arxivcs.ARcs.LGcs.OS2026-07-16
ExaGEMM: Exploration Framework for CPU-Driven ML Inference via Associative In-Register Computing for Low-Bit GEMM
Hyunwoo Oh, Suyeon Jang, Hanning Chen, Sanggeon Yun, Ryozo Masukawa, Mohsen Imani
Low-bit GEMM is increasingly central to efficient ML inference, yet very-low-bit execution remains a poor fit for conventional CPUs. Practical deployment spans fragmented regimes-from 1/2/4-bit weights to varying activation precision-whose feasibility, reuse opportunity, and supp…