Fast and Fusiest: An Optimal Fusion-Aware Mapper for Accelerator Modeling and Evaluation
Tanner Andrulis ∗, Michael Gilbert ∗, Vivienne Sze, Joel S. Emer · 2026
Read Fast and Fusiest: An Optimal Fusion-Aware Mapper for Accelerator Modeling and Evaluation by Tanner Andrulis ∗, Michael Gilbert ∗, Vivienne Sze, Joel S. Emer (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-15166.txt.
Description
Andrulis and colleagues introduce the Fast and Fusiest Mapper, an optimal fusion-aware mapper for accelerator design. Tensor algebra accelerators must schedule data movement and operations well, and fusion, holding data on-chip between computation steps, cuts energy and latency by avoiding costly off-chip movement. The best fusion choice varies with workload and shape, so a mapper that searches mappings can improve both metrics markedly.
Prior mappers cannot find optimal fused mappings in feasible time, since the count of fused mappings grows exponentially with computation steps. FFM shrinks the search space by pruning partial mappings proven never to belong to optimal mappings, then joins surviving partial mappings into optimal fused mappings. Against TransFusion, a state-of-the-art accelerator with hand-optimized fusion, FFM reduces energy-delay product. It also finds mappings orders of magnitude faster than TileFlow and SET, and at equal runtime reduces energy-delay product further.
Catalog record
| CREATOR | Tanner Andrulis ∗, Michael Gilbert ∗, Vivienne Sze, Joel S. Emer |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 46 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-15166