NANQ Quantization Boosts Analog AI Compute-in-Memory Accuracy.

Analog compute-in-memory (CIM) offers significant energy efficiency for neural network inference, but its performance can be severely hampered by device variations and read noise, especially with low-bit quantized models. Existing quantization methods for CIM primarily focus on minimizing ideal quantization error, often overlooking the practical limitations imposed by hardware noise. This oversight leads to inefficient precision allocation. To address this, researchers propose NANQ, a novel noise-aware mixed-precision non-uniform quantization framework specifically designed for analog CIM. NANQ models the magnitude-dependent weight noise observed in eFlash CIM arrays and translates this noise profile into an adaptive quantization density. This allows for finer resolution in low-noise regions while avoiding wasted precision in noise-dominated areas. Furthermore, NANQ assigns layer-wise bit-widths by identifying each layer's precision saturation point under hardware noise. On-chip experiments demonstrate that NANQ significantly improves vision model accuracy by 8.05 percentage points and reduces language model perplexity by 54.7% on average compared to other methods, achieving these gains with only 3.2-3.8 equivalent bits.