Evaluation of a Ferroelectric Capacitor-based Compute-in-memory Framework

Abstract

Non-volatile memory (NVM) based compute-in-memory (CiM) architecture has evolved as a promising energy-efficient solution for mitigating the memory wall problem, widely explored for machine learning applications. Among various emerging NVM technologies, ferroelectric (FE) capacitor (FeCAP) based crossbar with a non-destructive read operation promises further improvement in terms of both energy and overall chip-area. However, comprehensive performance analysis for such FeCAP-based CiM multi-tile architecture is not well explored, specifically for large neural-network (NN) workloads, requiring large cross-bar structures to map the weights for inference. Therefore, in this work, we evaluate FeCAP-based cross-bar performance and benchmark it against other NVMs such as RRAM and STT-MRAM using a multi-tile simulator framework, NeuroSim, for diverse NN workloads like VGG8, ResNet18, and ResNet50. Results show that compared to other NVMs, the FeCAP-based multi-tile architecture offers at least 1.55x, 1.67x, and 1.7x lower total inference energy along with a high energy-efficiency for VGG8, ResNet18, and ResNet50 NN workloads, respectively.