You are on page 1of 5

This article has been accepted for publication in IEEE Transactions on Circuits and Systems--II: Express Briefs.

This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TCSII.2023.3324584

3.6-pJ /spike, 30-Hz Silicon Neuron Circuit in


0.5-V, 65 nm CMOS for Spiking Neural Networks
Srikanth Vuppunuthala, Student Member, IEEE, Vijay Shankar Pasupureddi, Senior Member, IEEE,

Abstract—In large-scale neuromorphic systems, a spiking neu- the standard current-mode log-domain circuit [9] by enabling
ral network (SNN) provides a promising solution for energy- control of circuit gain through an additional independent
efficient computing. Optimizing SNN building blocks like neurons bias voltage. However, high supply voltage and short-circuit
and synapses can further enhance computing efficiency. Due to
their simplicity and computational efficiency, integrate-and-fire currents flowing through the inverters during the switching
neuron (I&F) models are widely used in SNNs. In the past time cause significant power consumption. Several neuron
proposals of I&F neuron models, the main issue is the short- circuits [10][11] have recently been implemented using the
circuit currents in CMOS inverters, which significantly inhibit FDSOI process. This choice is driven by the ultra-low leakage
the circuit’s optimal performance and energy efficiency. This currents offered by the FDSOI process, despite it not being
paper presents an energy-efficient differential-pair integrator
(DPI) based I&F silicon neuron (SiN) circuit by utilizing the cost-effective to produce biologically plausible dynamics with
subthreshold source-coupled logic (STSCL) circuit topology as a time constants in the order of seconds. In recent years, research
replacement for CMOS inverters and incorporating a low supply on emerging stochastic artificial neurons (SANs) [12]-[16] has
voltage to mitigate leakage currents effectively. The proposed SiN surged due to their simpler circuits, inherent randomness, and
is implemented in 65 nm CMOS technology with a supply voltage smaller size compared to CMOS neurons. While SANs can
of 0.5-V and has an energy consumption of 3.6 pJ/spike with a
spiking frequency (rate) of 30 Hz. replicate basic functions like firing, refractory periods, tunable
output frequencies, and frequency adaptation, they cannot
Index Terms—After-hyperpolarization (AHP), differential-pair mimic advanced features of biological neurons, such as lateral
integrator (DPI), integrate-and-fire neuron, silicon neuron, spik-
ing neural network, subthreshold source-coupled logic. inhibition, variable spiking modes, and chaos [17]. CMOS
neurons are cost-effective and mimic the advanced features of
biological neurons, however, the deployment of subthreshold
I. I NTRODUCTION CMOS inverters as reported in [7][8], causes a brief period

T HE human brain’s exceptional energy efficiency and


massive parallelism in complex cognitive tasks encourage
building systems that mimic the biological brain [1]. The third-
during the transition between logic levels when both the pull-
up and pull-down transistors simultaneously conduct, leading
to significant short-circuit power consumption.
generation artificial neural network (ANN) is a brain-inspired To address this issue, this paper presents a DPI-based I&F
spiking neural network (SNN) [2] that emulates biological neuron circuit based on the STSCL topology with drain-
brain functions with better bio-plausibility than other neural substrate shorted pMOS load devices [18]. STSCL circuits,
networks under tight power constraints. This can paves a new with their differential structure, can minimize the overlap of
path for the implementation of forthcoming intelligent devices conducting paths, thus reducing short-circuit power. Moreover,
for ultra-low power real-time computing applications such as STSCL circuits operate at lower supply voltages, effectively
health monitoring, IoT, and mobile devices. Many SiNs are reducing subthreshold leakage current and dynamic power
needed in SNNs to imitate complex brain-like functionality. consumption. The STSCL topology used offers additional
So, individual SiN hardware must be energy-efficient, com- advantages. Firstly, the delay of the gate can be controlled
pact, and robust. by the tail bias current. Secondly, the load resistance of the
Numerous VLSI models of neuromorphic SiNs have been STSCL gate can be controlled by the pMOS load device
implemented in the past ranging from simple integrate-and- source-gate voltage, providing control over the gate’s delay.
fire models to sophisticated conductance-based models [3]-[6] These benefits result in two extra degrees of control over the
depending on the intricacy and degree of emulation. In [7], neuron spiking frequency modulation. The proposed STSCL-
a leaky integrate-and-fire (I&F) neuron circuit is presented, based I&F neuron is implemented in 65 nm CMOS with
integrating spike-frequency adaptation, a tunable refractory a supply voltage of 0.5-V and energy consumption of 3.6
period, and voltage threshold modulation. However, significant pJ/spike with a biologically plausible spiking rate of 30 Hz.
power dissipation occurs due to the short-circuit currents This paper is organized in the following manner. The
flowing through the inverters during the switching time and neuron circuit topology is presented in section II, followed by
the DC current flowing through the source-follower circuit. In simulation results and discussions in section III, and finally,
[8], a DPI-based I&F neuron is introduced for implementing conclusions are presented in section IV.
tunable dynamic conductances, which offers an advantage over
II. T HE S ILICON N EURON C IRCUIT T OPOLOGY
Srikanth Vuppunuthala and Vijay Shankar Pasupureddi are with the
School of Electrical Sciences, Indian Institute of Technology Bhubaneswar, This section presents the proposed design of the neuron
Bhubaneswar, Odisha, India (email: sv35@iitbbs.ac.in; vijay@iitbbs.ac.in). circuit. To design the neuron circuit HVT (High threshold

Authorized licensed use limited to: Indian Institute of Technology - BHUBANESWAR. Downloaded on February 06,2024 at 10:23:56 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Circuits and Systems--II: Express Briefs. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TCSII.2023.3324584

voltage) FETs have been preferred due to their low subthresh- From the EKV model [19], the subthreshold pMOS device
old leakage currents for the applied gate-to-source voltage, as I–V characteristics are given by
illustrated in Fig. 1 to get the required low currents.  V 
 
VBG −VT 0 SB VDB
ISD = I0 · e np UT
e UT − e UT (1)

Where ISD is the source to drain current, np is the subthresh-


old slope factor, UT is the thermal voltage, and VT 0 is the
threshold voltage of the pMOS. VSB , VDB , VBG are the source
to the body, drain to the body, and body to gate voltages of
pMOS, respectively. For drain-substrate shorted pMOS device
(MN 1,N 2 ) VDB = 0, hence
  
VDG −VT 0 VSD
ISD = I0 · e n p UT
e UT
−1 (2)
(a) (b)
Fig. 1. Leakage current analysis of (a) nMOS and (b) pMOS with different Differentiating the ISD with respect to the VSD results in
threshold voltages in subthreshold region.
inverse of the load resistance, RSD of the STSCL gate, and it
is given by
 −1
∂ISD
RSD = (3)
∂VSD
 −1
np UT V
(np −1) npSD
−VSD
RSD = (np − 1) e UT
+e np UT
(4)
Ib
where  
VSG −VT 0
Ib = I0 e n p UT
(5)
Fig. 2. Block diagram of DPI-based silicon neuron circuit.
From (4) and (5), it can be demonstrated that when VSD =
0, the load resistance, RSD of pMOS load (MN 1,N 2 ) is finite
A. Neuron circuit design and equal to
The block diagram of the DPI-based silicon neuron circuit
UT −
 
VSG −VT 0
is shown in Fig. 2 and its corresponding circuit level imple- RSD |VSD =0 = ·e n p UT
(6)
mentation is shown in Fig. 3. The proposed neuron circuit I0
comprises an input DPI circuit (ML1 -ML5 ) that represents This implies that the load devices (MN 1,N 2 ) possess the
the leak conductance behaviour of the real neurons, a spike capability to elevate the output node VN or VP to VDD (Fig.
generating STSCL-based positive feedback mechanism (MN 1 - 3). This property of the STSCL circuit helps to make Iadp and
MN 10 ) that models the Sodium (Na+) channels of neurons. If b to zero when Vmem is less than the Vth , which eventually
A spike reset circuit (MK1 -MK6 ), which models the Potas- reduces the static power dissipation of the neuron circuit.
sium (K+) channels of a biological neuron, and an AHP The typical delay (td ) of the STSCL gate is given by Eq.
circuit (MA1 -MA6 ) implements a spike-frequency adaptation 7, where VSW = RSD · Iss .
mechanism that models the neuron’s Calcium conductance.
Cmem is an integrating capacitor that models the membrane ln(2) · VSW · CL
td = (7)
capacitance of a neuron. All the capacitors used in the design Iss
are MIM (Metal-Insulator-Metal) caps. The values and sizes
Therefore, from Eq. 7, it is evident that the delay of the STSCL
of capacitors are shown in Table I.
is dependent on RSD and Iss . Hence, by tuning the Iss through
Vbn N a and RSD through Vbp N a as shown in Fig. 3, the
TABLE I
C APACITANCE VALUES AND DIMENSIONS EMPLOYED IN THE DESIGN delay of the neuron is adjusted which in turn tunes the spiking
frequency of the proposed neuron circuit.
Cmem Cahp Cr
Value 135 fF 665 fF 135 fF
Length 8 µm 18 µm 8 µm III. R ESULTS AND D ISCUSSION
Width 8 µm 18 µm 8 µm
The proposed DPI-based silicon neuron is implemented in
65 nm CMOS with a supply voltage of 0.5 V. The simulation
In the STSCL circuit shown in Fig. 4a, the pMOS devices
results of the proposed neuron circuit, such as spike frequency
with their drain-body contacts intentionally shorted, are em-
adaptation, energy consumption per spike, Monte Carlo, and
ployed as gate-controlled load devices with high resistivity.
PVT analyses are presented in this section.
Authorized licensed use limited to: Indian Institute of Technology - BHUBANESWAR. Downloaded on February 06,2024 at 10:23:56 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Circuits and Systems--II: Express Briefs. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TCSII.2023.3324584

Fig. 3. Proposed STSCL-based integrate-and-fire(I&F) neuron schematic.

(a) (b) (a) (b)


Fig. 4. (a) Subthreshold SCL gate and (b) Membrane voltage (Vmem ) over Fig. 5. (a) Spiking rate variation of Vmem with Vth and (b) Spike-frequency
time. adaptation: Vmem and Vahp over time.

A. Neuron circuit simulation results by the proposed neuron circuit for a Vth of 230 mV is shown
Isyn from Fig. 3 represents the presynaptic currents re- in Fig. 4b along with the spiking period of Vmem variation
ceived by the neuron’s dendrites. The input DPI integrates with Vth is shown in Fig. 5a. The spike-frequency adaptation
the received currents and the current provided by the AHP behaviour obtained by appropriately adjusting the parameters
mechanism leading to an increase in the membrane voltage of the AHP block of Fig. 3 and the neuron’s response to
(Vmem ). the constant injected current are demonstrated in Fig. 5b.
Z Vth Z t The after-hyperpolarization voltage (Vahp ) increases with each
Cmem dVmem = (Isyn − Iahp ) dt (8) spike, reducing the overall current that charges Cmem . Once
0 0 the Vahp attains a stable state, the spiking frequency of the
As Vmem approaches the switching threshold Vth of the neuron stays constant to a lower value than its initial value.
STSCL circuit, the voltage of the output node VN decreases, The neuron’s response to a step input for various settings
leading to an increase in the current If b through MN 6 and of the parameters such as Vbn na , Vbp na , Vack and Vgain
MN 8 . This, in turn, causes a sharp rise in Vmem , contributing illustrated in Fig. 6. The increase in Vbn na increases Iss ,
to the spike generation mechanism. During the spike genera- reducing both the delay in the STSCL gate and the node
tion time the Cmem charges with a current of Isyn −Iahp +If b . voltage VN . This reduction in VN leads to increased currents
Immediately after the spike is generated, the decrease in node If b and Ireset , ultimately resulting in higher charging and
VN voltage starts to increase the current Ic flowing through discharging rates of the Cmem . As a result, the spiking rate
the MK2 , consequently charging the capacitor Cr . As the of the neuron increases, as illustrated in Fig. 6a. As the
capacitor attains a sufficiently high voltage, Ireset initiates an Vbp na increases, it induces a decrease in the voltage VN .
increase and pulls Vmem down to zero, thereby establishing Consequently, the neuron’s maximum spiking rate increases,
the reset mechanism. During the spike reset mechanism the shown in Fig. 6b. Increasing the parameter Vack decreases
Cmem discharges with a current of Isyn − Iahp + If b − Ireset . the current responsible for charging Cr , leading to a slow-
The transistor sizes are adjusted in order to ensure that Ireset down reset mechanism. As a result, the neuron’s spiking rate
has the capability to pull Vmem to zero. The Vmem generated decreases, shown in Fig. 6c. As expected, increasing Vgain

Authorized licensed use limited to: Indian Institute of Technology - BHUBANESWAR. Downloaded on February 06,2024 at 10:23:56 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Circuits and Systems--II: Express Briefs. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TCSII.2023.3324584

Fig. 7. Energy per spike versus spiking rate

(a) (b)
currents through STSCL topology and mitigating leakage
currents with low supply voltage.

C. Impact of non-idealities
In order to evaluate the impact of the non-idealities on
the proposed design, Monte Carlo simulations have been
performed to evaluate the device mismatches and PVT anal-
yses have been performed to evaluate the effect of process,
temperature, and supply variations on the spiking rate of the
neuron. Monte Carlo analysis has been performed for the
neuron circuit with a sample count of 500, injecting DC
(c) (d)
current (Isyn ) into the LEAK block through ML1 , as shown
Fig. 6. Spiking rate versus input current for different values of (a) Vbn na , in Fig. 3. Moreover, the bias currents were adjusted to attain a
(b) Vbp na , (c) Vack , and (d) Vgain .
spiking rate of 120 Hz while activating the AHP circuit. The
Monte Corlo results with mean is around the desired value
reduces the Igain so the most of Iin flow into the Cmem
(≈120 Hz) with a standard deviation of 39.36 Hz as shown
which eventually reduces the integration time and increases
in Fig. 8a. The deviation in the spiking rate is due to the
the neuron’s spiking rate shown in Fig. 6d.
subthreshold operation of the transistors.
The process corner simulations have been performed to
B. Energy per spike observe the influence of process corner variation on the spiking
The energy consumed per spike (Es ) is an essential figure rate of the proposed neuron. The proposed neuron can produce
of merit (FoM) for neuron circuits to implement massively the spiking in all the process corners for the corresponding
parallel large-scale neuromorphic processors. It is given by input stimulus current (Isyn ) with some variation in the spiking
rate. The variation in spiking rate with corresponding input
Pavg
Es = (9) stimulus current at different corners is shown in Fig. 8b.
fspike The effect of temperature and supply voltage variation
where Pavg is the average power consumption of the neuron on the spiking rate of the proposed neuron is observed by
and fspike is the spiking frequency. The proposed neuron adjusting bias currents to attain a stable spiking rate of 120
circuit’s simulated Es corresponds to the spiking rate shown Hz at 25◦ C and at a supply voltage of 0.5 V. Then the change
in Fig. 7. in spiking rate is observed by varying the temperature from
The proposed neuron’s Es compared with the previous -10◦ C to 70◦ C and the simulated results is shown in Fig. 9a.
works in Table II. The proposed neuron circuit achieved better The change in spiking rate with supply variation from 450
Es compared to previous works by eliminating short circuit mV to 550 mV is shown in Fig. 9b. The variation in the

TABLE II
P ERFORMANCE COMPARISON OF PROPOSED NEURON WITH PREVIOUS WORKS

Silicon Neurons (SiNs) Emerging Stochastic Artificial neurons (SANs)


Reference This Work [7] [8] [5] [6] [10] [12] [13] [14] [15] [16]
HfO2 Antiferro- V O2
Pt/HfAlOx GST
Material Silicon Silicon Silicon Silicon Silicon Silicon (Hafnium magnetic (Vanadium
TiN/Ag/Pt (Ge2 Sb2 T e5 )
Oxide) SOT 1 Dioxide)
MTJ PCM MIT
65 nm 350 nm 350 nm 180 nm 28 nm 22 nm Ferro- (Magnetic (Phase- (Metal-
Technology Filament
CMOS CMOS CMOS CMOS CMOS FDSOI electric Tunneling Change Insulator
Junction) Memory) Transition)
Type Mixed Mixed Mixed Mixed Mixed Mixed - - Digital Digital -
Vdd 0.5 V 3.3 V 3.3 V 1.8 V 0.7-1 V 0.8 V - - - - -
Frequency 30 Hz 100 Hz - 30 Hz - 30 Hz - 50 kHz Tens of kHz 35-40 kHz 3 GHz
Energy/spike 3.6 pJ 900 pJ 267 pJ 883 pJ 2.3-30 nJ 14 pJ 16 f J 1-10 pJ - 5 pJ 5.6 f J

Authorized licensed use limited to: Indian Institute of Technology - BHUBANESWAR. Downloaded on February 06,2024 at 10:23:56 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Circuits and Systems--II: Express Briefs. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TCSII.2023.3324584

R EFERENCES
[1] M. Davies et al., “Loihi: A Neuromorphic Manycore Processor with
On-Chip Learning,” in IEEE Micro, vol. 38, no. 1, pp. 82-99, Jan-
uary/February 2018.
[2] W. Maass, “Networks of spiking neurons: The third generation of neural
network models,” Neural Netw., vol. 10, no. 9, pp. 1659–1671, 1997.
[3] M. Mahowald and R. Douglas, “A silicon neuron,” Nature, vol. 354, pp.
515–518, 1991.
[4] M. F. Simoni, G. S. Cymbalyuk, M. E. Sorensen, R. L. Calabrese and
S. P. DeWeerth, “A multiconductance silicon neuron with biologically
matched dynamics,” in IEEE Transactions on Biomedical Engineering,
vol. 51, no. 2, pp. 342-354, Feb. 2004.
[5] S. Moradi, N. Qiao, F. Stefanini and G. Indiveri, “A Scalable Multicore
(a) (b) Architecture With Heterogeneous Memory Structures for Dynamic Neu-
Fig. 8. (a) Monte Carlo results for the distribution of the spiking rate of romorphic Asynchronous Processors (DYNAPs),” in IEEE Transactions
proposed neuron circuit and (b) Spiking rate versus input current at different on Biomedical Circuits and Systems, vol. 12, no. 1, pp. 106-122, Feb.
process corners. 2018.
[6] C. Mayr et al., “A Biological-Realtime Neuromorphic System in 28
nm CMOS Using Low-Leakage Switched Capacitor Circuits,” in IEEE
spiking rate of the proposed neuron is because the leakage Transactions on Biomedical Circuits and Systems, vol. 10, no. 1, pp.
currents of subthreshold transistors depend exponentially on 243-254, Feb. 2016.
both temperature and supply voltage [20] [21]. [7] G. Indiveri, E. Chicca and R. Douglas, “A VLSI array of low-power spik-
ing neurons and bistable synapses with spike-timing dependent plasticity,”
in IEEE Transactions on Neural Networks, vol. 17, no. 1, pp. 211-221,
Jan. 2006.
[8] P. Livi and G. Indiveri, “A current-mode conductance-based silicon
neuron for address-event neuromorphic systems,” 2009 IEEE International
Symposium on Circuits and Systems, Taipei, Taiwan, 2009, pp. 2898-
2901.
[9] D. R. Frey, “Log-domain filtering: An approach to current-mode filtering,”
IEE Proceedings G: Circuits, Devices and Systems, vol. 140, no. 6, pp.
406–416, December 1993.
[10] A. Rubino, C. Livanelioglu, N. Qiao, M. Payvand and G. Indiveri,
“Ultra-Low-Power FDSOI Neural Circuits for Extreme-Edge Neuromor-
phic Intelligence,” in IEEE Transactions on Circuits and Systems I:
Regular Papers, vol. 68, no. 1, pp. 45-56, Jan. 2021.
(a) (b) [11] H. Eslahi, T. J. Hamilton and S. Khandelwal, “Compact and Energy
Fig. 9. Spiking rate with (a) temperature variation and (b) supply variation. Efficient Neuron With Tunable Spiking Frequency in 22-nm FDSOI,” in
IEEE Transactions on Nanotechnology, vol. 21, pp. 189-195, 2022.
[12] Lu, Yi-Fan and Li, Yi and Li, Haoyang and Wan, Tian-Qing and Huang,
D. Layout Xiaodi and He, Yu-Hui and Miao, Xiangshui, “Low-Power Artificial
Neurons Based on Ag/TiN/HfAlOx/Pt Threshold Switching Memristor
The layout of the proposed neuron circuit is shown in Fig. for Neuromorphic Computing,” in IEEE Electron Device Letters, vol. 41,
10. The layout of the neuron circuit, including the bias circuit, no. 8, pp. 1245-1248, Aug. 2020.
occupies an active area of 55 × 37 µm2 . [13] Dutta Sourav, Schafer Clemens, Gomez Jorge, Ni Kai, Joshi Siddharth,
Datta Suman, “Supervised Learning in All FeFET-Based Spiking Neural
Network: Opportunities and Challenges,” Frontiers in Neuroscience, vol.
41, 2020.
[14] Kurenkov, Aleksandr and DuttaGupta, Samik and Zhang, Chaoliang and
Fukami, Shunsuke and Horio, Yoshihiko and Ohno, Hideo, “Artificial
Neuron and Synapse Realized in an Antiferromagnet/Ferromagnet Het-
erostructure Using Dynamics of Spin–Orbit Torque Switching,” Advanced
Materials, vol. 31, April 2019.
[15] Tuma, Tomas and Pantazi, Angeliki and Gallo, Manuel and Sebastian,
Abu and Eleftheriou, Evangelos, “Stochastic phase-change neurons,”
Nature Nanotechnology, vol. 11, May 2016.
[16] Yi, Wei and Tsang, Kenneth and Lam, Stephen and Bai, Xiwei and
Crowell, Jack and Flores, Elias, “Biological plausibility and stochasticity
in scalable VO2 active memristor neurons,” Nature Communications, vol.
9, November 2018.
Fig. 10. Layout of the proposed neuron circuit, including the bias circuit. [17] Li, Zong-xiao and Geng, Xiao-ying and Wang, Jingrui and Zhuge,
Fei, “Emerging Artificial Neuron Devices for Probabilistic Computing,”
Frontiers in Neuroscience, vol. 15, 2021.
IV. C ONCLUSIONS [18] A. Tajalli, E. J. Brauer, Y. Leblebici and E. Vittoz, “Subthreshold Source-
Coupled Logic Circuits for Ultra-Low-Power Applications,” in IEEE
This work presented an energy-efficient STSCL-based in- Journal of Solid-State Circuits, vol. 43, no. 7, pp. 1699-1710, July 2008.
tegrate and fire neuron circuit topology implemented in 0.5- [19] C. Enz and E. Vittoz, Charge-Based MOS Transistor Modeling: The
V, 65 nm CMOS with a refractory period, spike frequency EKV Model for Low-Power and RF IC Design. New York: Wiley, 2006.
[20] Sourikopoulos, Ilias and Hedayat, Sara and Loyez, Christophe and
adaptation, spiking frequency modulation mechanisms. The Danneville, François and Hoel, Virginie and Mercier, Eric and Cappy,
proposed neuron produces biologically feasible neural dynam- Alain, “A 4-fJ/Spike Artificial Neuron in 65 nm CMOS Technology,”
ics by matching the time constants to those observed in natural Frontiers in Neuroscience, vol. 11, 2017.
[21] L. Zhang, J. Zhang, Y. Wang, X. Zhang and R. Huang, “A Multi-Mode
signals like speech. Therefore, the proposed neuron topology is Silicon Neuron Circuit With High Robustness Against PVT Variation,”
a good candidate for deployment in large-scale neuromorphic 2018 IEEE International Symposium on Circuits and Systems (ISCAS),
systems for energy-efficient computing. pp. 1-4, 2018.

Authorized licensed use limited to: Indian Institute of Technology - BHUBANESWAR. Downloaded on February 06,2024 at 10:23:56 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.

You might also like