PerfDojo: Automated ML Library Generation for Heterogeneous Architectures Conference: The International Conference for High Performance Computing, Networking, Storage and Analysis; November 16–21, 2025; St Louis, MO, USA The International Conference for High Performance Computing, Networking, Storag
Andrei Ivanov, Siyuan Shen, Gioele Gottardo, Marcin Chrapek, Afif Boudaoud, Timo Schneider, Luca Benini, Torsten Hoefler · 2025
Read PerfDojo: Automated ML Library Generation for Heterogeneous Architectures Conference: The International Conference for High Performance Computing, Networking, Storage and Analysis; November 16–21, 2025; St Louis, MO, USA The International Conference for High Performance Computing, Networking, Storag by Andrei Ivanov, Siyuan Shen, Gioele Gottardo, Marcin Chrapek, Afif Boudaoud, Timo Schneider, Luca Benini, Torsten Hoefler (2025) 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-2511-03586.txt.
Description
Andrei Ivanov and colleagues introduce PerfDojo, a training ground where language models learn to write fast machine learning libraries for mixed hardware. Modern models must run well across processors, graphics chips, and accelerators, each with quirks of instruction set and data type. Hand tuning is slow, and automatic tools often hide their logic in opaque forms.
PerfDojo frames optimization as a reinforcement game, with a method called PerfLLM guiding language models through it. Scores reward real speed on the target hardware, not code that merely looks right. Sparse and quantized model features enter the challenge set directly. Portable, self-tuning libraries could cut the expert labor now burned on each new chip generation.
Catalog record
| CREATOR | Andrei Ivanov, Siyuan Shen, Gioele Gottardo, Marcin Chrapek, Afif Boudaoud, Timo Schneider, Luca Benini, Torsten Hoefler |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 52 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2511-03586