Fault Analysis for a MTJ-based Spiking Neural Network
Résumé
Interest in Spiking Neural Networks (SNN), which mirror brain functionality, within the Artificial Intelligence (AI) domain stems from their potential for energy efficiency. Recent advancements have introduced novel synapse and neuronimplementation devices, promising further energy efficiency enhancements. However, the susceptibility of hardware-based SNNs to manufacturing defects raises concerns. This study investigates the impact of manufacturing defects on SNN performance, leveraging Magnetic Tunnel Junction (MTJ) devices for synapse and neuron implementation. Circuit-level simulations are conducted by building a basic SNN and intentionally introducing defects. Input spikes are applied to the SNN to evaluate its performance in the presence of defects. Furthermore, our investigation delves into the effects of timing alterations during integration and the synapse operation’s leakage windows on defective SNNs. The results from our analysis offer valuable guidance for formulating fault models that generate high-quality test vectors.
Origine : Fichiers produits par l'(les) auteur(s)