Impact of Circuit Limit and Device Noise on RRAM Based Conditional Generative Adversarial Network

Abstract

In this work, a Conditional Generative Adversarial Network (CGAN) [1] is demonstrated based on the Resistive Random Access Memory (RRAM). During training, the read noise of RRAM is utilized as a random bias source to enrich the diversity of the generator in CGAN. Further, we evaluate the impact of both read noise (RRAM as weight storage cell) and the resolution of the AD/DA circuit on the performance of CGAN through a comprehensive simulation.