Intel / article

Do AI Agents Understand Computer Architecture?

Source: arxiv.org
Author: Ambika Sharan, Grigory Chirkov, Soheil Abbasloo
Date: Published Sep 18, 2026

Context

AutoTuring tests whether an AI hardware designer benefits from knowing what its design choices mean. Researchers gave the same identical design options, presented either as named hardware settings with simulator measurements or as anonymous numbers. Keeping the possible designs and unchanged helps separate useful architectural knowledge from effective trial and error. Study abstract.

The informed agent performed better and used fewer simulator calls. But adding a critique loop recovered most of the anonymous agent's performance gap, suggesting that structured feedback can substitute for some domain knowledge in this setting. These preliminary findings cover one simulated accelerator and five to six runs per condition, rather than demonstrated understanding across different machines. Reported findings.

Key takeaways · AI-distilled

AutoTuring gives the same tuning agent the identical 15-dimensional accelerator design space twice -- once with named architectural knobs, once as anonymous [0,1] numbers -- isolating whether understanding the problem's meaning actually helps.
With meaningful framing, the agent beat a modeled H200 by 5.4% and beat its own 'blind' version by 12.3% on average, while needing 70.1% fewer simulator calls.
Adding a critic loop (structured self-critique) closed most of that gap for the blind agent but gave the meaning-aware agent nothing extra -- suggesting architectural knowledge and structured critique substitute for each other rather than stacking.
The authors flag this as preliminary: five to six runs per condition on a single modeled accelerator, with the comparison method itself, not the accelerator design, as the contribution.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.

Read the source arxiv.org