chips infrastructure
Hardware Knob Names Cut the Agent’s Search by 70%
The same accelerator space became easier when its variables carried architectural meaning.
Summary
The same accelerator space became easier when its variables carried architectural meaning.
AutoTuring gave the same agent the same 15-dimensional accelerator-design space twice: once with named architectural controls and simulator counters, and once as anonymous values between zero and one. On nine FP16 matrix-multiplication kernels, the architecture-aware agent beat its blind counterpart by 12.3% on average and used 70.1% fewer simulator calls. A critic loop recovered most of the gap for the blind agent but added nothing for the architecture-aware one. The authors call the result preliminary because each condition has only five or six runs on one modeled accelerator.
Why it matters
The same accelerator space became easier when its variables carried architectural meaning.
Limits and context
- The authors call the result preliminary because each condition has only five or six runs on one modeled accelerator.
Key claims
The same accelerator space became easier when its variables carried architectural meaning.
Qualification: The authors call the result preliminary because each condition has only five or six runs on one modeled accelerator.
Evidence: source-2026-09-20-005
Sources
- arXiv preprint 2609.19387arXiv · primary research
Corrections
No corrections have been recorded for this story.