[C100] Dataflow-Preserving, Overhead-Free Weight Remapping for Fault-Tolerant ReRAM-Based In-Memory Computing

Abstract

Resistive random-access memory (ReRAM)-based in-memory computing (IMC) systems offer high energy efficiency and storage density for deep neural network (DNN) acceleration. However, ReRAM devices suffer from various non-idealities, with stuck-at faults (SAFs) being a major source of reliability degradation. Weight remapping (WR) is a widely adopted fault mitigation technique, yet existing WR approaches either neglect dataflow consistency or rely on additional hardware support to restore dataflow, which results in non-negligible hardware overhead. To address these challenges, we propose FREEMAP, an overhead-free WR algorithm that preserves dataflow consistency without any runtime operations or hardware modifications. FREEMAP integrates two key components: layer-wise filter reordering (LFR), which reorders filters across an entire layer to enhance fault resilience, and row group remapping (RGR), which realigns the next layer’s weight rows to match the reordered outputs. Experimental results across diverse models and datasets demonstrate that FREEMAP eliminates the substantial hardware overhead of conventional WR, reducing area and energy consumption to 0.02x - 0.06x and 0.04x - 0.14x, respectively. Compared to the state-of-the-art dataflow-aware WR method, FREEMAP further reduces area and energy to 0.45x - 0.47x and 0.52x - 0.63x, respectively, while maintaining comparable accuracy.