Preprint
Sep 2026
Entwine: Coordinating Tiled Computation and Fine-Grained Communication across GPUs
Entwine is presented, which coordinates tile computation order, fine-grained communication, and SM resource allocation to minimize overall completion time and achieves a geomean speedup of 1.232x over cuBLAS+NCCL, and outperforms state-of-the-art overlap baselines by 3.8% in geomean.
Kai Ma, Quan-Feng Lv, Jing-Guo Ge et al.
· 0 citations