COMET: Communication and Memory Co-Design for Fine-Grained AI Inference in MCM Accelerators

Jan 31, 2026 by Taishu Sheng, Guangyu Sun, Dezun Dong (International Symposium on High-Performance Computer Architecture)

DOI 10.1109/HPCA68181.2026.11408536

We built COMET, a chiplet-aware communication and memory co-design that tames the tiny, bursty DMA traffic AI accelerators actually generate, dynamically aggregating requests and re-mapping on-chip memory so chiplets stop fighting over bandwidth. It’s the first practical model+optimizer for fine-grained chiplet DMA patterns and it unlocks big real-world wins in latency and bandwidth across DNNs and LLMs.

source S2, crossref