Neuromorphic Computing with Phase-Change Memory Arrays for Ultra-Low-Power Edge AI Inference

摘要

Edge AI inference on battery-powered devices demands computing efficiency orders of magnitude beyond what von Neumann architectures can provide. We present NeuroPhase, a neuromorphic inference accelerator based on 256 × 256 phase-change memory (PCM) crossbar arrays that performs analog matrix-vector multiplication in-memory, eliminating the data movement bottleneck. NeuroPhase achieves 12.4 TOPS/W (tera-operations per second per watt) on ResNet-50 inference — 28× more energy-efficient than state-of-the-art digital accelerators — while maintaining 97.1% of the baseline FP32 accuracy through a hardware-aware quantization and drift compensation scheme. A 28 nm prototype chip consuming 8.3 mW classifies ImageNet images at 142 frames/second, enabling continuous visual AI on coin-cell batteries for over 1 year.