Memristor Crossbar Scaling Limits and the Implementation of Large Neural Networks

Abstract

Memristor crossbar-based neural networks perform parallel operation in the analog domain. Ex-situ training approach needs to program the predetermined resistance values to the memristor crossbar. Because of the stochasticity of the memristor devices, programming a memristor needs to read the device resistance value iteratively. Reading a single memristor in a crossbar (without isolation transistor) is challenging due to the sneak path current. Programming a memristor in a crossbar to either RON or ROFF state is relatively straightforward. A neural network implemented using higher precision weights provides higher classification accuracy compared to a Ternary Neural Network (TNN). This paper demonstrates the implementation of memristor-based neural networks using only the two resistance values (RON, ROFF). We have considered the crossbar scaling limits and proposed a novel technique to implement a large neural network using multiple smaller crossbar arrays. We have proposed novel neuron circuits to achieve higher weight precision. Our experimental result shows that the proposed higher precision synapses are easy to program and provide better classification accuracy compared to a TNN. Proposed technique of implementing a large neural network on memristor crossbar circuits has a slight loss in the classification accuracy compared to the software implementation. But the memristor-based implementation uses only 51.7% of the synapses used in the software implementation.