ANALOG INTEGRATED CIRCUITS AND SIGNAL PROCESSING An International Journal Volume 9, No. 2, March 1996 Translinear Circuits in Subthreshold MOS .............. Andreas G. Andreou and Kwabena A. Boahen' 141 PHOTOCOPYING.In the U.S.A.: Thisjournal is registered at the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, U.S.A. Authorization to photocopy items for internal or personal use, or the internal or personal use of specific clients, is granted by Kluwer Academic Publishers for users registered with the Copyright Clearance Center (CCC) Transactional Reporting Service, provided that the base fee of $8.50 per copy per article is paid directly to CCC. For those organizations that have been granted a photocopy license by the CCC, a separate system of payment has been arranged. The fee code for users of the Transactional Reporting Service is 09251030/95/$8.50. Authorization does not extend to other kinds of copying, such as that for general distribution, for advertising or promotional purposes, for creating new collective works, or for resale. In the rest of the world: Permission to photocopy must be obtained from the copyright owner. Please apply to Kluwer Academic Publishers, P.O. Box 17,3300 AA Dordrecht, The Netherlands. Printed on acid-free papel: @ 1996 by Kluwer Academic Publishers. Manufactured in The Netherlands. Analog Integrated Circuits and Signal Processing, 9, 141-166 (1996) @ 1996 Kluwer Academic Publishers, Boston. Manufactured in The Netherlands. Translinear Circuits in Subthreshold MOS ANDREAS G. ANDREOU Electrical and Computer Engineering, Johns Hopkins University, Baltimore MD 21218 USA KWABENA A. BOAHEN Computation and Neural Systems, California Institute of Technology, Pasadena CA 91125 USA Abstract. In this paper we provide an overview of translinear circuit design using MOS transistors operating in subthreshold region. We contrast the bipolar and MOS subthreshold characteristics and extend the translinear principle to the subthreshold MOS ohmic region through a draidsource current decomposition. A fronthack-gate current decomposition is adopted; this facilitates the analysis of translinear loops, including multiple input floating gate MOS transistors. Circuit examples drawn from working systems designed and fabricated in standard digital CMOS oriented process are used as vehicles to illustrate key design considerations, systematic analysis procedures, and limitations imposed by the structure and physics of MOS transistors. Finally, we present the design of an analog VLSI “translinear system” with over$590,000transistors in subthreshold CMOS. This performs phototransduction, amplification, edge enhancement and local gain control at the pixel level. \ 1. Introduction The Translinear principle [ 11 exploits the exponential current-voltage non-linearity in semiconthctor devices and offers a powerful circuit analysis and synthesis [2] framework. Originally formulated for bipolar transistors El], this principle enables the design of analog circuits that perform complex computations in the current-domain including products, quotients, and power terms with fixed exponents [ 11, [2]. Translinear circuits perform these computations without using differential voltage signals and are amenable to devicelevel circuit design methodology. Most of the work on translinear circuits todate, use bipolar transistorsand the emphasis is on high precision and high speed. One fascinating aspect of translinear circuits is their insensitivity to isothermal temperature variations, though the currents in its constitutive elements (the transistors) are exponent$ally dependent on temperature. The effect of small lock1variations in fabrication parameters can also be shown to be temperature independent. An excellent up-to-date overview of translinear current-mode analog circuits using bipolar transistors can be found in [3]. The increased commercial interest in analog CMOS LSI and VLSI has renewed interest in the translinear principle for MOS circuit design. A generalized form of the translinear principle was recently proposed for MOS operating above threshold [4]; this extension however does not follow the original definition of a translinear circuit [11. This extension is simply a design principle that exploits conservation of energy (KVL) around circuit loops which have specific topological properties. A novel class of translinear circuits that employs multiple input gates, with floating gate MOS transistors in subthreshold has been recently proposed and experimentally demonstrated [20]. Another exciting research area that emerged the last few years, is the synthesis of analog VLSI for sensory information processing systems [7], [S] employing MOS transistors operating in subthreshold region [5], [6], [7]. We have been exploring translinear circuits in subthreshold MOS for use in analog neuromorphic LSI and VLSI systems [113, [121, [9], [101. In this biologically motivated computationalparadigm, high processing throughput is attained through a tradeoff between massive parallelism and lower speed in the circuits and therefore subthreshold CMOS operation is possible. Such architectures often necessitate the computation of linear and non-linear functions, and if a current-mode [ 111, [ 121 design methodology is adopted, the translinear principle offers an effective way for synthesizing circuits [ 131, [ 141 and systems [15], [16], [17], [18]. / 142 A. G.Andreou and K. A. Boahen In this paper, we discuss experimental circuit designs based on the translinear properties of subthreshold MOS transistors in the saturation and ohmic regions. Our objective is to present a comprehensive overview on this subject, beginning with the basic devices and circuits, and following it through to the system level. The discussion of subthreshold MOS models, and their characteristicsand limitationscan be found in other excellent references (for example [5], [6]). However, a basic review of subthreshold MOS and bipolar operation, is provided since the large signal properties of the devices are key to the subject matter. Most of the circuit examples given, have been used in analog LSI and VLSI systems that have been fabricated and tested functional. The value of these circuits can only be fully appreciated in the context of the systems that employ them; references to the original journal articles are given. The paper is divided into six sections. Section 2 contrasts the translinear properties of bipolar transistors with those of MOS transistors in subthreshold. Basic circuit techniques that employ MOS transistors in subthreshold saturation and ohmic regimes are introduced in section 3. In the same section, we discuss both translinear loops (TL,) composed of generalized diodes and current sources, and translinear networks (TN) that include voltage sources as well. Section 4, focuses on an analog VLSI translinear system, a contrast-sensitive, silicon retina [171. A discussion of MOS device limitations and deviations from the first order large signal models that ultimately affect circuit and system performance is presented in section 5. Section 6 concludes the paper. 2. Translinear Devices We begin the discussion of translinear circuits in subthreshold MOS technology with the basic devices. A translinear element is a physical device whose transconductance and current through the device are linearly related, that is, the current is exponentially dependent to the controlling voltage. A two terminal p-n junction (diode), with its exponential I-V characteristics, is a translinear element and used often as an example in circuits [3]. Voltage gated, ion channelsconductances- are also translinear devices. Three-terminal devices are termed “translinear” if the relationship between the current and the controlling voltage is exponential and the two terminals across which the controlling. voltage is amlied exhibit true ,‘ diode-like behavior, i.e., increasing the voltage on one terminal is exactly equivalent to decreasing the voltage on the other terminal by the same amount. In this case, a loop of such devices consists of voltage drops across pairs of control terminals and we exploit the linear transconductance-current relationship. Bipolar transistors have both properties whereas MOSFETs do not. The large-signal device model equations for both the bipolar transistor and MOSFET in subthreshold are discussed in Appendix A where the approximations made during their derivations are clearly stated and the symbols are defined. In the active-forward region of operation, the function of a bipolar transistor as a transconductance amplijïer is captured by the following equation: IC = Is e V B -VE (1) VI where V, = ( k T / q ) and is^ is defined in Appendix A. The magnitude of the transconductance from the base is identical to the magnitude of the transconductance from the emitter: gm aIC E -=-- ~VBE %IvE=. - a IC avE IC -- vt I vB=c (2) We now contrast the operation of a bipolar transistor as a translinear element with that of an MOS transistor operating in subthreshold. Much like a bipolar transistor, the MOSFET in subthreshold has exponential voltage current characteristics (see Figure 1). There are however, two fundamental differences between MOSFET and bipolar devices that have implications in the design of translinear circuits. l. Unlike a bipolar transistor, the current in a MOSFET is controlled by the surface potential, which is capacitively-coupledto the gate (front-gate) and bulk (back-gate) terminals. 2. The MOSFET, is symmetric with respect to the source and drain terminals while a bipolar is not. In summary, the MOS transistor is a four terminal device with symmetric drain and source terminals, as result of lossless channel conduction, and an isolated control potential capacitively set by one or more control gates. As we will see in subsequent sections, the latter property of the MOS transistor is a mixed blessing when the design of translinear circuits is considered. Translinear Circuits in Subthreshold MOS 143 Equation 4 can be re-written as a function of dimensionless current quantities iG and i g . Each of these currents would correspond to the device current if the surface potential @S could assume the voltage at the gate or bulk terminal. In essence these currents correspond to ideal diode junctions between the source and surface potential weighted by the appropriate capacitive divider ratio. Therefore, the equation for the drain current can be written as: 0.5 1 VGS, VBE 1.5 [Volts] 2 2.5 Fig. I. Measured current ZDS and Zc versus controlling voltage VGS and VBE respectively. The MOS transistor has dimensions of (16 x 16pm2) and is fabricated in a 1.2pm n-well CMOS process and is biased at a drain-source voltage, V~s=1.5Volts. The current is measured at two different substrate voltage bias conditions. The bipolar transistor is a vertical device with an emitter area of (16 x 16pm2)fabricated in a 2pm n-well CMOS process and biased with Vc~z1.5Volts. T = 301.5 K. It should be pointed out that the voltage difference that controls the current in a MOSFET to yield the translinear behaviour, is the potential difference between the channel surface potential @, and the potential at the source VS and or drain VDso that the current between the drain and source for an NMOS is given by: Since the MOS transistor has two “gates” the relationship between @,(V,, V,) and the bulk or gate terminal voltages V, and VG can be obtained using the simple capacitive divider model depicted in Figure 2. The introduction of the parameter K EZ C,3,/(Cix C:ep) is convenient for modeling the effect of the two gates. Note that K is a function of the surface potential @S as Ciepis a function of the applied gate and substrate voltages. In saturation i.e. when VDS L. 4Vt, and when the current controlling voltages are referenced to source, Equation 3 simplifies to: + In subsequent sections, we will see how the latter formulation facilitates the analysis of MOS translinear circuits and an extension of it will be used to analyze FGMOS translinear loops. Since the dimensioneless current quantities are related to the surface potential the will be called @-currentsor psi-currents. The transconductance from the gate is given by: I gm EZ - a V,, vS=c -- K IDS vt (6) and from the local substrate (backgate) terminal: gmb a I VG,vS=c - (1 - K ) I D S V, (7) The conductance g, at the source is given by: The transconductances depicted in Equations 6,7 and 8 are linear functions of the current -to a first order- and hence each MOS transistor in saturation has the equivalent of three different translinear elements. Note, that the source transconductance is equal to the gate transconductance by shorting the local substrate and the source of the transistor (VB = V S )in which case Equation 4 for the current becomes: In subsequent sections, we will see how shorting of the substrate to the bulk, partially circumventsthe nonidealities in the translinear properties of MOS transistors and enables the design of near ideal loops. The translinear properties of the bipolar and MOS transistors in subthreshold are evident in Figure 1. In 144 A. G. Andreou and K. A. Boahen in the elements connected in the Clockwise (CW) direction is equal to the corresponding product for elements connected in the Counter Clockwise (CCW) direction. As an example let us consider the circuit of Figure 4 consisting of four ideal diodes in the loop X-Z-Y-WX. Following the translinear principle, we can write: G P C ’b E Note that the translinear principle is derived by beginning with Kirchoff’s voltage law or the principle of conservation of energy, so that: i=N/2 C i=l Fig. 2. (a) Symbol for an NMOS transistor. The current between the drain and source is controlled by the difference between the surface potential and the potential at the source terminal. The surface potential is set by the potential at the gate and bulk terminals through the capacitive divider between C,, and c&pshown in (b). An NPN bipolar transistor is shown in (C). The current between the collector and emitter is controlled by the voltage difference between the base and emitter nodes. a logarithmic current scale, the transfer characteristics show linearity with respect to the controlling voltages. Plots of the normalized transconductance ( g m / I ) are shown in Figure 3 and demonstrate how the bipolar device is an ideal translinear element while the MOS transistor in subthreshold only approximates it over a limited range. 3. Translinear Circuits In this section we discuss translinear circuits that employ translinear elements, both MOS operating in subthreshold and bipolar transistors. We follow the convention proposed by Barrie Gilbert in [3] and make a distinction between a Translinear Loop (TL) and a Translinear Network (TN). 3.1. TranslinearLoops In “strictly” TLs the translinear principle [l] can be stated as follows: In a closed loop containing an equal number of oppositely connected translinear elements, the product of the current densities i=N/2 VD(2i-1) - C VD(2i) = 0 (11) i=l Equation 10follows from Equation 11if the voltages are summed around loops of translinear devices. In a circuit graph composed of two terminal elements such as ideal diodes (see Figure 4), there is a direct relationship between the voltage difference among each pair of nodes transversed by the translinear device, and the current in the arc that joins the nodes. This is a consequence of having the voltage nodes that control the current be the same as the current-output nodes of the device. In practical systems, the ideal diodes in Figure 4, would correspond to base-emitter junctions of bipolar transistors with shorted collector-base terminals. Analogous behaviaur can be obtained using translinear three terminal devices such as bipolar transistors, MOS transistors in subthreshold, or any other device that yields diode-like characteristics. However, in three terminal devices, the diode-control nodes in the circuit need not correspond to the current path. In bipolar transistors the diode control nodes are available and thus they can be used explicitely in constraint equations such as Equation l l. This is not true for MOS transistors! As we have seen already, one of the diode control nodes, (namely the node corresponding to the surface potential is not directly accessible. The situation becomes even more complex in MOS transistors with a floating gate (FGMOS) (see Figure 5) coupled to one or more controlling gates (see [ 191, [20] and references therein). At first sight, the floating gate, appears to make the situation worse, but actually it opens the possibility for a new class of translinear circuits Translinear Circuits in Subthreshold MOS 145 ¿ Fig. 4. A Translinear loop using ideal p-n junctions (diodes). VBE WOllSl Fig. 3. Normalized transconductance curves. The transconductance is computed through numerical differentiationof the data in Figure 1, and subsequent smoothing. (Top) For the MOS transistor; (Bottom) for the bipolar transistor. The dramatic decrease of the transconductance in the MOS transistor at low gate-source voltages is attributed to the leakage current. 9 S proposed and experimentally demonstrated recently by Minch et.al. [20]. Essentially the physical structure of FGMOS transistors offer an extra degree of freedom which can be exploited systematically through another set of constraint equations of the form: where V F G i , QFGi are the floating gate voltage and charge on the (ìth) transistor and V G j is the voltage of the ( j th ) control gate. The total capacitance seen in the floating gate is C T i and A i j is a design parameter that depends on the ratio of the control gate to floating gate capacitance,i.e. A i j Cfgj/ C T i . The details of a systematic analysis procedure for FGMOS translinear circuits can be found in [20]. 3.l.l. Analysis of Translinear Circuits with MOS Transistors in Saturation The current mirror is a trivial example of a translinear circuit; it has a single loop with two translinear elements, one CCW and the other cw. Fig. 5. Capacitive model and symbol for a floating gate MOS (FGMOS) transistor. The device depicted in this figure has three control gates G 1, G2 and G3. Two currrent mirrors implemented with complementary devices and connected back-to-back yield the circuit shown in Figure 6. This loop includes four threeterminal devices and corresponds to the ideal diode example of Figure 4. The circuit can be readily recognized as a BiCMOS implementation of an AB stage in a digital oriented CMOS process where only one type (NPN) of bipolar transistors is available [23]. A composite strÚcture made of an MOS in subthreshold and an NPN bipolar yields a pseudo-PNP device with good driving capabilities. Translinear loops using both PNP and NPN bipolar transistors were first studied by Fabre [2 11. Applying the translinear principle to the loop XZ-Y-W-X of Figure 6 yields the following constraint A. G. Andreou and K. A. Boahen 146 In Out I junctions and obtain or The above relationship can also be derived by summing the voltages around the loop (conservation of energy) V1 I Out Fig. 6. The translinear loop of Figure 4, implementedusing composite bipolar and subthreshold MOS transistors. The loop is employed in a current conveyor configuration where the bidrectional output current Zout equals to the bidirectional current l i n . + V2 - V3 - V4 = o Replacing the gate-source voltages for M l , M2, M3, M4 with their respective drain-source currents using Eq. (9) (assuming all devices are in saturation, have VSB = O, have negligible drain conductance, have identical K , have identical IO and geometry S ) , we obtain: or I 1 Fig. 7. A translinear circuit that performs one-quadrant normalized multiplication. I I , 12 and 13 are the inputs and 14 is the output. equation for the currents in the circuit: This classical four junction loop can be combined with two current mirrors to implement a current conveyor [22] where Iout = l i n and Vz = v i n . Our second example is the MOS transistor onequadrant multiply-divide circuit shown in Figure 7. A large number of these CMOS multipliers have been employed in the implementation of a correlation-based motion-sensitive silicon retina [24]. Applying the translinear principle to the loop GNDA-B-C-GND, we find a total of four equivalent diode from which Eq. (14) readily follows. Note that the assumption of identical K holds true to a first order because VSB = O and the gate of all transistors are within a few hundred millivolts from each other. Yet another way of viewing the function of this circuit is that of a log-antizog block. Transistors M1 and M2 do the log-hg, M 4 does the antilog-ing and M3 is a level shifter. Another single quadrant multiplier is shown in Figure 8. This circuit was proposed and its function experimentally demonstrated in [29]. The operation of the circuit can be understood by noting that a single transistor (M4) can perform a single quadrant multiplication because the voltages on the gate and bulk control the current in a multiplicative fashion (see Equation 4). Since in subthreshold the transistors saturate at only a few & of drain source voltage, the bulk terminal of the device can be connected to the drain without turning on the bulk-source junction. An expression for the output current I4 can be obtained by applying the translinear principle around the four loops (Vdd-A-Vdd), (Vdd-B-Vdd),(Vdd-CVdd),(Vdd-D-Vdd) to obtain the following equations Translinear Circuits in Subthreshold MOS I Vbias A I I I I I I M2 A 147 F M5 '2 Fig. 8. A four transistor translinear circuit that performs a onequadrant normalized multiplication and exploits the back-gate in an MOS transistor. Device pairs M l , M 3 and M 2 , M4 share local substrate terminals (in this case n-wells). 1 1 , 12 and 13 are the inputs and 14 is the output. for the psi-currents introduced in Equation 5: The actual currents in the four MOSFETs M1 ,M2,M3,M4, can be written as a function of the psicurrents: z1 = IDSI = SI no iK1 i('-'1) G1 B1 where the devices have been assumed to have the same S and Zno. If now the assumption is made that K I = ~2 = ~3 = ~ 4 Equations , 15 and 16 yield the. following expression for the output current 14 in terms of the input currents 11,12 and 13 : In the original implementation [29] it was suggested that to improve accuracy, the voltage on the local substrate (n-well) of devices M3 and M1 should be set at a value close to that of node B , the local substrate of of devices M2 and M4. This is indeed necessary, as the bulk voltage determines K of all transistors, which Fig. 9. Circuit that converts a bidirectional current on a single wire into two unidirectional currents on separate wires. This is a currentmode absolute value circuit. The sign of the bidirectional input current is assumed to be positive when it adds positive charge to node C . The bidirectional current Z B D is the input, the unidirectional currents 11 and 12 are the outputs, and Z B sets the operating point of the circuit. was assumed to be the same for all transistors. Another implicit assumption here is that the gate voltage is approximately the same for all transistors. Our next example addresses the problem of converting a bidirectional current to two unidirectional currents which is the equivalent to a current-mode half-wave rectification. A translinear circuit that computes this nonlinear function is shown in Fig. 9. The bidirectional current ZBD is steered through transistor M3 when ZBD > O and through transistors M4,5 when ZBD < O. Concentrating on transistors M1,2,3,4, we identify a loop (VDD-A-B-C-VDD) and apply the translinear principle to yield: form a simple current mirror. These equations may be solved for 11 and 12 in terms of ZB and ZBD: M5 Which shows that 11 E ~ Z B DI and 12 2c O when ZBD >> ZB and vice versa. The absolute value is ob- 148 A. G. Andreou and K. A. Boahen and Vdd AI* = IB AI Ii n + similarly for A I* I3 - 14, I B E I3 14.The differential output current A I* is a scaled version of the differential input current A I . The voltage between node B and Vdd should be such that the current source IB stays in saturation. The mirror composed of transistors M5 and M6 converts the unidirectional differential signal AI to the bidirectional signal Iout so that: AI Iout = -IB lin Fig.10. A translinear circuit that computesthe normalized difference of two current signals. II and 12 are the inputs and the bidirectional current Zout is the output normalized to 11 12. + tained by connecting the two output wires together in which case: This circuit has been employed in a CMOS integration of an autoadaptive linear recursive network for the separation of sources [ 141. The next translinear circuit performs a current-ratio computation. This functional block, is part of the readout amplifier in an analog VLSI system that integrates monolithically a one dimensional array of photodiodes and selective polarization film to form a polarization contrast retina [25]. The simple translinear circuit in Figure 10 is excellent for rescaling differential current signals and thus computing the contrast. I1 and 12 represent currents fi-om two selected photodiodes. The heart of the computation circuit will be recognized as a Gilbert gaincell [3] implemented in subthreshold MOS. The analysis of this circuit is typical for translinear circuits that involve differential current signals. Application of the translinear principle around the loop A-B-C-D-A yields: 3.1.2. Analysis of Circuits with MOS Transistors in the Ohmic Regime In this subsection, we extend the translinear principle to subthreshold MOS transistors operating in the ohmic region. In Appendix A, (see Fig. 25), we show how the source-drain current of a MOS transistor can be decomposed into a source component I Q ~and a drain component I Q ~and , that these components superimpose linearly to yield the actual current ISD E Ies - I Q ~ . In the ohmic region, these components are comparable. Decomposition and linear superposition may be used to exploit the intrinsic translinearity of the gatesource and gate-drain “junctions.” This is the basis for extending the translinear principle to the ohmic region. On the otherhand, in the saturation region, we can exploit the translinearity of the gate-source “junction” directly because the drain component is essentially zero and decomposition is of no consequence. The translinear circuits based on subthreshold ohmic operation are only possible because of the symmetry between drain and source operation of an MOS transistor. One could argue that decomposition is also possible with bipolar devices. However, while the difference of two exponentials is the exact form for MOS devices, it is only an approximation for bipolars, due to the fact that the forward and reverse current gains of the device never reach unity [46], [47] (see Equations 45 in Appendix A). This distinction is fundamental and important difference between MOS and bipolar transistors arising from lossless transport in a MOS channel versus lossy transport in bipolars due to recombination in the base. It is possible to use CMOS compatible lateral bipolar transistors as symmetric devices [26] but at the expense of a large base current that increases the power dissipation in the system. Translinear Circuits in Subthreshold MOS Fig. II. A translinear circuit that employs subthreshold MOS transistors in saturation and ohmic regime and computes the product of two input currents 11 and 12 normalized to 11 12. y + To demonstrate the application of the translinear principle to circuits that include MOS transistors in the ohmic regime, consider the one-quadrant currentcorrelator circuit in Fig. 11. Transistor M2 operates in the ohmic region. Proper circuit operation requires that the output voltage is high enough to keep M3 in saturation. This circuit was first introduced by Delbrück [27] and later incorporated in a larger circuit that implements the non-linear Hebbian learning rule in an autoadaptive network [28], [30] and in a micropower autocorrelation system [313. An expression relating the output current, 13, to the input currents, I1 and 14,can be derived by treating the source-gate and the drain-gate “junctions”of the ohmic device as separate translinear elements and applying the translinear principle. For the two loops formed by nodes GND-A-GND and GND-A-B-C-GND in Fig. 11, we obtain 149 GND Fig. 12. A current-mode circuit -translinear network- that implements a normalized cubic non-linearity. 11 is the input, 12 is the output and the voltage source VR normalizes the result. and (20), the output current is given by: I3 Il 14 =Il + 14 3.2. Translinear Networks In the previous section we have discussed “strictly” translinear loops (TL).Translinear networks [3] differ from translinear loops in that they contain independent voltage sources and the following equation can be employed in their analysis: or source-drain current of the MOS transistor can be decomposed into a source component IQ,and a drain component IQd-COntrOkdby their respective “junction” voltages VQ, and Ved. These opposing components superimpose linearly to give the actual current passed by M2, i.e., where E is the independant voltage source and Ç is a constant coefficient that lumps device design and fabrication parameters. The above extension to the translinear principle was proposed by Hart [33]. We begin the discussion of TNs using a simple circuit that has the topology of a current mirror and incorporates a voltage source in the loop (see Figure 12). If the 150 A. G.Andreou and K. A. Boahen 81 T - a a GND e Fig. 13. A translinear circuit using FGMOS transistors to compute the ratio of a cubic to square functions. 11 and 12 are the inputs and 13 is the output. input current is I I , the output current is 12, and VR is a constant voltage source, application of the translinear principle around the loop (GND-A-B-C-D-GND) yields: The voltage source is necessary for circuit operation and it normalizes appropriate the output current. This circuit has been employed in a small system that implementsthe Herault-Jutten independent component analyzer [141. An FGMOS-based circuit that has the same functionality as the circuit in Figure 12, is shown in Figure 13. We begin the analysis of the circuit by noting that the current in the channel of an FGMOS, is controlled by multiple gates that can be thought as extensions to the front gate of the transistor. As such, Equation 5 can be re-written for an N-input NMOS transistor as: where it has been assumed that all N gates of the i-th transistor have the same strength, i.e. have the same coupling capacitance to the floating gate. The charge Q F G on the floating gate is incorporated through a geometry related multiplicative constant SQ so that when the charge Q F Gis zero, S Q = l. An expression for the output current I3 can be obtained by applying the translinear principle around three loops that include the floating gates and source of transistor M 3 together with the floating gate nodes and sources of the other transistors. When the three loops are traversed, the following equations for psi-currents are obtained: i3a = i2a i3a = ìla i3b = ilb (25) We have adopted a notation where for example, i3a denotes the psi-current in device 3 controlled by the voltage on its gate a. Using Equation 24, the current at the source of M1 , M 2 , M 3 , can be expressed as functions of psi-currents so that: where the devices are assumed to have the same S and I n o ; the back-gate contribution to the current in each device is eliminated as all transistors have the source shorted to the substrate. Now, by making the assumption that K I = ~2 = ~3 = ~ 4 and , no charge on the floating gates, Equations 25 and 26 yield the following expression for the output current I3 in terms of the input currents I I and 12: The assumption of equal K is reasonable so long the voltage on the floating gate is such that all devices stay in subthreshold. An alternative way of obtaining a functional description of the circuit can be found in the paper by Minch et. al [20]. A TN that incorporates bipolar transistors, an MOS in subthreshold, and an independent voltage source Vxu is shown in Figure 14. We will use Equation 22 to derive the relationship between I l , 1 2 , I z w and Vxy. The MOS transistor is assumed to be ideal with K = l. A discussion of networks with non-ideal devices will be done in the next section. A ratio relationship between I Z w and I1 - 12 can be derived by employing Equation 22 applied to the two loops (X-Z-Y-X) and (X-Y-W-X) to yield the following equations: Translinear Circuits in Subthreshold MOS Fig. 14. A BiCMOS translinear network that exploits the MOS subthreshold ohmic characteristics. Using the adopted conventions for current decomposition in NMOS transistors (see Figure 25), I Qand ~ IQd together with Equation 28 we obtain the following expression that relates the currents in the circuit. This is a lumped parameter model where G1 and G2 correspond to resistances per unit length. The voltages on nodes P and Q referenced to ground, represent the state of the network and can be read out using a differential amplifier with the negative input grounded. The equivalent circuit using idealized non-linear conductances is shown in Figure 15(right). The difference in currents through the diodes D1 and D2 are linearly related to the current through the dimsor MOS transistor.' This relationship can be derived from Equation 55 describing subthreshold conduction, and the ideal diode characteristics where ID= Is exp[VD/ Vt]. An expression can be derived for the current IPQin terms of the currents I p and I Q , the reference voltage V, and the bias voltage V,, when diodes are replaced by transistors: The current In0 and S is the zero intercept current and geometry factor respectively for the diffusor transistor M h . Is is the reverse saturation current for the diode that is assumed to be ideal. The currents in these circuits are identical if C1 = G2 It is immediately apparent that the current ratio Izw/(I1 - 12) can be controlled both by a fixed parameter (S)that is designed prior to the fabricaition of the circuit and a variable quantity (e-vxy/K)that can be programmed (post-fabrication) during circuithystem operation. This property will be utilized in the design of linear MOS transistor-only spatial averaging networks. 3.2.l. Translinear Spatial Averaging Networks Often, models of neural computation necessitate the realization of spatial averaging networks [7]. To demonstrate the analogies between linear and translinear networks as well as their subtle and important differences, we begin with networks that employ linear conductances, voltages and currents and contrast them with translinear current-mode [ 161 networks. A voltage-mode circuit model for a loaded network is shown in Figure 15(left) for which: 151 (F) "1 exp Increasing VC or reducing V, has the same effect as increasing G1 or reducing G2. The state of this network is represented by the charge at the nodes P and Q . Since the anode of a diode is the reference level (zero negative charge), the currents I p and IQrepresent the result. Unfortunately, the anode of a diode or a diode connected transistor is not a good current source. When diodes are not explicitly available in the process, diode connected PMOS or NMOS transistors can be used as shown in Figure 16. When the loads are PMOS, the current current I ~ isQgiven in terms of voltages normalized to ( k T / q ) : When NMOS transistors are used as loads, there is the additional benefit, that of exploitingthe current conveying properties of a single transistor [ 161, to obtain the current outputs I p and IQ, on nodes that are low A. G. Andreou and K. A. Boahen 152 Fig. 15. Building blocks for linear loaded networks. Using segments that employ ideal (left) linear and (right) non-linear elements. P P where s h and Sv are geometry parameters for transistors M h and M,, respectively. The one dimensional MOS transistor-only network corresponding to the Helmholtz equation shown in Figure 17 can model the averaging that occurs at the horizontal cells layer of the outer retina. This is equation is the basis of the well known silicon retina architecture proposed by Mahowald and Mead [34], [7]. Summing the currents at node j we get: Using the results from the previous section for the currents Iij and Ijk given by Equation 32 substituted in Equation 33 yield: Vr I I Fig. 16. Current-mode building blocks for linear loaded networks using (top) PMOS transistor implementation, (bottom)NMOS single transistor current-conveyor implementation. conductance (the drain terminal are now excellent outputs for the currents). Using Equations 8.45 in [1619 the current I P Q is given as: Normalizing internode distances to unity the above equation can be written on the continuum as: I*@) =I(x) d 2I (x) +h dx2 This equation yields the solution to the following optimization problem: Find the smooth function I (x) that best fits the data I * @ ) with the minimum energy in its first derivative. Input is the currents I * @ ) and output the currents I (x). The parameter h = exp& vc - K ~ V is~ the ) cost associated with the derivadve energy-relative to the squared-error of the fit. The diffusive network in Fig. 17 was recently described in terms of “pseudo-conductances” [35]. We (2) Translinear Circuits in Subthreshold MOS 153 Fig. 17. A one-dimensional MOS translinear network to perform local aggregation -spatial averaging-. The back-gate terminals of all devices are connected to the substrate. have used the chargelcurrent-based formulation first proposed in [ 151 to explain its behaviour. This &mentmode approach relies an intuitive understanding of the device physics and yielded the insight which enabled us to extend the translinear principle to subthreshold MOS transistors in the ohmic region as well as the decomposition of the current into dimensioneless components correspondingto ideal junction. We now have a comprehensive current-mode approach for ailalyzing subthreshold MOS circuits. The essence of this approach is the representation of variables and parameters by charge, current, and diffusivity. Voltages and conductances are not used explicitly. Bult and Geelen proposed an identical network for linear current division above-threshold and used it in a digitally-controlledattenuator [36]; they also analyzed its subthreshold behavior. However, they stipulate that all gate voltages must be identical and control the division by manipulating the geometrical factor W / L of the devices. We have shown here, and previously in [ 151, that this constraintmay be relaxed in subthreshold without disrupting linear operation. This is a real bonus because it allows us to modify the divider ratio or space constant of the network after the chip is fabricated by varying (V, - Vr). Tartagni et al. have demonstrated a current-modecentroid network [37] using subthreshold MOS devices whose operation is described by the current division principle. 3.3. A general resultfor MOS translinear loops Three of the circuits discussed in the previous subsection, namely the translinear multiplier of Figure 8, the MOS implementation of the Gilbert gain stage in Figure 10, and the current correlator (Fig. 11) have been experimentally shown to exhibit near “exact” translin- Fig. 18. Translinear loop composed of five MOS transistors in subthreshold. All devices are in saturation except device M5 which is in the ohmic regime and therefore can be decomposed as two devices in saturation ,back to back sharing same gate and substrate. ear behaviour even though they are build from MOS transistors and they do not have their source connected to the local substrate. A recent result by Eric Vittoz [32] can be employed to partially explain this rather surprising behaviour. He considers translinear loops constructed from MOS transistors in subthreshold saturation with common substrate connection (similar to the one shown in Figure 18). If the pairing of transistors in the CW and CCW direction, is such that they have their gates connected to gates and sources connected to sources and they are alternated (much like even-numbered and oddnumbered devices in Figure 18) Vittoz shows that the translinear loop does not suffer from the MOS transistor non-ideal translinear behaviour. He notes also that loops containing transistors in the ohmic regime can also be included in this formulation as they can be decomposed as two parallel connected saturated de- 154 A. G. Andreou and K. A. Boahen vices sharing common gate and comrnon substrate (see Figure 25.) However, to account for the near “exact” operation of the multiplier in Figure 8, Vittoz’s argument must be extended to include loops that go through the back gate of the MOS transistors as illustrated in Figure 18. The global substrate restriction can thus be removed and replaced by a local substrate connection, and the result still holds true. In a standard CMOS process, this will of course be possible only for one type of devices. Now, we will re-examinethe operation of the circuits in Figure 10, Figure 8, and Figure 1l . Consider the largest loop (A-B-C-D-A) in Figure 10. Devices M1 and M2 have common gate and common bulk and so do devices M3 and M+ When adjacent devices are paired in different ways we observe that M3 and M2 share the same source and bulk which is the case also for M1 and M4. In the the largest loop (A-B-C-D-A) of translinear multiplier circuit of Figure 8 in we can verify that transistors M1 and M2 as well as M3 and M 4 share common gate and source. The alternative pairing, finds M 4 and M2 sharing same bulk and source which is also the case for M3 and M l . When devices in the loop are operating in the ohmic regime, such as M2 in the circuit of Figure 11, we can verify that the loop (GND-B-C-GND) incorporates two adjacent sets of devices M3 and M2 share same bulk and source/drain while M3 and M 4 share bulk and gate; the bulk in this circuit is the same for all devices. 3.4. Translinear circuit dynamics The dynamics of translinear circuits and systems have not been discussed in this paper. However, it was pointed out in [45], that in networks with non-linear conductances without complementary non-linear reactances, the state equations that describe the dynamics of the system are non-linear. Given an architecture and a particular network, a method was outlined to test for stability [45]. 4. A TranslinearSystem: A ContrastSensitiveSilicon Retina Image acquisition under naturally occuring, uncontrolled lighting conditions is required by autonomous robotis, in prosthetic devices for the blind, and autonomous motor-vehicle navigation. Today this task is accomplished in two separate steps. First the light intensity is recorded through a standard imager such as a CCD camera. The intensity field is subsequently processed outside the camera to discard any absoluteluminance information and form a representation where only relative illumination, i.e. contrast, is retained. Additional processing such as edge extraction and or low bit-rate encoding may follow. However, even though the precision necessary for these tasks rarely exceeds 8 bits, the signal itself has a very large dynamic range, many orders of magnitude, which makes the problem difficult. This issue becomes acute when the illumination varies within a single frame something not uncommon in natural scenes (see Figure 19). The detrimental effects of non-uniform illumination in the performance of a face recognition system have been investigated experimentally by Buhman, Lades and Eeckman [38]. We will now present one solution to the problem of robust image acquisition and preprocessing under variable illumination conditions: a neurornorphic analog VLSI silicon retina. This is a contrast-sensitive edge-enhancing imager that includes a rudimentary, yet effective, local gain control mechanism at the pixel level. The architecture is inspired by the processing performed in the outer plexiform layer of the vertebrate retina [ 151, [171. The resulting image captured with such a system is shown in Figure 20. The biologically motivated solution is attractive from a computational perspectiye because contrast, an invariant representation of the visual world, has been obtained with a front-end that is robust, small, and extremely low power (a few mW). There is also another benefit; the output representation has limited range and therefore subsequent processing/communication stages are not burdened with handling and processing signals of wide dynamic range. A performance comparison between the contrast sensitive silicon retina front end [15] and a conventional camera, in a face recognition experiment is reported in [38]. 4.1. Biological Organization The analog silicon system in the core of the array is modeled after neurocircuitry in the distal part of the vertebrate retina-called the outer-plexiform layer. Figure 21 illustrates interactions between cells in this layer [39]. The well-known center/surround receptive field emerges from this simple structure, consisting of just two types of neurons. Unlike the ganglion cells Translinear Circuits in Subthreshold MOS : : 155 : ’ 120 150 180 210 Fig. 19. (Bottom) “Mark” & captured by a conventional camera. (Top) Intensity profile at image line 110 (white line). The light souIce is positioned to the right side of the image and it introduces a large gradient in illumination within a single frame. This is clearly shown in the intensity histogram. The dynamic range of the scene exceeds the dynamic range of the camera. Aperture control on the camera provides a rudimentary global gain control mechanism. Information in this imagé is lost at this very first step because there is no gain control (adaptation) at the pixel level. Fig. 20. (Bottom) “Mark” as captured by the translinear silicon retina. (Top) Histogram for the output of the system. The light source is again positioned to the right side of the image and it introduces a large gradient in illumination within a single frame. The image captured by the silicon retina discards absolute illumination and preserves only local contrast information through local gain control at the pixel level. Unlike the image in Figure 19, the presence of a large illumination gradient does not degrade image acquisition here. in the inner retina and the majority of neurons in the nervous system, the neurons that we model here have graded responses (they do not spike); thus this system is well-suited to analog VLSI. The photoreceptors are activated by light; they pro- duce activity in the horizontal cells through excitatory chemical synapses. The horizontal cells, in turn, suppress the activity of the receptors through inhibitory chemical synapses.. The receptors and horizontal cells are electrically coupled to their neighbors by electri- 156 A. G. Andreou and K. A. Boahen Photo-receptors Electrical Synapses f’ HorizontalCells Fig. 21. One-dimensional model of neurons and synapses in the outer-plexiform layer. Based on the red-cone system in the turtle retina. cal synapses. These allow ionic currents to flow from one cell to another, and are characterized by a certain conductance per unit area. In the biological system, contrast sensitivity -the normalized output that is proportional to a local measure of contrast- is obtained by shunting inhibitiori. The horizontal cells compute the local average intensity and modulate a conductancein the cone membrane proportionately. Since the current supplied by the cone outer-segment is divided by this conductance to produce the membrane voltage, the cone’s response, ,will be proportional to the ratio between its photoinpùt and the local average, i. e. to contrast. This is a very simplified abstraction of the complex ion-channel dynamics involved. The advantage of performing this complex operation at the focal plane is that the dynamic range is extended (local automatic gain control). The biological system, is mapped onto silicon using circuits of minimal complexity that exploit native properties of subthreshold MOS transistors. High computational throughput at low levels of energy dissipation is achieved by employing low precision analog processing in a massively parallel analog architecture that exploits the translinearproperties of subthreshold MOS in saturation and ohmic regime. 4.2. Silicon System Architecture The core of the silicon retina is an array of pixels with a six-neighbourconnectivity (see Figure 22). The wiring is included in the layout of the cell (see Figure 24) so that they may be tiled in a hexagonal tesselation to form the focal plane processor. This is a mesh processor architecture where two layers of processors, C and H , communicate both intra and inter layer through local paths. This parallel processing scheme features locality of reference and thus minimizes communication costs. I X I Fig. 22. Floorplan and system organization. It comprises of two functional components, the core, and the support circuitry. Focal plane processing is performed in the core area. Fig. 23. One-dimensionalimplementationof outer-plexiformretinal processing. There are two diffusive networks implemented by transistors M4 and M5, which model electrical synapses. These are coupled together by controlled current-sources(devices M1 and M2) that model chemical synapses. Nodes H in the upper layer correspond to horizontal cells while those in the lower layer ( C ) correspond to cones. The bipolar phototransistor QI models the outer segment of the cone and M3 models a leak in the horizontal cell membrane. Note that the actual system has a six neighbor connectivity. Support circuitry in the periphery extracts the data from the core and interfaces with the display. The chip incorporates a video pre-amplifier and some digital logic for scanning the processed images out of the array. This circuitry is discussed in detail in the paper by Mead and Delbrück [41]. Standard NTSC video is produced off-chip using an FPGA controller and a video amplifier. The basic analog MOS circuitry for a one dimensional pixel with two neighbor connectivity is shown in Figure 23. We begin with the non-linear aspects of system operation, its contrast sensitivity. The nonlinear operation that leads to a local gain-control mechanism in the silicon system is acheived through a mechanism that is qualitatively similar to the biological counterpart, but quantitatively different (see discus- Translinear Circuits in Subthreshold MOS 157 0 Fig. 24. (Left) Photomicrograph of the chip. The surface is covered by second metal except where there are openings for the phototransistors (the dark square areas). Note the hexagonal connectivity of the pixels. (Right) Layout df the basic cell. sion in [ 151). Refering to Figure 23, the output current I c ( x m , y,) at each pixel, can be given (approximately) in terms of the input photocurrent I (x, y,) and a local average of this photocurrent in a pixel neighborhood ( M , N ) . This region may extend beyond the nearest neighbor. The fixed current I , supplied by transistor M3 normalizes the result and KP is a parameter. At any particular intensity level, the outer-plexiform behaves like a linear system that realizes a powerful second-order regularization algorithm [40] for edge detection. This can be seen by performing an analysis of the circuit about a fixed operating point. To simplify the equations we first assume that g = ( I h ) g , where ( I h ) is the local average. Now we treat the diffusors (devices M4) between nodes C and C’ as if they had a fixed diffusitivity jj. The diffusitivity of the devices M5 between nodes H and H’ in the horizontal network is denoted by h. Then the simplified equations describing the full two-dimensional circuit on a square grid are: i=mfl j=nfl Ic(xm, Y,> = ~u + h x { I h ( X m , y,> - ~ h ( x i y, j ) } i=mfl j=nfl Using the second-difference approximation for the laplacian, we obtain the continuous versions of these equations M x , y) = I(& y) + jjV21,(x,Y ) (36) with the internode distance normalized to unity. Solving for Zh (x, y), we find This is the biharmonic equation used in computer vision to find an optimally smooth interpolating function I h (x, y) for the noisy, spatially sampled data I (xi, y j ) ; it yields the function with minimum energy in its second derivative [40]. The coefficient h = jjh is called the regularizing parameter; it determines the trade-off between smoothing and fitting the data. 4.3. Layout Considerations The two-layer architecture for the silicon retina can be accomodated in a cell area of 80h x 94h using a single poly two metal technology. In the implementation reported in [15] and here, a double poly, double metal technology is used and the cell area is 66h x 73h. First metal and polysilicon wires are used for interconnects; second metal is used to cover the entire array, shielding the substrate fi-om undesirable photogenerated carriers. Transistors are implemented using both polysilicon layers. The system has been fabricated with 230 x 210 pixels on a 9.5 x 9.3 mm die in a 1.2pm n-well double metal, double poly, digital oriented CMOS technology. The chip incorporates 590,000 transistors in the 48,000 pixels and support circuitry, with the core operating in subthreshold/transition region consuming less than 1OOmW. 158 A. G. Andreou and K. A. Boahen : S+ 4 B' 'Qs D 'Qd I I ) 'SD Fig. 25. Large signal model for an NMOS transistor (left). Adopted conventions for current decomposition in PMOS and NMOS devices (right). < 4.4. Discussion The silicon retina, presented in this paper is essentially an analog floating-point processor. As a first step, the system computes the range (the voltages on the H nodes correspond to the value of the exponent in floatingpoint data representation). This is the operating point of the system and is a function of the spatial coordinates and this is how local automatic gain control is achieved. At an operating point, sophisticated spatial filtering is performed to smooth the sampled data and enhance the edges. Having separated the problem of precision and dynamic range, the signal processing within the range can be done with low precision analog hardware. The benefit of a robust architecture on the ultimate system performance is evident in the design of the silicon retina. The regularization properties inherent in the architecturemitigate the intrinsic random variations in the device characteristics, leading to robust performance. This methodology allows analog VLSI implementations using the poorly matched, small geometry, nano-power devices available in garden-variety digital VLSI CMOS technology. Thus we see how a neuromorphic architecture can account for the properties of the computational substrate, and yield robust operation in the presence of noise (structural variability) in the individual devices, The translinear propeties of the MOS transistor in subthreshold ohmic and saturation are key to an area efficient implementation which is commensurate with integration at the focal plane. 1,7 - I !I I !I I 13 +................i...................1...................t...................I................... ................L1 1 5 '1 1 1 O I I I I I 0.5 1 1.5 2 2.5 l/sqrt(Vs,) 3 ( V-"2 ) Fig. 26. Experimentally determined variation of parameter function of the substrate voltage V S B . K as a tr bslinear circuit design. This was also shown to be gi applicable to MOS transistors in subthreshold as well. The absence of a base current makes it in a sense an "ideal" element. However, one must be aware of certain characteristics of the MOS transistor that have a detrimental effect on circuit behaviour. Some of them are already depicted in the large signal model equations. Device characteristics that are not modeled in the simplified transistor model are discussed with the help of experimental data from measurements on transistors operating in subthreshold. 5. Subthreshold MOS Device Limitations 5.1. Transconductance Limitations As we have seen earlier, linear relationships between conductance/transconductances and current, and the equivalence of controlling the current from the collector and emitter terminals, are the basis of bipolar Unlike bipolar transistors, the gate and the source of an MOS transistor are not equally effective in controlling the current. Increasing their voltages by the same amount, does not change the current by the same Translinear Circuits in Subthreshold MOS amount because the transconductances are different (compare Eqs. (6) and (8)). Whereas changing the source voltage changes the the barrier by an equal amount, only a certain fraction ( K ) of changes in the gate voltage affects the surface potential and hence the barrier height. This behaviour is depicted in Figure 1 and manifests itself as a slope for the current characteristics that has lower value compared to that of a bipolar transistor. An even lower value (1 - K)for the slope is seen from the back gate terminal. The experimental data from Figure 1 suggest that K can be pushed closer to unity by biasing the device with a large surface potential (i.e. large VSBand VGB).The dependence of K on the substrate voltage is measured and plotted in Figure 26. As the substrate is reversed bias, the depletion capacitance C&,p, in Equation 58 becomes smaller and hence the gate has larger influence on the channel conductance. Under these conditions, the subthreshold MOS device becomes closer to an ideal translinear element whose transfer characteristics come closer to those of bipolars. Another way to circumvent the potential divider problem is to move the local substrate voltage together with the source voltage-which is exactly equivalent to increasing VGB by the same amount and does not change VSB.In practice, this may be achieved simply by shorting the source to the local substrate. Clearly, this is only possible with devices in separate wells. For this reason, it is preferable to bias the device at large values of surface potential as described in the previous paragraph, since that works for both types of transistors and is more area efficient (there is no need for separate wells and a triple well process). Another departure from the simplified model equations presented earlier is evident when we consider the dependence of the drain,current on the voltage at the source of the transistors. According to the model (equation 55) this dependence should follow the exponential law and the parameter in the exponent is the inverse of the thermal voltage. An experiment to verify this was conducted and the results are shown in Figure 27 where the slope of the curve has the value of only 35.4 V-’ . This suggests that the source conductance cannot be adequately described though Equation 8 and that parameter r] m Os9 must be introduced in of the source and drain voltages in the original device Equation 55. Both NMOS and PMOS transistors in different fabrication processes show similar behaviour that is correlated to the zero bias leakage current of the junctions. 159 107 10.8 10-9 10 1 O I I I -0.05 I I I I O 0.05 0.1 0.15 i 0.2 0.25 VSB (Volts) Fig. 27. Measured drain current IDS as a function of the source voltage VSBfor a (16 x 16pm)NMOS transistor fabricated in a 2pm n-well CMOS process. The solid line is an exponential function fit to the data. The experimentswere performed at a temperature such that the thermal voltage V, = 0.0259 Volts and hence 1/ V, = 38.5V-’. (;C 10 -’ 1o -0 n S ~ -O 10 -I0 10 -” O 0.2 0.4 0.6 0.8 1 1‘( Fig. 28. Drain saturation characteristics of an NMOS FGMOS transistor. The dimensions of the device are W = 8pm and L = 4pm and the first-poly second-poly coupling area is 2pm x 2pm. Note that the current is plotted in a logarithmic scale. 160 A. G. Andreou and K. A. Boahen 1.1 ou 30 , , --Iu 25 ' . * I I Above Threshold 20 15 Subthreshold -1 n 10 --1 oop 5 --IOP O 1 O 5 10 15 20 25 0.2 30 0.4 0.6 0.8 1 1.2 VGS 1.4 [V] Fig. 29. (Left) Density plots of currents in a 32 x 32 array of 4 p m x 4 p m NMOS transistors. Each transistor is represented by a square pixel. Current level is coded by the shade of gray, where the minimum and maximum values are represented by black and white, respectively. The current at a nominal current level of lOOnA is obtained by setting VGSto be the same for all transistors in the array. The devices are biased in saturation. (Right) Measured drain current ZD versus gate-source voltage VGSfor 32 small geometry transistors (4 x 4p.m) fabricated in a 2 p m n-well CMOS process; drain-source voltage of V~s=1.5Volts. The fuzziness in the current, (mismatch between ¿levices), is constant in subthreshold (on a log(Z) scale) and decreases as the device ent!qs the transition and above threshold regime. \ , 5.2. Output Conductance Limitations The non-zero output conductance, or what is often called the Early effect also degrades the performance of the circuits. All well known circuit techniques that reduce the output conductance, such as cascoding and regulated cascoding are beneficial to translinear circuits. A second effect contributes to the drain conductance of FGMOS transistors. As pointed out in [19], the parasitic coupling of the drain and source to the floating gate through capacitances Cfgd and Cfgs(see Figure 5) yields an exponential dependance of the output current on the drain voltage. This is evident in the experimental data shown in Figure 28 for NMOS devices; PMOS transistors exhibit similar behaviour. The experimental data in Figure 28 are fit to an output conductance model and the overlap capacitances Cfgd and Cfgsestimated as 0.15fFlpm [19]; this number could be as large as O.SfF/pm. 5.3. Device Matching Limitations Another important area of concern in designing translinear computational circuits, is the poor matching characteristics of MOS transistors in subthreshold (much worse than bipolars). This is more acute in analog VLSI systems applications, where small geometry transistors must be used, typically 4pm x 4pm or 6pm x 6pm, (in a one micron process) to achieve high densities. Low power design condiserations suggest that it is preferable to operate the devices in the region where the transconductance per unit current is maxi- mum [16], i.e., in the subthreshold and transition regions. Small device geometries and high transconductance per unit current makes the drain current strongly dependent on spatial variations of process-dependent parameters, particularly l o . Characterization of the fabrication process and the qhtching properties of the basic devices is thus of paramount importance because it provides the necessary information for designing low power systems. The experimental data in Figure 29 show that there are three different effects that can be responsible for the poor matching characteristics of MOS transistors in subthreshold; these were discussed in [42]. After discounting the two deterministic effects, we are left with the random variations. Random mismatch in the subthreshold region can be characterized in terms of the simple model parameters Io and K . The parameter K is very stable, with a normalized standard deviation of o ( K ) / ( K ) M 0.3%, where (.) denotes mean value. The small variations in K suggest that doping and gate-oxide thickness are extremely uniform. The fuzziness in the data points in Figure 29(Right) is therefore not due to changes in slope. This implies that characteristics are displaced from one transistor to the next, implicating the flatband voltage which depends on the contact potential, and this fixed interface charge density Qg (implanted ions and trapped electrons). The variations in the current can of course be related to this fixed-charge distribution through the transconductance and the gate-oxide capacitance. (39) Translinear Circuits in Subthreshold MOS Fig. 30. Dependence of normalized standard deviation of IDS on transistor size; the lines are best fits to the data. All devices have square geometries with area A = L2. Notice that the normalized standard deviation of IDS saturates at large transistor geometries. Figure 30 shows the dependence of the normalized standard deviation of the drain current, O ( I ) / (I ) , on transistor size. Each data point represents measurements from approximately 1000 transistors. The normalized standard deviation of the current is inversely proportional to the square root of the device area, A, and is given by: where 00 is the mismatch per unit length for a given device type and process. 5.4. Noise Limitations Shot noise in the MOS transistor operating in subthreshold has a one-sided power spectrum given in [49]p (41) Si,shot (LC)) = 4qIDS For a device in saturation the noise is exactly half. For sub-threshold currents between 1 nA and 100 nA, flicker noise for mid-to-low frequencies must be included. A model for flicker noise is given by [49]. Mg; si,f l i c k ( w ) = C&WL 2n LC) (42) M a process-dependent constant with a typical value of 4.0 x 1026C2/rn2. 161 Fig. 31. Power spectral density for a PMOS transistor in saturation, where W = 1148 pm and L = 4 pum. The model is given by solid lines, the data are marked by x's. The three curves correspond to nominal current values of (a) 1 nA, (b) 10 nA, and (C)100 nA for an equivalent square device. Some amount of excess noise is evident at low current levels. Assuming shot and flicker noise are independent, a complete noise model for a transistor operating in the subthreshold region is Figure 31 shows the noise power spectral density for a PMOS transistor. One free process-dependent paramciter M is used to model the flicker noise. Note that, at low enough current levels, flicker noise cannot be detected within the audio frequency range. This property is seen for curve (a) of Figure 31 in which there is little evidence of flicker noise for frequencies above 50 Hz. 5.5. Bandwidth Limitations The maximum useful frequency of operation possible with an MOS transistor, is determined by its transition frequency f~ defined [49] as (gm/ 2 n C ) where C is the total input capacitance i.e. the capacitance per unit area times the area of the device: so that for subthreshold operation: 162 A. G. Andreou and K. A. Boahen The maximum value of drain current IDSmax with the MOS transistor still in subthreshold region is given by 151: a sabbatical leave at Caltech. We thank Carver Mead for his continuing support and encouragement. Chip fabrication was provided by MOSIS. r r Appendix A: Device Models From the above equations a maximum transition frequency in subthreshold f T m a x can be approximated to: (44) where is the effective carrier mobility and L iy$he device channel length. The transition frequency ‘of a device is essentially the bandwidth (as determined by the internal gain and parasitic capacitances of the transistor). For six to ten micron length devices (typical in analog VLSI today), functional systems in the hundreds of kHz range are possible while for submicron devices, the limit extends to the MHz range. 6. Conclusions In this paper we have provided a comprehensive overview of the application of the translinear principle to MOS circuits operating in subthreshold. Our research was aimed at exploring different ideas on neuromorphic analog network computations and their VLSI implementations. The results of our investigation are encouraging; analog circuits designed with components of limited precision, when assembled in large networks following a design methodology, based on translinear circuit techniques, can successfully perform linear and non-linear computation with energetic eficiency unmatched by any other digital counterparts. Our 590,000 transistor analog VLSI, contrast sensitive, silicon retina is another step towards the direction envisioned by Barry Gilbert: [44] “. . .convergence of IC technology capabilities and neural network requirements makes wafer-scale integration of meganetworks a very real possibility . . .”. Acknowledgements The research was partially supported by NSF grant ECS-9313934; Paul Werbos is the program monitor, and by a contract from the Army Night Vision Laboratory at Fort Belvoir. The final version of this document was prepared while one of the authors (AGA) was on A*[email protected] The Ebers-Moll model [46], [47] for an npn bipolar transistor is: IE = -IF +CYRIR (45) IC = ~ F I F IR , I F = I E s (q eV B7E - 1) QVBC IR = I c s ( e 7 - 1) FIES = ~ R I C S (47) where IC and I E are the collector and emitter currents respectively and V B E is the base to emitter voltage, VBCis the base to collector voltage, IES is the saturation current of emitter junction with zero collector current, ICS is the saturation current of collector junction with zero emitter current, a~ common-base current gain. aR common-base current gain in inverted mode, i.e. with the collector functioning as an emitter and the emitter functioning as a collector. By convention, the currents for bipolars are positive when flowing into its terminals. Combining Eqs. (49, (46), and (47), the collector current can be expressed as: (e* - 1) - - ( e qvBc y - 1)) UR (48) For an ideal device with common-base current gain, C Y F , and common-base current gain in inverted mode, CXR,very close to unity, the above equation becomes: IC = I E s (qvBE e 7 - e 4vBC r ) (49) However, regular bipolar transistors do not have both near unity. CXF and U R Translinear Circuits in Subthreshold MOS When the collector to base voltage equals zero or the collector is reverse biased with respect to the base, the above equation simplifies to the familiar: where A E is a design parameter, the area of the emitter junction. JES and Is are the saturation current density and current for the emitter respectively. In this case, IR << I F and the equations above give IC = - a I~E . Using the relation I E IC IB = O (KCL) we get the familiar result + + where ß F is the common-emitter current gain. A.2 MOS Transistor Model A charge-based formulation [7], [ 161 that preserves the symmetry between the source/drain terminals of an MOS transistor is presented. Which terminal of the device actually serves as the source or the drain is determined by the circuit, the bias conditions-and even the input signals. This symmetric view of an MOS transistor enabled us to extend the translinear principle to operation in the subthreshold ohmic regime [ 161. The MOS device has a very simple current-charge relationship because diffusion and drift are both proportional to the concentration gradient. As shown in Appendix A of [7] and in [ 161, this yields a quadratic expression for the current that consists of two independent opposing components I Q and ~ Zed-in the absence of velocity saturation and channel-length modulation effects. These components are related to charge densities at the source Q: and QZ at the drain of the device. The device drain-source current can thus be written as: I E les - I 163 tances Cix and C:ep are the gate oxide and depletion area capacitances of the channel. A key property of the MOS device that makes this possible is lossless channel conduction. Unlike a bipolar transistor, the controlling charge on the gate is isolated from the charge in transport by the almost infinite gate-oxide resistance. Therefore, there is no recombination between the current-carrying charge in the channel and the current-modulating charge on the gate. The familiar ohmic/saturation dichotomy introduced in voltage-mode design can be reformulated in terms of the opposing drain and source driven current components. In saturation, [ZedI << II Q I~and I M I Q ~ and therefore the current is independent of the drain ~ Zed and voltage. In ohmic, Zed N IQs and I = I Q therefore the current depends on the drain voltage as well as the source and gate voltages. The functional dependence of the current components on the terminal voltage is fixed and remains the same throughout the ohmic and saturation regions. The charge densities at the source and drain terminals can be related to the terminal voltages. In general the charge-voltage relationship is much more complicated than the current-charge one because both the mobile charge and the depletion charge are involved in the electrostatics. The device current in Equation 52 can thus be written as a function F,of the terminal voltages with a general functional form for the current-voltage relationship valid for all the regions of operation given by: This functional form was first introduced by [48] for above threshold operation and is also discussed in [49]. q$r an n-type device, F is a nonpositive, monotonically decreasing function of VGBand a monotonically increasing function of VSB. In subthreshold region, the following factorization of F : is also possible [6], [7]. Q ~ C W is the width, L is the length of the channel and (P) is the effective channd mobility. The capaci- where Ç and 3-1 are exponential functions. This shows that the source-driven and drain-driven componentsare controlled independently by VSB and VDB.However, - i / c B , acting through the surface potential, also controls both components in a symmetric and multiplicative fashion. In this mode of operation the MOS transis- A. G. Andreou and K. A. Boahen 164 tor has been called a difisor [ 151 in analogy with the variable conductance electrical junctions in biological systems. An expression for the current in an NMOS transistor operating in subthreshold can thus be written [6], [7] as: divider ratio closer to unity. A larger surface potential also reduces C&p. The parameter K takes values between 0.6 and 0.9. ID l. The diffusor is a term adopted in [151to describe the exploitation of diffusion transport in MOS transistors to spread signals in a manner analogous to gap junctions between neural cells. E IDS = I n 0 Sexp(~nVGB / Vt 1 x [exp(-VSB/h) - exp(-VDB/Vt)l (55) References and for a PMOS ID ISD = IpO Sexp(-Kp VGB/Vt) x [exp(VSB/Vt) - exp(VDB/Vt)] (56) The terminal voltages VGB, VSB, VDB are referenced to the substrate. The constant IOdepends on mobility (P) and other silicon physical properties. S is a geometry factor, the width W to length L ratio the device. For devices that are biased with‘ VDS 2 4 V,, (saturation) the drain current is reduced to: IDS = S I n o exp(1 - Kn)vBS/vt) exp(KnVGS/Vt) (57) This shows explicitly the dependence on VBS and the role of the bulk as a back-gate that underlies this. This equation, having only the dependence on VGS, and VBs, is used for circuit designs where devices operate in saturation as transconductance amplifier. However, channel-length modulation (Early effect) -which we have ignored completely-becomes significant in saturation. So the device equations must be augmented with terms that model this effect to accurately predict the output conductance. The parameter K is defined as Cl% K = Czx +G e p (58) The physical significance of K is apparent if the observation is made that that the oxide and depletion capacitances form a capacitive divider between the gate and bulk terminals that determines the surface potential [7]. Lighter doping reduces C&,p, and pushes the 1. B. Gilbert, “Translinear circuits: A proposed classification.” Electronics Letters, 11(1), pp. 14-16, 1975; errata, 11(6), p. 136. 2. E. Seevinck, “Analysis and synthesis of translinear integrated circuits.” Studies in Electrical and Electronic Engineering, vol. 31. Elsevier: Amsterdam, 1988. 3. B. Gilbert, “Current-mode circuits from a translinear viewpoint: a tutorial.” In Chapter 2 of C. Toumazou, F. J. Lidgey and D. G. Haigh, eds., Analogue IC Design: The CurrentMode Approach, IEEE Circuits and Systems Series, vol. 2. Peter Peregrinus Ltd: London, 1990. 4. E. Seevinck and R. J. Wiegerink, “Generalizedtranslinear circuit principle.” IEEE Journ. of Solid-state Circuits, SC-26(8), PP. 1198-1102, August 1991. 5. E. A. Vittoz and J. Fellrath, “CMOS analog integrated circuits based on weak inversion operation.” IEEE Journ. of Solid-state Circuits, SC-12(3), pp. 224-231, June 1977. 6. E. A. Vittoz, “Micropower techniques.” In Y. P. Tsividis and P. Antognetti, eds., VLSI Circuits for Telecommunications. Prentice Hall: 1985. 7. C. A. Mead, Analog VLSI and Neural Systems. AddisonWesley: Reading, MA, 1989. 8. C. A. Mead, “Neuromorphicelectronic systems.” Proceedings IEEE, 78(10), PP. 1629-1636, Oct. 1990. 9. A. G. Andreou, “Low power analog VLSI systems for sensory information processing,” in B. Sheu, Edgar Sanchez-Sinencio and M. Ismail, eds., Microsystems Technologiesfor Multimedia Applications: An Introduction. IEEE Press: Los Alamitos, CA, 1995. 10. K. A. Boahen, “Retinomorphic vision systems.” Proceedings of MicroNeuro-96. IEEE Computer Society Press: Los Alamitos, CA, 1996. 11. A. G. Andreou and K. A. Boahen, “Synthetic neural circuits using current domain-signalrepresentations.’’Neural Computation, 1, pp. 489-501, 1989. 12. A. G. Andreou, “Synthetic neural systems using current-mode circuits.” Proceedings of the IEEE 1990 International Symposium on Circuits and Systems, New Orleans, pp. 2428-2432, May 1990. 13. A. G. Andreou, K. A. Boahen, P. O. Pouliquen, A. PavasoviC, R. E. Jenkins, and K. Strohbehn, “Current-modesubthreshold MOS circuits for analog VLSI neural systems.” IEEE Trans. on Neural Networks, 2(2), pp. 205-213, March 1991. 14. M. Cohen and A. G. Andreou,“ Current-mode subthreshold MOS implementation of the Herault-Jutten autoadaptive network.”IEEE Journ. ofSolid-State Circuits, 27(5), pp. 714-727, May 1992. Translinear Circuits in Subthreshold MOS 15. K. A. Boahen and A. G. Andreou, “A contrast sensitive silicon retina with reciprocal synapses.” In J. E. Moody, S. J. Hanson, and R. P.Lippmann, eds., Advancesin Neural Information Processing Systems 4. Morgan Kaufmann Publishers: San Mateo, CA, 1992. 16. A. G. Andreou and K. A. Boahen, “Neural information processing (II).” In Chapter 8 of Analog VLSI Signal and Znformation Processing, M. Ismail and T. Fiez, eds. McGraw-Hill: 1994. 17. A. G. Andreou and K. A. Boahen, “A 48,000 pixel, 590,000 transistor silicon retina in current-modesubthresholdCMOS.” Proceedings of the 37th Midwest Symposium on Circuits and Systems, pp. 97-102, Lafayette,-Louisiana, August 1994. 18. A. G. Andreou and K. A. Boahen, “A 590,000 transistor, 48,000 pixel contrast sensitive,edgeenhancingCMOS imagersilicon retina.” Proceedings of the 16th Conference on Advanced Research in V U I , pp. 225-240, Chapel Hill, North Carolina, March 1995. 19. K. Yang and A. G. Andreou, “A multiple input differential amplifierbased on charge sharing on a floatinggate MOSFET.” Journal of Analog Integrated Circuits and Signal Processing, 6, pp. 197-208, November 1994. 20. B. A. Minch, C. Diorio, P. Hasler, and C. A. Mead, “Translinear circuits using subthreshold floating-gateMOS transistors.” Journal of Analog Integrated Circuits and Signal Processing, this issue. 21. A. Fabre, “Dual translinear voltage/current convertor.” Electronics Letters, 19(24), pp. 1030-1031, 1983. 22. A. Fabre, “Wideband translinear current convertor.” Electronics Letters, 20, pp. 241-242, 1984. 23. P. M. Furth and A. G. Andreou, “A high-drive low-power BiCMOS buffer using compound PMOSLNPN transistors.” Proceedings of the 36th Midwest Symposium on Circuits and Systems, pp. 1369-1372, Detroit, Michigan, August 1993. 24. A. G. Andreou, K. Strohbehn, and R. Jenkins, “A silicon retina for motion computation.” Proceedings of the IEEE 1991 International Symposium on Circuits and Systems, Singapore, pp. 1373-1376,June 1991. 25. Z. Kalayjian, A. G Andreou, L. Wolff, and J. Williams, “Integrated l-D polarization imagers.” Proceedings of the 29th Annual Conference on Information Sciences and Systems, Baltimore, MD, March 1995. 26. X. Arreguit, E. A. Vittoz, F. A. van Schaik, and A. Mortara, “Analog implementation of low-level vision systems.” Proceedings of European Conference on Circuit Theory and Design, pp. 275-280. Elsevier Publishing: September 1993. 27. T. Delbrück, “‘Bump’ circuits for computing similarity and dissimilarity of analog voltages.” Proc. Znt. Joint Con. on Neural Networks, Seattle WA, pp. 475-479, 1991. 28. M. Cohen and A. G. Andreou, “MOS circuit for nonlinear Hebbian learning.” Electronics Letters, 28(9), pp. 809-810, April 1992. 29. M. van der Gevel and J. C. Kuenen, “Square-root X ciqcuit based on a novel, back-gate-using multiplier.” Electronics%tters, 30(3), pp. 183-184,February 1994. 30. M. H. Cohen and A. G. Andreou, “Analog CMOS integration and experimentation with an autoadaptive independent component analyzer.” ZEEE Transactions on Circuits and Systems: Part II: Analog and Digital Signal Processing, 42(2), pp. 6577, February 1995. 31. T. S. Lande, J. A. Nesheim, and Y. Berg, “Auto correlation in micropower analog CMOS.” Journal of Analog Integrated Circuits and Signal Processing, 7, pp. 61-68, January 1995. 165 32. E. A. Vittoz, “Analog VLSI implementation of neural networks.” In Handbook of Neural Computation, Institute of Physics Publishing and Oxford University Press, USA. 33. B. L. Hart, “Translinear circuit principle: a reformulation.” Electronics Letters, 15(24), pp. 801-803, Nov. 22, 1979. 34. Mead, C. A. and Mahowald, M. A., “A silicon model of early visual processing.” Neural Networks, 1, pp. 91-97, 1988. 35. E. Vittoz and X. Arreguit, “Linear networks based on transistors.,’ Electronics Letters, 29, pp. 297-299, February 4, 1993. 36. K. Bult and G. J. G. M. Geelen, “An inherently linear and compact MOST-only current division technique.” IEEE J. SolidState Circuit, SC-27(12), pp. 1730-1735, December 1992. 37. M. Tartagni and P. Perona, “Computing centroids in currentmode technique,”Electronics Letters, 21, October 1993. 38. J. M. Buhmann, M. Lades, and R. Eeckman, “Illuminationinvariant face recognition with a contrast sensitive silicon retina.” In J. D. Cowan, G. Tesauro, and J. Alspector, eds., Advances in Neural Information Processing Systems 6. Morgan Kaufmann Publishers: San Mateo, CA, 1994. 39. J. E. Dowling, The Retina: An Approachable Part of the Brain. The Belknap Press of Harvard University: Cambridge, MA, 1987. 40. T. Poggio, V. Torre, and C. Koch, “Computational vision and regularization theory.” Nature, 317, pp. 314-319,1985. 41. C. A. Mead and T. Delbrück, “Scanners for visualizing activity in analog VLSI circuitry.’’ Analog Integrated Circuits and Signal Processing, 1(2), Oct. 1991. 42. A. PavasoviC, A. G. Andreou, and C. R. Westgate, “Characterization of subthresholdMOS mismatch in transistors for VLSI systems.” Journal of Analog Integrated Circuits and Signal Processing, 6, pp. 75-84, June 1994. 43. B. Gilbert, “A monolithic 16-channelanalog array normalizer.” IEEE Journ. of Solid-state Circuits, SC-19(6), 1984. 44. B. Gilbert, “Nanopower nonlinear circuits based on the translinear principle.” In Workshop on: Hardware implementation of neural nets and synapses, San Diego, CA, January 1988. 45. F. J. Pineda and A. G. Andreou, An Analog Neural Network Inspired by Fractal Block Coding, Advances in Neural Information Processing Systems, Vol. 7. Morgan Kaufmann Publishers: San Mateo, CA, 1995. 46. J. J. Ebers and J. L. Moll, “Large-signal behavior of junction transistors.” Proceedings IRE, 42, pp. 1761-1772, December 1954. 47. R. S . Muller and T. I. Kamins,Device Electronicsfor Integrated Circuits. John Wiley and Sons: New York, 1977. 48. J. E. Meyer, “MOS models and circuit simulation.” RCA Review, 32, pp. 42-63, March 1971. 49. Y. Tsividis, Operation and Modeling of the MOS Transistor. McGraw-Hill: New York, 1987. 166 A. G.Andreou and K. A. Boahen Andreas G. Andreou received the M.S.E. and Ph.D. in electrical engineering and computer science from Johns Hopkins University, Baltimore, in 1983 and 1986, respectively. From 1987 to 1989 he was a postdoctoral fellow and associateresearch scientist at Johns Hopkins where he became Assistant Professor in 1989 and Associate Professor in 1993. His research interests are in the areas of device physics, integrated circuits and neural computation. Dr. Andreou is a member of Tau Beta Pi and a member of I.E.E.E. Kwabena A. Boahen is a Ph.D. student at Caltech in the Computation and Neural Systems program after completing a B.S.N.S.E. degreein electrical and computer engineering at Johns Hopkins University, Baltimore, MD. His research at Caltech involves analog VLSI models of biological computation, with an emphasis on retinal computation and chip-to-chip communication. Mr. Boahen is a member of Tau Beta Pi and a student member of the I.E.E.E.

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

Download PDF

advertisement