Impact of RTN and Variability on RRAM-Based Neural Network

Abstract

Resistive switching memory devices can be categorized into filamentary RRAM or non-filamentary RRAM depending on the switching mechanisms. Both types of RRAM devices have been studied as novel synaptic devices in hardware neural networks. In this work, we analyze the amplitude of Random Telegraph Noise (RTN) and program-induced variabilities in both TaOx/Ta2Os filamentary and TiO2/a-Si (a-VMCO) non-filamentary RRAM devices and evaluate their impact on the pattern recognition accuracy of neural networks. It is revealed that the non-filamentary RRAM has a tighter RTN amplitude distribution than its filamentary counterpart, and also has much lower programed-induced variability, which lead to much smaller impact on the recognition accuracy, making it a promising candidate in synaptic application.