Impact of memristor switching noise in a neuromorphic crossbar

Abstract

Many existing memristor models have a direct relationship between resistance change and the voltage pulse applied. However, this results in a memristor model that can be tuned nearly to a floating point value if a small enough voltage pulse is applied. This paper discusses how noise can be added to the dynamic resistive switching component of a memristor model in SPICE. The proposed memristor model has a tunable degree of stochastic behavior during switching. Therefore, each time an identical voltage pulse is applied to a memristor device, a varying amount of resistance change will occur. This provides a much more realistic model of memristor behavior. Furthermore, this model is used in a neuromorphic circuit simulation to show that stochastic memristor devices can be trained according to a learning algorithm. The amount of switching noise in the memristors was varied to see what impact this may have on a neuromorphic circuit.