Accelerator Architecture: Fusion-Aware Mapper (MIT)
Researchers from MIT have developed the Fast and Fusiest Mapper (FFM), a novel mapper that efficiently finds optimal mappings in a comprehensive fused mapspace for tensor algebra workloads. The FFM focuses on fusion, which involves holding data on-chip between computation steps to reduce energy and latency. By quickly eliminating suboptimal mappings and constructing optimal fused mappings, the FFM outperforms prior approaches in finding optimal mappings for Transformers. This new mapper addresses the challenge of exponentially increasing search space with the number of computation steps, scaling its runtime linearly.
SemiEngineering•