LeeYeji1
ShahArati Kumari2
KangMyounggon3
ChoSeongjae1*
-
(Department of Electronic and Electrical Engineering, Ewha Womans University, 52 Ewhayeodae-gil,
Seodaemun-gu, Seoul 03760, Republic of Korea)
-
(Ewha Womans University, in the same department, and currently is with Philophos, Inc.,
Bundang, Republic of Korea)
-
(Department of Intelligent Semiconductor, University of Seoul, Seoul 02504, Republic
of Korea)
Copyright © The Institute of Electronics and Information Engineers(IEIE)
Index Terms
Digital integrated circuit, integrate-and-fire neuron, hardware implementation, neuromorphic chip, FPGA
I. INTRODUCTION
In recent years, the demand for artificial intelligence (AI) has increased owing to
its widespread application and the expansion of its scope in various fields. Simultaneously,
there is a growing need for advanced computing architectures with processing parallelism
to process vast amounts of data effectively and make rapid decisions in complex situations.
To address these challenges, neural computing inspired by the human brain has attracted
considerable attention [1]. This approach uses highly interconnected synthetic neurons and synapses to model
complex neural processes and solve difficult machine-learning problems [2]. By mimicking the distributed topology of the brain, neural systems provide significant
advantages in terms of speed and energy efficiency, making them suitable for sophisticated
information processing tasks [3,4]. In this approach, spiking neural networks (SNNs) have the potential to replicate
the parallel processing capabilities and energy efficiency of the biological nervous
system by mimicking the way neurons transmit information through electrical synapses
in the form of spikes [5,6,7]. At the core of the SNN and neuromorphic architecture is the integrate-and-fire (I&F)
neuronal circuit, which serves as a basic unit of computation that closely mimics
the behavior of biological neurons [8,9,10,11]. In recent years, the digital implementation of I&F neuron models has attracted considerable
attention as a promising alternative to analog models, providing flexibility, scalability,
and smooth integration with conventional CMOS technologies [12,13,14,15]. These digital I&F models play an important role in building complex neural architectures,
and serve as essential components of SNNs that can perform a wide range of cognitive
tasks, such as recognition, classification, and decision-making [16,17].
Thus, in this study, we explore the implications of implementing digital I&F neuron
models in neuromorphic computing systems. By utilizing the inherent parallelism and
low-power characteristics of spiking neurons, these systems demonstrate considerable
potential for the efficient and scalable implementation of AI algorithms [12,18]. Despite these developments, studies have continuously explored the future direction
and prospects of digital I&F neuron models in the SNN domain. The continued development
of digital circuit design combined with neuroscientific insights is expected to lead
to the development of more sophisticated and energy-efficient neuromorphic systems,
ultimately narrowing the gap between AI and biological intelligence [1,19,20]. This study explores the intersections of digital I&F neuron models, and evaluates
their underlying principles. In addition, we investigate the ability of these models
to accurately capture the spiking dynamics of biological neurons using field-programmable
gate array (FPGA)-based programming and implementation efforts.
II. DESIGN OF DIGITAL NEURON MODEL
A. Digital integrate-and-fire neuron model
The I&F neuron model is a fundamental element of the SNN. This model integrates incoming
signals and triggers an output spike when the membrane potential ($V_{\rm mem}$) reaches
a specific threshold. After the spike is generated, it is transmitted to the connected
neurons, resetting $V_{\rm mem}$ and preparing the neuron for the next cycle of integration
and firing. In the digital implementation of the I&F neuron model, the analog behavior
of biological neurons is replicated using digital circuits. The core components of
the digital I&F neuron model include an integrator that accumulates input signals,
a threshold comparator that determines when the accumulated potential exceeds a predefined
voltage ($V_{\rm th}$), and a reset mechanism that restores $V_{\rm mem}$ to its initial
state after spike generation.
Fig. 1 shows the logical flowchart used to achieve the digital equivalent circuit of an
I&F neuron. The integrator unit accumulates the incoming spike signals (spike_in)
over time, each of which is weighted by its synaptic weight. As these signals are
integrated, $V_{\rm mem}$ increases until it surpasses the preset $V_{\rm th}$. When
this threshold is exceeded, the comparator triggers an output spike (spike_out). Immediately
after the spike is generated, the membrane potential is reset to 0 V, allowing the
neurons to begin a new integration process. This cycle of integration, threshold comparison,
and spike generation verifies the essential behaviors of an I&F neuron circuit. This
digital implementation provides accurate control over neuronal parameters, such as
time constants and critical voltages, and can be tailored to specific applications.
Owing to the flexibility and adaptability of the digital I&F neuron circuit, it is
suitable for a wide range of neural computing tasks ranging from basic pattern recognition
to complex cognitive functions. This compact design also enables efficient expansion,
making it easier to achieve a large-scale SNN in the neural system.
Fig. 1. Workflow of the digital equivalent I&F neuron model.
B. FPGA implementation of an I&F neuron
The I&F neuron model was implemented in a Zynq multiprocessor system-on-chip (MPSoC)
FPGA using the Xilinx Vivado Design Suite. First, a hardware description language
(HDL) representation of the I&F neuron model was developed. This HDL code defines
neuron behavior, including signal integration, threshold detection, spike generation,
and membrane potential resetting. Then, an analog-to-digital converter module was
used to digitize the continuous input current fed into the analog I&F neuron. This
allowed the HDL-based design methodology to build an I&F neuron circuit by replacing
analog components by equivalent digital ones.
In the simulation, a reset signal was first enabled to create the initial conditions.
After a short delay of $1\times$ unit, the reset signal was toggled, starting with
the main simulation sequence. The program algorithm proceeded in a loop simulation
clock cycle. The input spike signal was enabled and toggled after each $1\times$ unit
interval. The integral $V_{\rm mem}$ was monitored. If it exceeded a given $V_{\rm
th}$, the comparator triggered the generation of an output spike (spike_out). $V_{\rm
mem}$ was then reset to 0 V, allowing the neurons to resume the integration process.
Subsequently, we simulated a neuron reset event by enabling the neuron_reset signal
for $1\times$ unit of the clock signal. This sequence was repeated until the clock
signal stopped, as shown in Fig. 2. The simulation results validate the functionality of the digital I&F neuron circuit,
and provide important insights into its dynamic behaviors. These results demonstrate
the ability of this circuit to accurately replicate the spiking dynamics of biological
neurons, making it a suitable candidate for the hardware-oriented nervous system.
The FPGA implementation is optimized to minimize resource usage, allowing the digital
I&F neuron circuit to be scaled to accommodate larger neuronal networks.
Fig. 2. Simulation of the designed digital equivalent of an I&F neuron using the Vivado
Design Suite.
III. RESULTS AND DISCUSSION
The FPGA implementation of the digital I&F neuron model was thoroughly analyzed to
evaluate its performance in terms of resource utilization, power consumption, and
overall design efficiency. These results provide a deeper understanding of the behaviors
of neurons and enhance the reliability, cost-effectiveness, and compactness of the
model. Table 1 provides an overview of the FPGA resources used to implement the I&F neuron digital
circuit. The design utilizes only a minimal portion of the available resources, ensuring
a simple and efficient implementation. Specifically, only 11 of the 274,080 lookup
tables (LUTs) were used, representing a utilization rate of only 0.01%. Similarly,
only 33 flip-flops (FFs) were required out of the 548,160 available units, corresponding
to a utilization rate of 0.01%. Of the 328 pins provided by the FPGA system for input/output
(IO), only five (1.52%) were sufficient to implement the neuron circuit. In addition,
only one clock buffer (BUFG) was used out of the 404 available units, representing
only 0.25% of the total. These results indicate that the design is highly resource-efficient,
with ample room for further optimization and enhancement. As shown in Fig. 3, the total on-chip power consumption was measured to be 0.71 W, with the junction
temperature maintained at 25.7 $^\circ$C. The power consumption is clearly lower than
the values obtained from the comparable FPGAs in the recent reports, 4.5 W, 4.6 W,
and 10.5 W [21,22,23]. These metrics highlight the efficiency of the design, and suggest that the digital
I&F neuron circuit, successfully implemented and verified on the Zynq MPSoC FPGA,
is well suited for integration into fully CMOS-based SNN chips. Also, although the
number of components making up the FPGA and the usage portions can differ, the implementation
design in this work can be adopted to other FPGA platforms with minor adjustments
such as clocking and resource optimization due to its modular structure.
Table 1. Summary of the resources used for the neuron implementation in comparison
with the available amount.
Resources
|
Utilization
|
Available
|
Utilization %
|
LUT
|
11
|
274,080
|
0.01
|
FF
|
33
|
548,160
|
0.01
|
IO
|
5
|
328
|
1.52
|
BUFG
|
1
|
404
|
0.25
|
PLL
|
1
|
8
|
12.50
|
Fig. 3. Power consumption per function in the digital equivalent I&F neuron circuit
for an SNN realized using an FPGA.
The successful implementation and validation of this digital neuron circuit underscore
its potential as a platform for neuromorphic computing applications. Its compact design,
combined with low power consumption and minimal resource utilization, makes it a promising
candidate for scaling into larger and more complex SNNs. Although there have been
substantial challenges in optimizing the bit widths, reusing the computational units,
and balancing efficiency and precision, trade-off relation has been precisely controlled
to obtain permissible precision and scalability, achieving resource and power efficiencies.
Replacing analog components introduces quantization and discretization effects due
to finite bit widths, slightly impacting precision. However, optimal bit-width selection
ensures functional accuracy and repeatability, which is advantageous for large-scale
systems. This study not only confirms the functionality of the I&F neuron model, but
also demonstrates its feasibility for future neuromorphic systems that require efficient
and scalable hardware solutions.
IV. CONCLUSION
In this study, we successfully designed, implemented, and verified the digital equivalent
circuit of an I&F neuron model using the FPGA technology. The functionality and performance
of the neuron circuit were validated, demonstrating that it is effective in replicating
biological neuron behaviors. Furthermore, the compact design and detailed power analysis
of individual functional blocks highlight the efficiency of the proposed circuit and
its strong potential for integration into hardware-oriented neural systems. A quantitative
evaluation of component usage and power consumption further confirms the viability
of the proposed circuit for scalable and energy-efficient applications in advanced
neural computing and its chip implementation.
ACKNOWLEDGMENTS
This work was supported by National R&D Program through the National Research
Foundation of Korea (NRF) funded by the Ministry of Science and ICT of Korea (MSIT)
under Grants NRF 2022M3I7A078936 and RS-2024-00402495.
References
C. Mead, ``Neuromorphic electronic systems,'' Proceedongs of IEEE, vol. 78, no. 10,
pp. 1629-1636, Oct. 1990.

C. D. Schuman, T. E. Potok, R. M. Patton, J. D. Birdwell, M. E. Dean, G. S. Rose,
and J. S. Plank, ``A survey of neuromorphic computing and neural networks in hardware,''
arXiv:1705.06963, 2017.

S. Furber, ``Large-scale neuromorphic computing systems,'' Journal of Neural Engineering,
vol. 13, no. 5, pp. 051001-1--051001-14, Aug. 2016.

D. Markovic, A. Mizrahi, D. Querlioz, and J. Grollier, ´ “Physics for neuromorphic
computing,” Nature Reviews Physics, vol. 2, pp. 499-510, Jul. 2020.

Q. T. Pham, ``A review of SNN implementation on FPGA,'' Proc. of 2021 International
Conference on Multimedia Analysis and Pattern Recognition (MAPR), pp. 1-6, Hanoi,
Vietnam, Oct. 15-16, 2021.

W. Maass, ``Networks of spiking neurons: The third generation of neural network models,''
Neural Networks, vol. 10, no. 9, pp. 1659-1671, Aug. 1997.

J.-Q. Yang, R. Wang, Y. Ren, J. Y. Mao, Z.-P. Wang, Y. Zhou, and S.-T. Han, ``Neuromorphic
engineering: From biological to spike-based hardware nervous systems,'' Advanced Materials,
vol. 32, no. 52, pp. 2003610-1--2003610-32, Dec. 2020.

S. Yang, P. Liu, J. Xue, R. Sun, and R. Ying, ``An efficient FPGA implementation of
Izhikevich neuron model,'' Proc. if 2020 International SoC Design Conference (ISOCC),
pp. 141-142, Yeosu, Korea, Oct. 21-24, 2020.

J. Han, Z. Li, W. Zheng, and Y. Zhang, ``Hardware implementation of spiking neural
networks on FPGA,'' Tsinghua Science and Technology, vol. 25, no. 4, pp. 479-486,
Jan. 2020.

S. Schmidgall, R. Ziaei, J. Achterberg, L Kirsch, S. P. Hajiseyedrazi, and J. Eshraghian,
``Brain-inspired learning in artificial neural networks: A review,'' APL Machine Learning,
vol. 2, no. 2, pp. 021501-1--021501-14, Jun. 2024.

A. N. Burkitt, ``A review of the integrate-and-fire neuron model: I. Homogeneous synaptic
input,'' Biological Cybernetics, vol. 95, no. 1, pp. 1-19, Apr. 2006.

M. Kwon, M. Baek, S. Hwang, K. Park, T. Jang, T. Kim, J. Lee, S. Cho, and B.-G. Park,
``Integrate-and-fire neuron circuit using positive feedback field effect transistor
for low power operation,'' Journal of Applied Physics, vol. 124, no. 15, pp. 152107-1--152107-7,
Sep. 2018.

G. Indiveri, E. Chicca, and R. Douglas, ``A VLSI array of low-power spiking neurons
and bistable synapses with spike-timing dependent plasticity,'' IEEE Transactions
on Neural Networks, vol. 17, no. 1, pp. 211-221, Jan. 2006.

J. Schemmel, D. Brüderle, A. Grübl, M. Hock, K. Meier, and S. Millner, “A wafer-scale
neuromorphic hardware sys- tem for large-scale neural modeling,” Proc. of 2010 IEEE
International Symposium on Circuits and Systems (ISCAS), pp. 1947-1950, Paris, France,
May 30-Jun. 2, 2010.

J. Seo, B. Brezzo, Y. Liu, B. D. Parker, S. K. Esser, and R. K. Montoye, ``A 45 nm
CMOS neuromorphic chip with a scalable architecture for learning in networks of spiking
neurons,'' Proc. of IEEE Custom Integrated Circuits Conference (CICC), pp. 1-4, San
Jose, USA, Sep. 19-21, 2011.

M. Pfeiffer and T. Pfeil, ``Deep learning with spiking neurons: Opportunities and
challenges,'' Frontiers in Neuroscience, vol. 12, pp. 774-1--774-18, Oct. 2018.

H. Taherdoost, ``Deep Learning and Neural Networks: Decision-Making Implications,''
Symmetry, vol. 15, no. 9, pp. 1723-1--1723-22, Sep. 2023.

P. Panda, S. A. Aketi, and K. Roy, ``Toward scalable, efficient, and accurate deep
spiking neural networks with backward residual connections, stochastic softmax, and
hybridization,'' Frontiers in Neuroscience, vol. 14, pp. 653-1--653-18, Jun. 2020.

G. Indiveri and T. K. Horiuchi, ``Frontiers in Neuromorphic Engineering,'' Frontiers
in Neuroscience, vol. 5, pp. 118-1--118-2, Oct. 2011.

G. Indiveri, B. Linares-Barranco, R. Legenstein, G. Deligeorgis, and T. Prodromakis,
``Integration of nanoscale memristor synapses in neuromorphic computing architectures,''
Nanotechnology, vol. 24, no. 38, pp. 384010-1--384010-13, Sep. 2013.

H. Fang, Z. Mei, A. Shrestha, Z. Zhao, Y. Li, and Q. Qiu, ``Encoding, model, and architecture:
Systematic optimization for spiking neural network in FPGAs,'' Proc. of 2020 IEEE
ACM International Conference on Computer-Aided Design (ICCAD), pp. 1-9, Virtual, Nov.
2-5, 2020.

X. Ju, B. Fang, R. Yan, X. Xu, and H. Tang, ``An FPGA implementation of deep spiking
neural networks for low-power and fast classification,'' Neural Computation, vol.
32, no. 1, pp. 182-204, Jan. 2020.

S. Yang, J. Wang, B. Deng, C. Liu, H. Li, C. Fietkiewicz, and K. A. Loparo, ``Real-time
neuromorphic system for large-scale conductance-based spiking neural networks,'' IEEE
Transactions on Cybernetics, vol. 49, no. 7, pp. 2490-2503, Apr. 2018.

Yeji Lee received the B.S. degree in electrical engineering from Hankyoung National
University, Anseong, Korea, in 2024. She is currently pursuing an M.S. degree at Ewha
Womans University. Her current research interests include resistive-switching random-access
memory (RRAM), synaptic devicese for neuromorphic systems, low-power neuron circuits
and CMOS integrated fabrication. She is a Student Member at the Institute of Electronics
and Information Engineers (IEIE) of Korea.
Arati Kumari Shah received her B.Tech. and M.Tech. degrees in electronics and communication
engineering from the North Eastern Regional Institute of Science and Technology, Arunachal
Pradesh, India, in 2017 and 2019, respectively. She received her Ph.D. degree at Gachon
University, Seongnam, South Korea, in 2024. She also worked as a Postdoctoral Researcher
at Ewha Womans University, South Korea, where she continued her research on CMOS technologies
and integrated circuits. Dr. Shah is currently working as a Senior Research Engineer
at Philophos Inc., South Korea. She received the Gold Medal for the highest score
during her master’s degree. Her research interests include memory devices, neuron
circuits for spiking neural networks, and circuit designs optimized for various synapse
arrays.
Myounggon Kang received the Ph.D. degree from the Department of Electrical and
Computer Engineering, Seoul National University, Seoul, Korea, in 2012. From 2005
to 2015, he worked as a Senior Engineer at Flash Design Team of Samsung Electronics
Company. In 2015, he joined Korea National University of Transportation as a Professor
of the Department of Electronics Engineering. His current research interests are CMOS
device modeling and memory circuit design. He has been working as a Professor at Department
of Intelligent Semiconductor Engineering, School of Advanced Fusion Studies, University
of Seoul, Seoul, Korea, since 2023.
Seongjae Cho received his B.S. and Ph.D. degrees in electrical engineering from
Seoul National University, Seoul, Korea, in 2004 and 2010, respectively. He worked
as an Exchange Researcher at the National Institute of Advanced Industrial Science
and Technology (AIST), Tsukuba, Japan, in 2009. He worked as a Postdoctoral Researcher
at Seoul National University in 2010 and at Stanford University, Palo Alto, CA, from
2010 to 2013. Also, he worked as a faculty member at the Department of Electronic
Engineering, Gachon University, from 2013 to 2023. He is currently working as an Associate
Professor at the Division of Convergence Electronic and Semiconductor Engineering,
Ewha Womans University, Seoul, Korea from 2023. His current interests include emerging
memory deviecs, advanced nanoscale CMOS devices, optical interconnect devices, and
novel devices for future computing.