Noise and Drift Models for Protonic ECRAM and Their Impact on Analog In‐Memory Computing

Abstract

Analog in‐memory computing (AIMC) with nonvolatile memories (NVMs) offers a promising route toward energy‐efficient deep learning. However, because AIMC relies on analog device states, understanding device‐level nonidealities, particularly noise and drift, is essential for assessing achievable computing performance. Among NVM candidates, the electrochemical random‐access memory (ECRAM) is attractive for AIMC owing to its symmetric switching, large number of accessible conductance states, low energy consumption, and small cycle‐to‐cycle variability. However, the physical origins of ECRAM read noise and conductance drift remain insufficiently understood. Here, we systematically investigate noise and drift in ECRAM devices, identify their dominant mechanisms, and develop physics‐based statistical models. Based on measured device characteristics, we construct compact device models for hardware‐aware simulations to evaluate ECRAM performance in AIMC. We show that the low noise of ECRAM enables high inference accuracy even in the low‐conductance regime with a small memory window. This feature offers a path toward improved energy efficiency and reduced leakage currents without sacrificing computational fidelity in high‐performance AIMC applications.