of High-Speed Modems
The proliferation of digital ISDN and Ethernet networks may actually encourage the
growth of high-speed modems. After all, even a 9,600-bps modem is slow compared to
56-Kbps ISDN or 10-Mbps Ethernet. It is generally acknowledged that both ISDN and
Ethernet will exist in the office environment long before they invade the general
switched network. While office workers will be able to communicate within a given
plant location at ISDN or Ethernet rates, communication outside that location will still
require the use of the switched network. Dataquest believes that the high rates at which
these networks transmit data will demand that switched-network transmission rates
increase to 9,600 bps.
Public data networks are both an indicator and an influence on the direction in
which switched-network modems are headed. Tymnet, one of the largest public data
networks, started using Concord Data Systems V.32 modems about three months ago.
Similarly, Dow Jones Retrieval Service has been using Concord's V.32 for almost two
years.
CompuServe has entered into a joint venture with Hayes to provide 9,600-bps
service using the Hayes V-Series 9,600-bps modems. The service is expected to be
operable within a few months. However, this Hayes modem is not V.32 compatible (as
described earlier). It is clearly an attempt by Hayes to use its clout in the modem
business to create a de facto standard using its noncompatible V.32 modem. It will be
interesting to see how events evolve for both Hayes and CompuServe as they appear to
be bucking what seems to be a growing tide toward adherence to the V.32 standard.
SIS Newsletter
© 1988 Dataquest Incorporated June
v.32 MODEM MARKET FORECASTS
Figure 2 illustrates Dataquest's projections for the growth of the dial-up modem
marketplace. The curve for 9,600-bps includes both the V.32 and the pseudo-V.32
modems that are incompatible with the specification. Of the estimated 43,000 9,600-bps
dial-up modems shipped worldwide in 1987, less than 10 percent were V.32-compliant.
However, Dataquest estimates that by 1990, nearly 90 percent of the estimated
200,000 units shipped will be V.32-compliant.
Figure 2
Estimated Worldwide Market for Dial-Up Telephone Modems
Thousands of Units
laoo1000
600
600
400
200
1987
•
•
•
9,600 bps
2,400-bps V.22 bis
1,200-bps 212A
1388
1989
1990
1991
Source: Dataquest
June 19SS
DATAQUEST CONCLUSIONS
The 2,400-bps V.22bis modems are now mass-market items, due in large part to the
low-cost DSP semiconductor chip sets used to build them. A modem that is faster,
downward compatible, and has a small price premium will displace lower-speed modems.
One can no longer find advertisements for 300-bps-only modems in personal computer
magazines. They have gone the way of the 16K dynamic RAM.
Dataquest believes that market forces are beginning to encourage 9,600-bps modem
transmission rates over the switched-telephone network. The question is no longer "if"
but "when" V.32 modems will exceed V.22bis modems in production volume for personal
computer applications. V.32 implementation issues are technically complex, but
curiously enough, the effort required is largely for algorithms and software, not
hardware. The final implementations will probably make almost exclusive use of
single-chip digital signal processors programmed for the V.32 function. Dataquest
expects many of the semiconductor companies to begin selling their own (or somebody
else's) DSP microprocessors as V.32 modem chips. There is plenty of incentive for
manufacturers to do this, as the modem market is extremely large. It is likely that
nearly 200,000 V.32 units will be shipped in 1990.
© 1988 Dataquest Incorporated June
SIS Newsletter
I
While the market for V.32 modems will be quite large, Dataquest does not expect
V.32 modems to be the only 9,600-bps modems in production. Other standards such as
V.29 will also be used, but mainly for specific applications such as facsimile.
(Statistical support for this newsletter was provided by Larry Cynar of Dataquest's
Telecommunications Industry Service.)
David M. Taylor
SIS Newsletter
© 1988 Dataquest Incorporated June
DataQuest
aomipanycrf
The Diin& Brad^reet Ccvporation
WmM
Research Newsletter
SIS Code:
1988 Newsletters:
DS5 1988-17
August
0001103
DATAQUEST DSP OPINION: CONDITIONS FOR SURVIVAL
IN THE GENERAL-PURPOSE DSP MARKETPLACE
SUMMARY
The digital signal processing (DSP) marketplace is partitioned by Dataquest into the
following four different integrated circuit product areas:
•
•
•
•
General-purpose DSP microprocessors (DSMPUs)
Microprogrammable DSP (MPDSP)
Special-function DSP (SFDSP)
Application-specific DSP (ASDSP)
Certainly the most visible of these product categories through the middle portion of
the 1980s has been DSMPUs. Revenue has blossomed from roughly $18 million in 1983 to
nearly $100 million in 1987, as shown in Figure 1. Dataquest expects DSMPU revenue to
accelerate rapidly past revenue for the relatively flat MPDSP market during calendar
year 1989. DSMPU revenue growth is expected to slow only minimally to about
47 percent from the 53 percent compounded growth it experienced from 1983 through
1987.
Figure 1
Historical and Anticipated Revenue Growth for
DSP Microprocessors from 1983 through 1992
1990
1991
1992
Source: Dataquest
August 1988
© 1988 Dataquest Incorporated August—Reproduction Prohibited
The content of this report represents our interpretation and analysis (^information generally available to the public or releasai by responsible individuals in the subject companies, but
is not guaranteed as to accuracy or completeness. It does not ccmtain material pmvided ta us in cor^idence by our dknls. Individual aympanies report^ (m and analyzed by Dataquest
may be clients cfthis andfor other Dataquest services. This ir^rmaiifm is not furnished in connection with a sale or o^r to sell securities or in connection with the solicitation of an
q^r to buy securities. This firm and its parent and/or their officers, stocklwlders, or members cf their families may, fivm time to rime, have a long or short position in the securities
mentitmed and may sell or buy such securities
Dataquest Incorporated, 1290 Ridder Park Drive, San Jose, CA 95131-2398 / (408) 437-8000 / Telex 171973 / Fax (408) 437-0292
Although such growth potential presents many opportunities for new entrants to
service this market, there are also plenty of risks involved. The remainder of this
newsletter examines some of .the issues important to the market success of
general-purpose DSP microprocessor architectures.
STATE OF THE ART
A minimum of 13 different manufacturers have introduced families of proprietary
DSMPU architectures. Although each processor family has architectural characteristics
that certainly help distinguish it from competitive products, these features often
introduce incremental advantages to designers of DSP systems. Illustrating this point,
three key DSP benchmarks are shown in Table 1 for four of the new floating-point
DSP processors introduced recently by AT&T, Motorola, Texas Instruments, and Zoran.
Table 1
Comparative Benchmarks for Four Recently Introduced
DSP Microprocessors*
Company
FIR Filter
(per tap)
AT&T DSP-32C
Motorola 96000
TI TMS320C30
Zoran ZR34325
80ns
75ns
60ns
80ns
IIR Filter
(2nd-orcler biquad)
IK Complex FFT
(radix-2 with
bit reversal)
3.2ms
2.45ms
4.12ms
1.7ms
400ns
375ns
360ns
400ns
•Performance benchmarks obtained from manufacturers
Source:
Dataguest
August 1988
As observed, the benchmarks indicating device performance are reasonably similar.
Of course, additional subtleties introduced by the architectures will affect other signal
processing operations. Generally, however, each of these processors can be used
effectively to solve a wide range of problems.
NECESSARY (BUT NOT NECESSARILY SUFFICIENT) CONDITIONS FOR SUCCESS
Dataquest believes that the long-term success of these architectures ultimately will
be determined by the product support provided by the manufacturer to the customer.
Support in this case refers to:
•
Training, education, and applications assistance
•
Hardware and software development tools including simulators, assemblers,
compilers, and emulators
•
Software libraries
© 1988 Dataquest Incorporated August
SIS Newsletter
DSP microprocessors require an even larger investment in customer development
assistance than do conventional microprocessors.
Training and Education
The required base knowledge by designers of DSP systems is quite high. The
traditional skill of microprocessor system design is essential, coupled with a knowledge
of digital signal processing theory, hardware, and software design techniques. The
number of designers possessing both backgrounds is relatively few. In order for
semiconductor companies to be successful in propagating and expanding their digital
signal processing products, they must help provide the appropriate training and education
for their customers.
Development Tools
As the architectures of many DSP products on the market are maturing, designers of
DSP systems are demanding more sophisticated hardware and software development
tools. These tools, analogous to those available for traditional microprocessor systems,
include C compilers and hardware emulators. DSP algorithm purists rightfully claim that
compilers generate less efficient code than can be generated by hand coding coupled
with an assembler. However, as the complexity of the software effort for DSP systems
increases, the traditional reasons for using high-level languages becomes ever more
important for DSP systems.
Additional tools also are important for DSP systems including software simulators
and development boards. These tools allow designers of DSP systems to develop and
debug many of their software algorithms prior to actually building a piece of hardware.
Software Libraries
DSP systems often require lengthy hardware and software development cycles.
However, shared by many systems are a few "key" DSP algorithms that are easily
identified, such as fast Fourier transforms, finite-impulse response, and infinite-impulse
response filters. Manufacturers can greatly aid the software effort of their customers
by publishing "standard" DSP algorithm code in readily available libraries. Observing the
large amount of software available for Texas Instruments DSP processors provide
testimonial to the importance of libraries.
As happened in the microprocessor marketplace in the early 1980s, and for similar
reasons, Dataquest expects the number of true winners in the DSMPU marketplace—as
judged by unit and revenue growth—to be limited to three or four. Other minor players
may exist but will be limited largely to specialty applications on the fringes of the
mainstream DSMPU marketplace.
SIS Newsletter
© 1988 Dataquest Incorporated August
J
The fallout in suppliers to the DSMPU market already has begun to occur. Probably
the most notable has been National Semiconductor's exit from the market in 1987.
Zoran Corporation has not exited the DSP business, but it is refocusing its strategy after
never having achieved the revenue growth expected after its visible market entry in 1986.
DATAQUEST CONCLUSIONS
It is important for DSMPU manufacturers to remember that a product is not simply
a device, such as a DSP processor, but rather a complete package purchased by the
customer. This package includes tangibles such as the device, development tools,
software libraries, and applications assistance; it also includes intangibles such as the
reputation and stability of the manufacturer.
Dataquest expects the fallout among general-purpose DSP microprocessor
manufacturers to accelerate over the next two years. Adoption and implementation of
the stated support requirements are not sufficient to guarantee success; they are
necessary conditions for manufacturers to be considered among the contenders for
success.
David M. Taylor
4
© 1988 Dataquest Incorporated August
SIS Newsletter
^ ^ ^ ^
Dataqyest
ilKDSn&
Corporation
Research Newsletter
SIS Code: GA4/DS2 1988 Newsletters:
0000040
April
DSP EVOLUTION EVIDENT AT 35TH ISSCC
SUMMARY
The International Solid State Circuits Conference (ISSCC), held February 17 through
19, 1988, in San Francisco, celebrated its 35th year of reporting continuing progress in
semiconductor technology. One hundred papers by more than 600 industry and academic
participants emphasized the rapid rate of evolution of IC functions. Significantly, a
record 30 percent of these were specifically oriented to digital signal processing (DSP)
chips, and another 50 percent described general-purpose functions that are directly
usable in DSP applications.
Dataquest broadly categorizes DSP IC functions as signal processing, graphics
generation, image processing, and scientific computing products. This newsletter
examines major ISSCC presentations in each of these areas, then discusses their impact
on trends in products, technology, and applications.
ANALYSIS AND DISCUSSION
Since the first 4-bit microprocessor of 1970, microprocessor chips have evolved to
32-bit word length, CMOS and GaAs processing, <1-micron design rules, clock rates
beyond 30 MHz, and throughputs measured in tens to hundreds of mops. Where
appropriate, on-board EPROM or EEPROM is used for microcontrol storage. The DSP
functions presented at this conference represent a continuation of this progress.
Some of the major developments discussed at the conference are summarized in
Table 1. All of the devices shown use CMOS processing, with transistor counts as high as
368,000. All incorporate two-layer metal and/or double-poly interconnections for
efficient use of chip area. Minimum geometries are on the order of 0.8 to 2.0 microns.
Power dissipation for each of the two largest functions is 4 watts.
© 1988 Dataquest Incorporated April—Reproduction Prohibited
The content (^ this rqjort represents our interpretation and armlysis cf ir^rmation generally available to the public or released by responsible individuals in the subject companies, but
is not guartwteed as to accuracy or completeness It does not contain rruuerial provided to us in corpulence by our clients- Individual companies reported on and arutlywl by Dataquest
may be clients cfthis and/or other Dataquest services. This ir^>rmation is not furnished in connection with a sale or offer to sell securities or in connection with the solicitation (^ an
{^ to buy securities Thisfirmand its parent and/or their cfficers, stockholders, or members of theirfamiliesmay, fivm time to time, htive a long or short position in the securities
mentioned and mtjy sell or buy such securities
Dataquest Incorporated, 1290 Ridder Park Drive, San Jose, CA 95131-2398 / (408) 437-8000 / Tfelex 171973 / Fax (408) 437-0292
Table 1
:tO>
1988 ISSCC—Major Silicon DSP Chips
Hitachi
Matsushita
F#aMi<>it
Video
signal
processing
3-D hidden
surface
processing
Video
signal
processing
Geometric
mapping
processing
Graphics
processing
Gate Count
9,000
H/A
H/A
105,000
N/A
Transistor Cottat
36,000
330,000
40,000
N/A
368,000
Process
CMOS
Twin-tub
CMOS
CMOS, 2L
poly
CMOS
Twin-tub
CMOS
o
Minimun) Fe4Lt;v^e
Size (u)
1.0
1.2
2.0
1.2
0.8
-1
tu
Metal
2L K\
2L
N/A
2L
2L
a
Die Size (cm^)
6.8 X 5.5
11.1 X 11.5
7.0 X 5.15
14.5 X 14.5
10.9 X 12.2
Clock Eate
N/A
20 MHz
27 M8z
50 MHz
40 MHz
Performance
50 mips
M/A
N/A
N/A
40M pixels/s
Pinouts (S + P|
80
57 + 7
37 + 3
280
144
Package
N/A
N/A
DIP
PIP
Ceramic FP
Voltage(s)
+5V
+5V
+5V, -5V
+5V
+5V
Power Consumpticua:
N/A
4W
500mW
4W
800mH
256-pixel
processing
Maximum
rate—30 MHz
©
00
00
NTT
National
Toshiba
D
<-f
£1
C
m
10
3
O
•3
o
r+
fD
>
n
1
V5
•—t
m
'Z
01
M
ff
Remarks
N/A = Not Available
fD
-»
Source:
A
Signal Processing Products
Companies and other organizations presenting information on developments in signal
processing products include AT&T Bell Laboratories, Fujitsu, General Electric, Hitachi,
Matsushita, National Semiconductor, NEC, Philips, Siemens, and Stanford University.
Notable are the AT&T communications signal processing chips, the 4Gb/s optical
repeater by Fujitsu, and Sony's CMOS thermal printer head LSI logic/driver chip
containing 11,448 transistors and parallel heating elementsNumerous improvements in A/D and D/A conversion were evident. Philips and
others continue to advance the state of the art of audio signal processing using digital
techniques. Several telecommunications functions were presented, including an echo
canceller and switching functions including cross-point switches. These are included in
Table 2.
Table 2
1988 ISSCC Developments in Signal Processing Chips
Chip Developments
Company
AT&T Bell Labs
45-MHz P/FLL (phase/frequency-locked
loop) timing recovery circuit,
146Mb/s time/space switch, 64 x 17
nonblocking cross-point switch,
2-GHz CMOS dual-modulus prescaler
Catholic University (Leuven, Belgium)
Micropower monolithic data acquisition
Fujitsu Opto Systems Lab
4Gb/s repeater chip set
General Electric
A mixed A/D chip for phased array
single processing
Hitachi
Voice-band DSP chip with A/D and D/A,
lOMb/s link-level CMOS processor
Matsushita ERL
GaAs programmable timer with 125ps
resolution
National Semiconductor
33Mb/s data-synchronizing phase-locked
loop (PLL)
NEC, BSR (Newton, Massachusetts)
18-bit oversampling A/D converter for
digital audio
Philips
Stereo 16-bit CMOS D/A for digital
audio, algorithmic 15-bit CMOS D/A
converter
(Continued)
SIS Newsletter
© 1988 Dataquest Incorporated April
Table 2 (Continued)
1988 ISSCC Developments in Signal Processing Chips
Chip Developments
Company
Siemens
2u CMOS equalizer for quadriture
amplitude modulation (QAM) digital
radio, self-calibrating 16-bit CMOS
A/D converter
Sony
Thermal printer head with CMOS logic
and drivers
Stanford University
Asynchronous multiplexer for
biotelemetry
University of California
at Berkeley ERL
250 Mbit/s CMOS cross-point switch
Source:
Dataguest
April 1988
Graphics-Generation Products
Charge-coupled device (CCD) chips topped the list of graphics-generation functions
discussed at the 1988 ISSCC. The largest of these, a 2 million-pixel chip by Toshiba, is
organized in high-definition television (HDTV) format, 1,920 horizontal by 1,036 vertical
pixels, allowing a 9:16 aspect ratio. The unit cell measures 7.3 x 7.6 microns, with a.
charge-handling capacity of 200,000 electrons per picture element. At F l l illumination,
a signal current of 300nA has been obtained.
Texas Instruments' 128Kx8 video RAM allows storage and manipulation of 256
individual colors or shades of gray at standard HDTV resolution and frequency. Philips
engineers have developed a 835Kb video serial memory and a 400K-pixel CCD imager,
both aimed at HDTV. Table 3 summarizes the graphics-generation product developments.
Table 3
1988 ISSCC Developments in Graphics-Generation Fimctions
Company
Chip Developments
Philips
835-Kbit video serial memory (VSM),
400K-pixel CCD imager
Sony
Comb filter
Texas Instruments
128Kx8, 70-MHz video RAM
Toshiba
2M-pixel CCD imager
Source:
© 1988 Dataquest Incorporated April
Dataguest
April 1988
SIS Newsletter
Image Processing Products
Video signal processing (VSP) functions have evolved to include 3-D manipulation
and elimination of hidden areas from processed images. Matsushita presented an
approach to thisoroblem that incorporates a skewed systolic architecture fabricated on
a 11.1 X ll.Smm^^die containing 330,000 transistors. This chip is detailed in Table 1.
Significant developments were described by the France's National Telecommunications Center, General Electric, Hitachi Central Research Laboratories,
Matsushita Central Research Laboratories, Nippon Telegraph and Telephone, Toshiba,
and Visual Information Technology, Inc. (VITI). The parallel image processor (PIP) chip
by VITI is a graphics computer incorporating the properties of two-dimensional DSP of
pixel data, pixel enhancement, and interactive processing rates. Performance of 50 mips
is achieved in CMOS with relatively conservative design rules of 2 microns. VITI's chip is
detailed in Table 1. Table 4 is a summary of image processing product developments
presented at the conference.
Table 4
1988 ISSCC Developments in Image Processing
Company
Chip Developments
French Government
27-MHz D/A VSP
General Electric
10-MHz ICs for graphics processing designed
on a silicon compiler
Hitachi CRL
20ns CMOS DSP core for VSP
Matsushita CRL
Hidden surface processor
NTT
50-MHz CMOS geometrical mapping processor
Toshiba
32-bit 3-D graphics processor with
lOM-pixels/s Gouraud shading,
40M-pixels/s graphics processor
VITI
PIP chip
Source:
SIS Newsletter
© 1988 Dataquest Incorporated April
Dataquest
April 1988
Scientific Computing Products
Table 5 identifies the major developments in scientific computing products
described by LSI Logic's Stanford Research Laboratory, Rockwell, a Stanford team, and
Texas Instruments.
Table 5
1988 ISSCC Developments in Scientific Computing Chips
Company
Chip Developments
LSI Logic SRL
30-mflops, 32-bit CMOS floating point processor
Rockwell
150-mops GaAs 8-bit slice
Stanford University
Pipeline 64x64 array multiplexer
Texas Instriunents
200-mips GaAs 32-bit microprocessor
Source:
Dataquest
April 1988
Technology Trends
While CMOS processing at 0.8- to 2.0-micron geometries represents the present
"workhorse" technology, advances in silicon emitter-coupled logic (ECL) and GaAs
processing were visible in the presentations. The yield data for GaAs devices at LSI
density levels indicate significant improvements during the last 12 to 18 months.
Dataquest believes that more applications will shift toward GaAs as speed becomes a
larger factor in new applications. Table 6 compares two approaches to DSP using GaAs
hardware developed by Rockwell and Texas Instruments.
© 1988 Dataquest Incorporated April
SIS Newsletter
Table 6
GaAs Microprocessor IC Comparison
Characteristic
Rockwell Design
Company
Texas Instruments/CDC Design
Function
150 mops 8-bit ALU
(1,750 type)
32-bit RISC microprocessor
Architecture
3-bus bit slice (expandable to 16 bits)
6-stage pipeline
Performance
150 mops
200 mips
Operations
Add, subtract, modified
Booth multiply, divide,
bit operations
4 address modes, 16 ALU
and 5 control instructions,
10 memory instructions
Power (Watts)
Low—4.2, high—9.2
Estimated 20 maximum
Die Size—mm
4.9 X 3.9
7.6 X 7.6
Chip Complexity
9,400 transistors,
3,000 diodes
12,872 g a t e s ( t y p i c a l
g a t e i s 1 t r a n s i s t o r and
5 resistors)
I/Os:
Signal + Power/Gnd
64 -«- 29
256 + 70
Process
l.Ou GaAs D-MESFET
1.5u GaAs HBT
Logic Form
Buffered FET logic (BFL)
I^L (modified RTL)
Cell Speed/Power
120ps/1.6mW (register)
160ps/2mW ( g a t e )
Yield Data
18 percent
Experienced 2 p e r c e n t ,
p r o j e c t e d 8 percent
. ' FJJ'^^ 1
Source:
Aztek A s s o c i a t e s
Product Trends
The development of several VSP devices by ISSCC participants is an indication of an
emerging product area for chips oriented toward solving the unique problems incurred in
high-definition video system designs. A similar evolution is occurring in
voice-processing hardware. An estimated 50 other papers described DRAMs and SRAMs,
microprocessors, and other DSP-related functions. These gave evidence of trends
toward continuing increase in SRAM, DRAM, and DSP function densities and speeds with
accompanying decreases in power per bit.
SIS Newsletter
© 1988 Dataquest Incorporated April
Application Trends
Enhancing human interface with systems is a major new thrust in semiconductor
applications. The main focus of this year's ISSCC developments in chip architectures is
toward improving the human-system interface. This was portrayed by the focus of many
papers on voice and graphics communication and processing. HDTV is a major
application addressed by many chip suppliers, followed by optical communications and
massive main memories. Dataquest expects many product announcements in these
applications areas in the near future.
DATAQUEST CONCLUSIONS
The IC world is rapidly evolving toward submicron processing of chips ranging above
1 million transistors in complexity (silicon CMOS), with GaAs now within a factor of 16
in complexity and 4 to 5 times faster in speed. The competing technologies are causing a
proliferation of application-specific DSP designs. Dataquest expects DSP functions to
play an increasingly important role in IC suppliers' product portfolios for at least the
next two to three years. Realignments among IC competitors are expected as these
developments accelerate the obsolescence of many existing products.
Gene Miles
Z
© 1988 Dataquest Incorporated April
SIS Newsletter
•er
r .
•3.qmc. "
r 3"r.-i--j'.;,
n? \t!c;

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