Cryogenic Analog 1T-ReRAM with Enhanced Dynamic Range and Suppressed Noise for Cold Neural Networks

Abstract

We present the first cryogenic characterization of read noise in 14 nm CMOS compatible analog resistive RAMs (ReRAMs) and evaluate the efficiency of analog in-memory (AIM) neural network (NN) training at 77 K using the optimized Tiki-Taka algorithm (TTv2). Compared to standard room temperature operation, cryogenic operation suppresses the read noise by an exceptional 88% and improves the analog dynamic range by 2200% owing to reduced stochasticity and improved heat confinement in the conductive metal oxide layer. The effectiveness of analog cryo-ReRAMs in training NNs is validated by simulations using TTv2 on handwritten digits yielding an accuracy of 96.5%, comparable to the floating-point baseline and the highest reported to date for non-volatile memories (NVMs) at cryogenic temperatures. The results highlight the potential for cryogenic ReRAM technology in power-constrained applications such as quantum computing.