Papernews
← back

OpenACM: An Open-Source SRAM-Based Approximate CiM Compiler

Jan 16, 2026 by Yiqi Zhou, Junhao Ma, Xingyang Li, Yule Sheng, Yue Yuan, Yikai Wang, Bochang Wang, Yiheng Wu, Shan Shen, Wei W. Xing, Daying Sun, Li Li, Zhiqiang Xiao (arXiv.org)

DOI 10.48550/arXiv.2601.11292



We built OpenACM, the first open-source, accuracy-aware compiler that maps AI workloads onto SRAM-based approximate compute-in-memory, pairing a library of tunable approximate multipliers with a transistor-level, variation-aware SRAM macro and a full OpenROAD/FreePDK45 physical flow for reproducible design exploration. Play with fine-grained accuracy-energy tradeoffs across CNNs and save up to 64% energy with negligible accuracy loss — source and flow are on GitHub.

source S2, openalex



dgfl, 2026