SpikeSim: An end-to-end Compute-in-Memory Hardware Evaluation Tool for Benchmarking Spiking Neural Networks

Abhishek Moitra, Abhiroop Bhattacharjee, Runcong Kuang, Gokul Krishnan, Yu Cao, Priyadarshini Panda

Abstract

SNNs are an active research domain towards energy efficient machine intelligence. Compared to conventional ANNs, SNNs use temporal spike data and bio-plausible neuronal activation functions such as Leaky-Integrate Fire/Integrate Fire (LIF/IF) for data processing. However, SNNs incur significant dot-product operations causing high memory and computation overhead in standard von-Neumann computing platforms. Today, In-Memory Computing (IMC) architectures have been proposed to alleviate the "memory-wall bottleneck" prevalent in von-Neumann architectures. Although recent works have proposed IMC-based SNN hardware accelerators, the following have been overlooked- 1) the adverse effects of crossbar non-ideality on SNN performance due to repeated analog dot-product operations over multiple time-steps, 2) hardware overheads of essential SNN-specific components such as the LIF/IF and data communication modules. To this end, we propose SpikeSim, a tool that can perform realistic performance, energy, latency and area evaluation of IMC-mapped SNNs. SpikeSim consists of a practical monolithic IMC architecture called SpikeFlow for mapping SNNs. Additionally, the non-ideality computation engine (NICE) and energy-latency-area (ELA) engine performs hardware-realistic evaluation of SpikeFlow-mapped SNNs. Based on 65nm CMOS implementation and experiments on CIFAR10, CIFAR100 and TinyImagenet datasets, we find that the LIF/IF neuronal module has significant area contribution (>11% of the total hardware area). We propose SNN topological modifications leading to 1.24x and 10x reduction in the neuronal module's area and the overall energy-delay-product value, respectively. Furthermore, in this work, we perform a holistic comparison between IMC implemented ANN and SNNs and conclude that lower number of time-steps are the key to achieve higher throughput and energy-efficiency for SNNs compared to 4-bit ANNs.

Introduction

Spiking Neural Networks (SNNs) are hailed as the energy-efficient future of AI because they process data using sparse, brain-like "spikes" over time. However, running them on traditional von-Neumann hardware leads to a significant "memory wall" bottleneck. While In-Memory Computing (IMC) using analog crossbars is a promising solution, existing simulators often overlook SNN-specific requirements like temporal processing and Leaky-Integrate-Fire (LIF) neurons. Furthermore, they fail to model how analog hardware non-idealities accumulate over multiple time-steps. To address these gaps, we propose SpikeSim, a tool for realistic performance, energy, latency, and area evaluation of IMC-mapped SNNs.

Methodology: The SpikeSim Framework

SpikeSim provides an end-to-end pipeline to bridge the gap between high-level SNN algorithms and physical hardware constraints.

Overview of the SpikeSim tool flow, integrating SNN modeling with hardware-realistic evaluation engines.

The framework consists of three primary components:

  1. SpikeFlow Mapping: A monolithic IMC architecture that maps SNNs onto tiled crossbars. It features a novel, fully digital DIFF module that handles signed operations without the area-heavy dual-crossbar approach.
  2. NICE (Non-Ideality Computation Engine): This engine performs circuit-level analysis to model resistive parasitics and device variations. It introduces NI-aware weight encoding to maximize high-resistance states, improving robustness against hardware noise.
  3. ELA (Energy-Latency-Area) Engine: A Python-based tool that tracks physical costs, including Network-on-Chip (NoC) overhead and neuronal storage.

Hardware Architecture: SpikeFlow

The SpikeFlow architecture is designed for efficient SNN mapping, utilizing a hierarchical structure to manage data flow and computation.

The hierarchical SpikeFlow architecture, showing the organization from Network-on-Chip (NoC) interconnects down to individual IMC crossbars.

As shown in the architecture diagram, the system is organized into tiles connected via a NoC. Each tile contains multiple Processing Elements (PEs), which in turn house clusters of IMC crossbars. This hierarchy allows for scalable deployment of large-scale SNN models like VGG-style architectures.

Experimental Results and Discussion

The tool was evaluated using a 65nm CMOS implementation on standard computer vision benchmarks: CIFAR-10, CIFAR-100, and TinyImageNet.

Key Insights

  • The Neuronal Area Bottleneck: Analysis revealed that the neuronal module (LIF/IF) consumes 11–30% of the total chip area due to the need to store membrane potentials across time-steps. Topological modifications, such as channel scaling, reduced this area by 1.24×.
  • Noise Accumulation: SNNs perform repeated analog operations over time, making them sensitive to error accumulation. SpikeSim's NI-aware encoding improved accuracy by over 70% for CIFAR-10 compared to vanilla encoding.
  • Topological Optimization: By reducing output channels in the first convolutional layer, we achieved a 10× reduction in Energy-Delay Product (EDP) with minimal accuracy loss.

Performance Comparison

At low time-steps (T=3 to 5), SNNs on IMC hardware can achieve better energy efficiency and throughput than 4-bit quantized ANNs. While SNNs have a significantly larger neuronal module area (roughly 1000× higher than ANNs), they remain competitive in overall performance when latency is minimized.

Conclusion

SpikeSim demonstrates that while SNNs offer massive potential for energy efficiency, the "memory tax" of neuronal states is a significant bottleneck. Future work should focus on IMC-aware SNN training algorithms that optimize for low-latency spikes and hardware-specific dataflows to further mitigate these overheads.

Summary updated

Abstract

SNNs are an active research domain towards energy efficient machine intelligence. Compared to conventional ANNs, SNNs use temporal spike data and bio-plausible neuronal activation functions such as Leaky-Integrate Fire/Integrate Fire (LIF/IF) for data processing. However, SNNs incur significant dot-product operations causing high memory and computation overhead in standard von-Neumann computing platforms. Today, In-Memory Computing (IMC) architectures have been proposed to alleviate the "memory-wall bottleneck" prevalent in von-Neumann architectures. Although recent works have proposed IMC-based SNN hardware accelerators, the following have been overlooked- 1) the adverse effects of crossbar non-ideality on SNN performance due to repeated analog dot-product operations over multiple time-steps, 2) hardware overheads of essential SNN-specific components such as the LIF/IF and data communication modules. To this end, we propose SpikeSim, a tool that can perform realistic performance, energy, latency and area evaluation of IMC-mapped SNNs. SpikeSim consists of a practical monolithic IMC architecture called SpikeFlow for mapping SNNs. Additionally, the non-ideality computation engine (NICE) and energy-latency-area (ELA) engine performs hardware-realistic evaluation of SpikeFlow-mapped SNNs. Based on 65nm CMOS implementation and experiments on CIFAR10, CIFAR100 and TinyImagenet datasets, we find that the LIF/IF neuronal module has significant area contribution (>11% of the total hardware area). We propose SNN topological modifications leading to 1.24x and 10x reduction in the neuronal module's area and the overall energy-delay-product value, respectively. Furthermore, in this work, we perform a holistic comparison between IMC implemented ANN and SNNs and conclude that lower number of time-steps are the key to achieve higher throughput and energy-efficiency for SNNs compared to 4-bit ANNs.
BibTeX
@article{moitra2023spikesim,
  author = {Abhishek Moitra and Abhiroop Bhattacharjee and Runcong Kuang and Gokul Krishnan and Yu Cao and Priyadarshini Panda},
  title = {{SpikeSim: An end-to-end Compute-in-Memory Hardware Evaluation Tool for Benchmarking Spiking Neural Networks}},
  journal = {IEEE Trans. Comput. Aided Des. Integr. Circuits Syst.},
  year = {2023},
  doi = {10.1109/tcad.2023.3274918},
}
Figure