Temperature-Resilient RRAM-Based In-Memory Computing for DNN Inference

Resistive random access memory (RRAM)-based in-memory computing (IMC) has emerged as a promising paradigm for efficient deep neural network (DNN) acceleration. However, the multibit RRAMs often suffer from nonideal characteristics such as drift and retention failure against temperature changes, leading to significant inference accuracy degradation. In this article, we present a new temperature-resilient RRAM-based IMC scheme for reliable DNN inference hardware. From a 90-nm RRAM prototype chip, we first measure the retention characteristics of multilevel HfO$\mathbf {_2}$2 RRAMs at various temperatures up to 120$^{\circ }$∘C, and then rigorously model the temperature-dependent RRAM retention behavior. We propose a novel and efficient DNN training/inference scheme along with the system-level hardware design to resolve the temperature-dependent retention issues with one-time DNN deployment. Employing the proposed scheme on a 256×256 RRAM array with the circuit-level benchmark simulator NeuroSim, we demonstrate robust RRAM IMC-based DNN inference where $>$>30% CIFAR-10 accuracy and $>$>60% TinyImageNet accuracy are recovered against temperature variations.