A Novel Exponential Pulse-Decay Modulation Multiplier for VLSI Neural

Download Full-Text PDF Cite this Publication

Text Only Version

A Novel Exponential Pulse-Decay Modulation Multiplier for VLSI Neural

Raed Althomali

Department of Electrical and Electronics Engineering, Yanbu Industrial College,Yanbu Al-Sinaiyah, KSA

Abstract:- This paper introduces a new EPDM multiplier circuit

t

based on the dynamic charge of MOS-differential configuration resistor. In addition, shows that the actual response of a

Q e

U(t) dt

(1)

conventional EPDM multiplier is not linear and pseudo linear resistor realization is VDS and VTN dependent. This active resistor has matched devices biased in the triode region and enables the multiplier to enhance the output linearity. The proposed technique removes the influences of bulk-source and drain-source voltages on circuit resistance by taking into account the effects of channel length modulation, making the developed new multiplier suitable enough for the architectures implementation of several analog VLSI neural networks.

, where U t are the step function and Q is the charge delivered. and are constant values controlled by the

input and weight signals respectively. The neuron input signal with its weight simultaneously enables two quadrant multiplication complementary arrangements to charge packet of the distribution bus.

I K\ W V V

1 V 2

(2)

Index TermsArtificial Neural Network, Synapse circuit design and Exponential Pulse Decay Modulation Technique.

  1. INTRODUCTION

    The application of artificial neural networks (ANN) in analog chip designing resulted in the solutions to problems of geometry, propagation delay, parametric nonlinearities and power consumption which were once considered difficult. [2] And [7] have shown that, due to its power consumption and critical matching, designing of an analog multiplier is among the challenges of analog circuit design. The multiplier finds it applications in communication circuits, including filters,

    TN N L n TN n TN 2 TN

    From equation (2) it is evident that device N2 always conducts in the saturation region and succeeds to regulate the dependence of overall currents of N1 and N2 on the square of n . On the other hand, the coefficient '' that represents the linear dependence of ID on VDS increases considerably for the very short transistor VDS , and has considerable effect on ID in the saturation region, as shown in the following equations.

    +VDD

    modulators and mixers, and can be implemented using both BJT and MOS technologies [14], [15], and [16]. Since BJT consumes more power, the MOS multiplier with active load resistor becomes strong candidate for ANN usage for low-

    Clock

    Vs

    Dynamic

    N4 Signal

    V's

    P4 Ccp

    P1 P2 P3

    NODE A

    power design. Under the conditions of negligible channel length modulation and mobility reduction, [3], [6] and [11] have shown linear response of an EPDM multiplier when compared to conventional analog techniques in terms of low power and area requirements. An EPDM multiplier

    Clock

    N5

    -VSS

    Vw

    Ccn

    NODE B

    N1 N2

    ExcOut

    InhOut

    Cbus N3

    independent of VDS and VTH changes is proposed. The paper is divided into following sections: Section II and III present the basic operation and design of a conventional and modified EPDM multiplier with double MOS configuration,

    Static Weight CW

    -VSS

    respectively. Section IV presents the analysis and results of SPICE simulation of both types of EPDM multipliers. Section V elaborates on the comparison of both EPDM circuits

    Figure 1: Circuit of the EPDM multiplier circuit.

    followed by the conclusion.

    I K\ W V V 2 1 V

    1. With

      D GS TH DS

      2L

  2. II. THE CONVENTIONAL EPDM CIRCUIT

    The conventional EPDM circuit, as shown in Figure 1,

    channel length modulation, the currents become:

    d

    modulates the input by a weight signal through the decay time of a single-sided exponential current pulse. The

    ITN

    CTN dt

    (4)

    delivered charge packet Q can be obtained using equation

    1.

    V2

    ID4

    M4 M3

    ID3

    VC2

    V1

    ID2

    M2 M1

    ID1

    V2 V1

    I2 I1

    +VDD

    VC2

    VC1

    VB

    I2 V3

    rac/2

    I1 V3

    rac/2

    V3

    VIN

    Dynamic Signal

    Ccp

    Ccn

    M9 M8

    NODE A

    NODE B

    M7 M6

    M10

    M5

    M11

    11

    Cbus

    M12

      1. (b)

        Figure 2: (a) Double-MOS differential configuration resistor

      2. The equivalent circuit of the double-MOS Differential resistor

    Vw M4

    Static Weight Cw

    M3 M2 M1

    VC1

    -VSS

    , where CTN is the total capacitance at point B.

    Figure 3: Schematic diagram of the new EPDM multiplier circuit.

    V V 3 V t

    C W DD TN

  3. MODIFIED EPDM CIRCUIT

    (t) V \ cn V V

    2 n

    n S C TN TN V V

    • 2V e

    TN

    V

    W DD

    V 3 V

    TN

    I1 ID1

    • ID3

      (8)

      W

      VTN

      DD 2

      TN

      (5)

      I 2 ID 2 ID 4

      (9)

      VW

      VDD 2VTN

      By using equations (8) and (9) to find (I1 – I2) gives:

      W

      1 W

      I I I

      K \ V

      V V V

      (10)

      Q I

      dt

      K \ 2 (t) dt

      (6)

      TN 1

      2 C1

      C 2 1 2

      n N 3 2 N L n

      L

      Clearly the amount of charge depleted over one clock cycle or

      The equivalent resistor rac can be found as:

      S

      S

      injected into the distribution bus is not linear and depends on the square ofV \ . This is because the pseudo linear resistor

      formed by matched devices is not linear and depends on V 2

      rac

      V1 V2

      I1 I2

      1

      C1 C2

      C1 C2

      K \ W V V

      L

      (11)

      DS

      and VTN as indicated in the equation (7).

      1 ITN

      K \ W [3 V 2

      • 2V V

        (7)

        All devices are assumed to be in the ohmic region. Equation

        R V

        L 2 DS

        DS TN

        (11) only holds when:

        DS

        V , V min V V

        , V V

        (12)

        V 2V

        V 2 ]

        1 2 C1 TH

        C2 TH

        GS TN 2 TN

        Because of the non-linearity in the distribution bus it is important to find an alternative circuit that not only makes the output linear, but also removes the influences of drain source voltage and bulk voltage, taking into consideration the channel-length modulation.

        To eliminate the influences of VDS, and VTN upon the AC resistor realization, a modification in the conventional EPDM circuit is made. The principle employed here is to use four identical devices biased so that the influences of VDS and VTN

        By substituting the previous circuit of the AC resistor in the EPDM synapse multiplier circuit to become the modified EPDM multiplier circuit, as seen in Figure. 3.

  4. ANALYSIS OF THE NEW EPDM MULTIPLIER

Referring to Figure 3 and the total current described by equation (10), it is interesting to compare the AC resistor realization between conventional and new EPDM circuits.

variations are canceled out. This design uses a double-MOS

V V V

(13)

differential configuration resistor, as shown in figure. 2.

TN L W

C1 1 2

1 X

1 X

The characteristics of the AC resistance can be obtained by assuming all of the devices are matched and operate in the

V V

and

V2 VX

  • . where is a small value,

ohmic region. I1 and I2 are obtained as follows:

represent the voltage drop across the active resistor of

transistor M5 and VX

is the instantaneous voltage at node B.

ITN becomes as indicated below

I K \ W (V V )

(14)

channel MOS transistors having W/L = 5/2 and model

TN L W C1

parameters of VTH = -.75V and KP = 8 A/V².

Using equations (11) and (14) results in

K \ WV V

d L W C1

n

. (15)

d t CTN

rac CTN

Using Laplace transformation, it can be shown that:

S n

(s)

. (16)

S

(a)

The initial condition of n (t) is

(t 0) V

Cc n . (17)

n I N C

TN

Therefore,

Ccn

(18)

n (t)

t VIN C

TN

Q I

dt 1 K\ W (V

V ) dt

(19)

2

2

n M

12 L

GS12 TH

2

2

(b)

GS

GS

V

12

VTH

VX VDD VTH

n (t)

(20)

Figure 4: SPICE simulation results showing the signal outputs on nodes A & B with different weight voltage of: (a) -3V, (b) +3V.

Q 1 K \ W 2 dt

(21)

n 2 L n

The double MOSFET differential resistor is found to be superior in linearity and can be operated under VDS and VTH variations. In the operating region (ohmic), and to avoid more

complications in the system model, there is no physical

justification to multiply 1 VDS with the drain current.

Therefore, since

VDS is so small in that ohmic region and

(a)

has little effect on ID, a transient multiplier characteristic of an EPDM circuit and quadrant multiplier modified synapse circuit has been simulated in Figures 4 and 5. With Vinput of 5 V and Vweight of 0V, the transfer curves of this synaptic circuit for weights from -3 V to +3 V are shown. The DC bias voltages are 5V, -5V, 4V, and -3V for VDD, VSS, VC1, and VC2 respectively.

Both circuits are simulated by SPICE a simulator using N- channel MOS transistors having W/L = 3/5 and model parameters of VTH = 0.75V and KN = 24 A/V², and P-

V. ADVANTAGE OF THE NEW EPDM MULTIPLIER

The modified EPDM circuit has added advantages when compared to the conventional circuit. Because there is no static path between bias voltages it will result in low power consumption when integrated at VLSI levels. It successfully reduced the influences of VDS and VBS by using the double- MOS differential configuration with matched devices operating in the ohmic region, as indicated in Figure 6. The modified circuit not only makes the AC resistor linear, but removes the influence of the bulk-source voltage. The effect of the channel-length modulation has been reduced as all

(b)

Figure 5: SPICE simulation results showing the signal outputs on nodes A & B at M11 & M12, with different weight voltage of: (a)

3V, (b)+3V.

devices are operating at the ohmic region. And, since VDS is so small in this region, it has little effect on ID. Due to neglect of the channel-length modulation and body effects, the output becomes more linear and the THD problem is ignored forever. The die area is small compared to the previous design because the smaller transistor dimensions can be chosen than those in the other circuit.

2

2

Although the linearity of the new circuit is very good over a wide range of input signals, the dependence of output charge

V

V

on S prevents this circuit from being an ideal linear multiplier. The combined trans-conductance (gm) of devices N1-N4 (or P1-P4) must be small enough to ensure a high

equivalent resistance between node B and VDD (or node A

V

V

S

S

and +VDD), to assure good capacitive coupling from \ to

node B (or to node A). It may happen to see an offset in the output due to non-ideal matching between devices in the aspect ratios, parasitic capacitances, and threshold voltages.

IV. PROPOSED EPDM ANN SYNAPSE CIRCUIT

The suggested EPDM multiplier circuit shown at Figure 7 became the more suitable synapse circuit for building the new synapse memory matching cell.

N

N

N

N

The proposed circuit is basically an enhancement EPDM with two or more gates. One of the gates is left floating and other select gates act as the gate of a regular enhancement MOSFET. Referring to Figure 8 and the total current described by equation (22), that controls the ON and OFF states of biological neurons transistors by calculates the weighted sum of all input signals.

Figure 7: Equivalent of the proposed EPDM floating gate circuits

VFGS

FG

N

N

  • Ci

    VG i VS

    The operation of the new suggested synapse memory cell

    shown in Figure 9 can be described by equation (23).

    C0

    Cj

    i 1 C0

  • Cj

    K V

    *V C

    • K V

      *V C

      \ \

      j 1

      j 1

      V

      1 IN1 W 1 G1 2 IN 2 W 2 G 2

      (23)

      Where FG is the initial charge stored on the floating gate

      out

      CG1

  • CG 2

C0

due to the programming process.

I 1 C W (V V )2

DS 2 OX L FGS TH

With two-inputs, one acts as an input vector and the other as a template vector. Each is applied to the corresponding input of its multiplier. The weight signals of each chosen EPDM

multiplier are equal in magnitude and opposite in direction.

(22) 1 C

W (

Ci V

N

N

V V )2

The output of the two matched multipliers is out of phase.

N

N

2 OX L

N

G i

i 1 C0 Cj j 1

S TH

When V1~ = zero, the proposed inverter does not turn on because the floating-gate potential is a little smaller than its inverting threshold. The inverter gives an output of logic 1,

S TH

S TH

K ( Wi VG i i 1

V V )2

which is input to the CMOS inverter.

The CMOS inverter is then activated to give logic 0 outputs. The opposite action of this work is done. When V1~Vn are non-zero signals.

Figure 6: Equivalent of programmed floating gate circuits

Figure 8: The signal voltage at the floating-gate voltage at selected gates inputs (VG1 and VG2).

Figure 9: Simulation results of the suggested synapse-memory cell at input signals VIN1=VIN2=10V and weight signals VW1=VW2=1V

The floating-gate potential of the proposed inverter is larger than its inverting threshold. It turns on and gives logic 0 output that turns off the CMOS inverter. The CMOS inverter gives an output of logic 1.

As shown at Figure 9, the results are accumulated on the MOS floating gate and the result is the difference between them.

The output illustrates either logic 1 or logic 0, according to the degree of matching between inputs.

  1. ANALYSIS OF THE PROPOSED synapse memory cell

    The proposed circuit has been utilized to produce the sum operation with no dc-current flows. As a result, a dc-current free operation of the cell has been established making it possible to build a low power system. The circuit is very simple in concept and constrution. It consumes a small chip area because there are no switches, as in most classical matching cells. The proposed synaptic memory cell design with floating-gate potential and real-time monitored during data writing achieves accuracy and high speed for analog voltage storage through the thresholding action of proposed inverter.

    Because no complicated control circuitry is required for weight updating, the new version cell allows high-density integration of a synapse cell as well as a high-speed learning operation of a proposed self-learning neural network. The static power dissipation in each inverter is zero in both of its states, where the dissipation due to the leakage currents is so small as to be neglible.

    A low-resistance path exists between the output terminal and ground or VDD. These low resistance- paths ensure the output voltage is independent of the exact values of W/L ratios or any other device parameters. Furthermore, the low-output resistance makes the inverter less sensitive to the effects of noise and other disturbances.

    V. CONCLUSION

    S

    S

    The proposed multiplier can modulate an input signal based on a locally stored weight voltage. The modified EPDM multiplier circuit can produce efficient computations in the analog domain with a reduction in power and area requirements. Although the linearity of the new circuit is very good over a wide range of input signals, the dependence of output charge on V2 prevents this circuit from being an ideal

    linear multiplier. The average actual power consumption of the new circuit will be less when compared to a conventional EPDM circuit. The combined trans-conductance (gm) of devices N1-N4 (or P1-P4) must be small enough to ensure a

    high equivalent resistance between node B and VDD (or node A and +VDD) to assure good capacitive coup- ling from V\ to

    S

    S

    node B (or to node A). There is a possibility of appearance of an offset in the output due to less than ideal matching between devices in the aspect ratios, parasitic capacitances, and threshold voltages.

    Practically, I observe that the output of the proposed matching cell is not exactly zero. This is due to the non-perfect matched device parameters and the parasitic capacitance coupled to the floating gate.

    The speed of operation of a digital system as a computer is determined by the propagation delay of the logic gates used to construct the system. To obtain lower propagation delay and faster operation, a higher process transconductance parameter should be utilized and the transistor W/L ratio increased. There are design tradeoffs because these conditions increase the dynamic power dissipation.

    REFERENCES

    1. R. Althomali, "A modified exponential pulse-decay modulation technique based on CMOS analog multiplier," 2017 International Conference on Communication and Signal Processing (ICCSP), CHENNAI, India, 2017, pp. 1967-1970. doi: 10.1109/ICCSP.2017.8286744

    2. C. Bartolozzi and G. Indiveri, Synaptic Dynamics in Analog VLSI, Neural Computation, vol. 19, no. 10, pp. 25812603, 2007.

    3. S. Moradi and G. Indiveri, An event-based neural network architecture with an asynchronous programmable synaptic memory, TBioCAS, pp. 110, March 2013.

    4. F. Galluppi, S. Davies, A. Rast, T. Sharp, L. A. Plana, and S. Furber, A hierachical configuration system for a massively parallel neural hardware platform, in Proceedings of the 9th conference on Computing Frontiers. ACM, 2012, pp. 183192.

    5. J. Seo et al., A 45nm CMOS neuromorphic chip with a scalable architecture for learning in networks of spiking neurons, in Proceedings IEEE CICC, 2011, pp. 14.

    6. R. J. Vogelstein, U. Mallik, J. T. Vogelstein, and G. Cauwenberghs, Dynamically reconfigurable silicon array of spiking neurons with conductance-based synapses, IEEE TNN, vol. 18, no. 1, pp. 253 265, 2007.

    7. F. Folowosele, R. Etienne-Cummings, and T. Hamilton, A CMOS switched capacitor implementation of the Mihalas-Niebur neuron, in BioCAS, 2009, pp. 105 108.

    8. M. Noack, M. Krause, C. Mayr, J. Partzsch, and R. Sch¨uffny, VLSI implementation of a conductance-based multi-synapse using switched capacitor circuits, in International Symposium on Circuits and Systems ISCAS 2014, 2014.

    9. E. T. Rolls, L. Dempere-Marco, and G. Deco, Holding multiple items in short term memory: A neural mechanism, PloS one, vol. 8, no. 4, p. e61078, 2013.

    10. C. Mayr and J. Partzsch, Rate and pulse based plasticity governed by local synaptic state variables, Frontiers in Synaptic Neuroscience, vol. 2, no. 33, p. 28, 2010.

    11. O. Levy, N. Ziv, and S. Marom, Enhancement of neural representation capacity by modular architecture in networks of cortical neurons, European J. Neuros. , vol. 35, pp. 17531760, 2012.

    12. M. Noack, C. Mayr, J. Partzsch, M. Schultz, and R. Sch¨uffny, A switched-capacitor implementation of short-term synaptic dynamics, in Proceedings MIXDES, 2012, pp. 214218.

    13. M. Noack, J. Partzsch, C. Mayr, and R. Sch¨uffny, Biology-derived synaptic dynamics and optimized system architecture for neuromorphic hardware, in 17th International Conference on Mixed Design of Integrated Circuits and Systems MIXDES 2010, 2010, pp. 219224.

    14. K. Ishida, K. Kanda, A. Tamtrakarn, H. Kawaguchi, and T. Sakurai, Managing subthreshold leakage in charge-based analog circuits with

      low-vth transistors by analog T- switch (AT-switch) and super cut- off CMOS (SCCMOS), Solid-State Circuits, IEEE Journal of, vol. 41, no. 4, pp. 859867, April 2006.

    15. C. Mayr, M. Noack, J. Partzsch, and R. Sch¨uffny, Replicating experimental spike and rate based neural learning in CMOS, in IEEE International Symposium on Circuits and Systems ISCAS 2010, 2010, pp. 105108.

    16. J. Hasler and B. Marr, Finding a roadmap to achieve large neuromorphic hardware systems, Frontiers in neuroscience , vol. 7, 2013.

    17. S. Brink, S. Nease, P. Hasler, S. Ramakrishnan, R. Wunderlich, A. Basu, and B. Degnan, A learning-enabled neuron array IC based upon transistor channel models of biological phenomena, Biomedical Circuits and Systems, IEEE Transactions on, vol. 7, no. 1, pp. 7181, 2013.

Dr. Raed Althomali was born in Saudi Arabia, in 1979. He received the B.S. and Electrical engineering from the Yanbu Industrial College, SA in 2008 and the

M.S. and Ph.D. degree in Electrical and Computer engineering in fiber optics and photonics from SIU, Illinois in 2014.

He currently is the Managing Director, Yanbu English Language & Prep Year Institute, Saudi Arabia. His main areas of research interest are signal processing, all- optical devices, VLSI, active filters, and neural networks. Dr. Raed is a member of the Saudi Council of Engineers and Optical Society of America.

Leave a Reply

Your email address will not be published. Required fields are marked *