OS-CIM: A 28-nm Output-Stationary SRAM Compute-in-Memory Macro Featuring Dynamic Look-Ahead Zero Weight Skipping and Runtime Partial Sum Quantization

View at publisher Summarise with VIRAT
Authors: Xiaofeng Hu, HanGyeol Mun, Yifan He, Jian Meng, Yuan Liao, Jae-sun Seo
Cited by: 0
References: 47

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Static random access memory (SRAM)-based compute-in-memory (CIM) designs have shown impressive energy efficiency for vector–matrix multiplications of artificial intelligence (AI) workloads. Conventional SRAM CIM macros naturally adopt the weight-stationary (WS) dataflow. While the WS-CIM maximizes the on-chip weight reuse, the expensive global partial sum (Psum) read/write operations largely degrade the WS-CIM system energy efficiency. In addition, the activation sharing and static weight mapping in WS-CIM make zero-weight skipping infeasible. To address these challenges, this work introduces OS-CIM, an output stationary SRAM-based CIM macro and system. The OS-CIM features an OS-CIM structure with a custom single-cycle read-and-write 8T SRAM cell to store the Psum in the CIM macro and perform the accumulation locally, eliminating the expensive Psum access to SRAM buffers, a dynamic look-ahead weight skipping (DLS) scheme, well aligned with OS dataflow, which achieves high latency reduction across various weight sparsity patterns and significantly improves workload imbalance, and a runtime Psum quantization (RPQ) method to mitigate the hardware cost of high-precision accumulation. The fabricated 28-nm OS-CIM macro shows 20.9–137.2 TOPS/W peak energy efficiency at 0.45 V across dense to 2:16 weight sparsity and INT8 precision for inference of ViT-Base model. Compared to state-of-the-art DCIM works, OS-CIM achieves$1.29\times $improvement on macro-level energy efficiency,$1.39\times $on system-level energy efficiency.

Fields of Study & Subjects

Advanced Memory and Neural Computing Low-power high-performance VLSI design Parallel Computing and Optimization Techniques

Publication details

Journal / source
IEEE Journal of Solid-State Circuits
Volume
61
Issue
8
Pages
4489–4502
Publication date
25 March 2026
Language
EN