Presentation
In-Memory ADC-Based Nonlinear Activation Quantization for Efficient In-Memory Computing
DescriptionThis paper presents Boundary Suppressed K-Means Quantization (BS-KMQ), a nonlinear (NL) quantization method to reduce analog-to-digital converter (ADC) resolution in the in-memory computing (IMC) systems. ReLU and clamping often cause value accumulation near distribution edges, leading to biased clustering and suboptimal quantization. BS-KMQ mitigates this by removing such outliers before clustering, yielding more informative quantization levels. It achieves at least 3X lower quantization error compared to linear, Lloyd–Max, CDF and K-means methods. The resulting NL references are implemented using a reconfigurable in-memory NL-ADC with 7X area improvements compared to previous two works. Evaluated on ResNet-18, VGG-16, Inception-V3, and DistilBERT, BS-KMQ improves up to 66.8%, 25.4%, 66.6%, and 67.7% higher post‑training quantization accuracy compared to linear quantization. After low-bit finetuning, it maintains competitive accuracy with significantly fewer ADC levels (3/3/4/4b). System-level simulation on ResNet-18 (6/2/3b) shows up to 4X speedup and 24X energy efficiency over existing IMC accelerators.
Event Type
Research Manuscript
TimeWednesday, July 292:34pm - 2:47pm PDT
LocationMtg Room 203AB