Bit-sliced quantization for robust memristor-based compute-in-memory

Abstract The escalating computational demand of modern AI is increasingly constrained by the data-movement overhead of von Neumann architectures, motivating memristor-based compute-in-memory (CIM) as an energy-efficient alternative. Yet extending CIM to reliable analog computation remains difficult because practical memristor arrays exhibit limited conductance resolution and stochastic device fluctuations, causing severe performance loss when conventional uniform quantization is directly deployed. Here we present bit-sliced quantization (BSQ), a hardware–software co-design framework for robust analog CIM inference. BSQ combines distribution-aware non-uniform post-training quantization, hardware-aligned noise injection, and crossbar-compatible bit-slice mapping onto stable binary device states. This design aligns model representation with device physics while avoiding costly retraining. We evaluate BSQ on three tasks with increasing complexity: handwritten digit classification, sentiment analysis, and large language model reasoning. Under programming noise, BSQ improves accuracy by +16.0% on 2-bit digit classification, +0.5% on 3-bit sentiment classification, and up to +20.08% on 3-bit LLM reasoning benchmarks. These results show that physically aligned quantization is critical for dependable analog inference, and establish BSQ as a practical route toward robust and scalable edge AI on memristor CIM hardware.