Exploring the Power – Prediction Accuracy Trade-Off in a Deep Learning Neural Network using Wide Compliance RRAM Device

Abstract

In this work, the quantitative impact of variability in the low and high resistance state distributions of Hafnium oxide based RRAM on the prediction accuracy of deep learning neural networks is explored over a wide range of current compliance ranging from 2 to 500micro Ampere. The device power versus prediction accuracy trade-off trend is examined for such a wide range of compliance for the first time. The weights of one of the layers of the convolutional neural network (CNN) are represented by the floating point binary representation where the binary bits are configured using the RRAM resistance distribution data on an AlexNet platform.