Abstract

A spiking neural network (SNN) is a computational model comprised of spiking neurons, inspired by the functionality of the human brain. These networks utilize discrete events called spikes to process information, enabling them to capture the dynamic temporal patterns inherent in data and facilitating efficient event-driven computation. This enables the development of low-power NNs, especially when coupled with bio-plausible local spike-timing-dependent plasticity (STDP) learning algorithms that are capable of encoding temporal information for complex machine learning tasks. The hardware implementation of STDP utilizing spintronic devices offers low-power solutions to enable efficient and high-performance neuromorphic computing tasks. However, these devices suffer from non-idealities associated with factors such as process variations, material properties, fabrication defects, and environmental conditions, which can degrade their performance. This work presents the impact of device-level variations in STDP using unsupervised learning techniques. The variability analysis incorporates variations in the thickness of the oxide layer, tunnel magnetoresistance, saturation magnetization, and damping of the free layer. Furthermore, the effect of the number of excitatory neurons on the classification accuracy of handwritten digits in the MNIST dataset is investigated. The results show that the accuracy of SNN drops by 9% due to device variations. Understanding the impact of device variations is crucial for optimizing the efficiency and reliability of spintronic-based SNNs, thereby advancing the field of neuromorphic computing.