A Unified Evaluation Framework for Spiking Neural Network Hardware Accelerators Based on Emerging Non-Volatile Memory Devices

TL;DR

This paper introduces a comprehensive evaluation framework for assessing the performance and energy efficiency of Spiking Neural Network hardware accelerators utilizing emerging non-volatile memory devices, with a focus on magnetic skyrmion-based components.

Contribution

It presents a unified methodology for evaluating hardware metrics of SNNs, addressing a gap in current research, and demonstrates its application using skyrmionic devices with high accuracy and low energy consumption.

Findings

Achieved approximately 98% inference accuracy on MNIST.

Demonstrated energy consumption in the picojoule range.

Provided a versatile framework applicable to various emerging devices.

Abstract

Spiking Neural Networks (SNNs) have emerged as a promising paradigm, offering event-driven and energy-efficient computation. In recent studies, various devices tailored for SNN synapses and neurons have been proposed, leveraging the unique characteristics of emerging non-volatile memory (eNVM) technologies. While substantial progress has been made in exploring the capabilities of SNNs and designing dedicated hardware components, there exists a critical gap in establishing a unified approach for evaluating hardware-level metrics. Specifically, metrics such as latency, and energy consumption, are pivotal in assessing the practical viability and efficiency of the constructed neural network. In this article, we address this gap by presenting a comprehensive framework for evaluating hardware-level metrics in SNNs based on non-volatile memory devices. We systematically analyze the impact of synaptic and neuronal components on energy consumption providing a unified perspective for assessing the overall efficiency of the network. In this study, our emphasis lies on the neuron and synaptic device based on magnetic skyrmions. Nevertheless, our framework is versatile enough to encompass other emerging devices as well. Utilizing our proposed skyrmionic devices, the constructed SNN demonstrates an inference accuracy of approximately 98% and achieves energy consumption on the order of pJ when processing the Modified National Institute of Standards and Technology (MNIST) handwritten digit dataset.