DS-CIM: Digital Stochastic Computing-In-Memory Featuring Accurate OR-Accumulation via Sample Region Remapping for Edge AI Models

Jan 10, 2026 by Kunming Shao, Liang Zhao, Jiangnan Yu, Zhipeng Liao, Xiaomeng Wang, Yi Zou, Tim Kwang-Ting Cheng, Chi-Ying Tsui (arXiv.org)

DOI 10.48550/arXiv.2601.06724

We built DS-CIM, a digital stochastic compute-in-memory fabric that marries SC's simplicity with DCIM throughput by turning signed MAC into an unsigned OR-based kernel and killing OR collisions with a 2D-partitioned shared PRNG and sample-region remapping. The result is a tiny, massively replicated circuit that fixes 1s saturation and delivers near-floating accuracy on ResNet models while pushing TOPS/W and TOPS/mm^2 through the roof — real energy- and area-win hardware for edge AI.

source S2, openalex