Overcoming Crossbar Nonidealities in Binary Neural Networks Through Learning

Abstract

The crossbar nonidealaties may considerably degrade the accuracy of matrix multiplication operation, which is the cornerstone of hardware accelerated neural networks. In this paper, we show that the crossbar nonidealities especially the wire resistance should be taken into consideration for accurate evaluation. We also present a simple yet highly effective way to capture the wire resistance effect for the inference and training of deep neural networks without extensive SPICE simulations. Different scenarios have been studied and used to show the efficacy of our proposed method.