Memristor Crossbar Scaling Limits and the Implementation of a Large Neural Network Using 3D Stacked Crossbars

Abstract

Memristor crossbar-based neural networks perform parallel operations in the analog domain. In ex-situ training, predetermined resistance values are programmed into the memristor crossbar. Due to the stochastic nature of memristor devices, programming a memristor in a crossbar needs to read the device resistance value iteratively. However, reading an individual memristor in a crossbar, especially without an isolation transistor, is challenging because of sneak path currents. Programming a memristor to either the RON or ROFF state is relatively straight-forward. Neural networks that use higher-precision weights typically achieve better classification accuracy than Ternary Neural Networks (TNNs). This paper presents a memristor-based neural network implementation that uses only the two resistance states (RON, ROFF). We have examined the impact of the device RON/ROFF ratio and driver size on the scalability of memristor-based neural network circuits. A large neural network is implemented using multiple smaller 3Dstacked crossbar arrays. Additionally, we have proposed novel neuron circuits to support higher weight precision. Experimental results show that the proposed high-precision synapses are easy to program and offer improved classification accuracy compared to a TNN.