Do AI Agents Understand Computer Architecture?
Ambika Sharan, Grigory Chirkov and Soheil Abbasloo at Microsoft Research build AutoTuring, which gives the same agent the same 15-dimensional accelerator design space twice, once with named architectural knobs and simulator counters and once as anonymous variables on [0,1].
Ask this paper
The framing is the only variable. Evaluator, legal space and reachable optima are held identical, so the gap between the two conditions measures whether the agent uses architectural meaning rather than searching competently over unlabeled knobs.
Naming pays 12.3 percent. On a nine-kernel FP16 GEMM basket the architect condition beats its blind counterpart by 12.3 percent on average and a modeled H200 by 5.4 percent, using 70.1 percent fewer simulator calls.
A critic loop substitutes for architectural knowledge. Adding a critic recovers most of that gap for the blind agent and buys the architect nothing, so structured critique and domain knowledge act as substitutes rather than complements.
Existing evaluations cannot make this distinction. They vary the agent while holding the problem framing fixed; this design does the opposite, which is what the authors take to be the contribution.
The result is preliminary by the authors' own statement. Five to six runs per condition on a single modeled accelerator, reported as such.
Abstract
Agents are increasingly asked to design hardware, and increasingly reported to succeed. Such reports establish that a design improved; they cannot establish why. An agent that improves an accelerator may be reasoning about the machine, or may be searching competently over knobs whose meaning it never recovers -- and only the first transfers to the next architecture. Existing evaluations cannot tell the two apart, because they vary the agent while holding the framing of the problem fixed. We do the opposite. AutoTuring hands the same agent the same 15-dimensional accelerator space twice: once as named architectural knobs with simulator counters, once as anonymous variables on [0,1], with the evaluator, the legal space and the reachable optima held identical, so that the only thing that varies is whether the problem means anything. The gap between the two is the measurement. On a nine-kernel FP16 GEMM basket, meaning pays: the architect beats a modeled H200 by 5.4% and its blind counterpart by 12.3% on average, with 70.1% fewer simulator calls. It does not pay uniquely: a critic loop recovers most of that gap for the blind agent and buys the architect nothing, so architectural knowledge and structured critique behave as substitutes rather than as complements. We report these as preliminary findings -- five to six runs per condition on a single modeled accelerator -- and take the comparison itself, not the accelerator, to be the contribution.