Design Limits of In-Memory Computing: Beyond the Crossbar

Abstract

Resistive random-access memory (RRAM)-based in-memory computing (IMC) architecture offers an energy-efficient solution for DNN acceleration. Yet, its performance is limited by device non-idealities, circuit precision, on-chip interconnection, and algorithm properties. Based on statistical data from a fully-integrated 65nm CMOS/RRAM test chip and a cross-layer simulation framework, we show that the IMC system's real bottleneck is not the RRAM device but the analog-to-digital converter (ADC) precision and the stability of DNN models. The results are summarized into a roofline model and demonstrated on CIFAR-10, SVHN, CIFAR-100, and ImageNet, helping understand RRAM-based IMC architectures' design limits.