ArchAgent: Agentic AI-driven Computer Architecture Discovery

Raghav Gupta, Akanksha Jain, Abraham Gonzalez, Alexander Novikov, Po-Sen Huang, Matej Balog, Marvin Eisenberger, Sergey Shirobokov, Ngân Vũ, Martin G Dixon, Borivoje Nikolic, Parthasarathy Ranganathan, Sagar Karandikar

Abstract:

Agile hardware design flows are a critically needed force multiplier to meet the exploding demand for compute.
Recently, agentic generative artificial intelligence (AI) systems have demonstrated significant advances in algorithm design, improving code efficiency, and enabling discovery across scientific domains.

Bridging these worlds, we present ArchAgent, an automated _computer architecture discovery_ system built on AlphaEvolve. We show ArchAgent's ability to automatically design/implement state-of-the-art (SoTA) cache replacement policies (architecting new mechanisms/logic, not only changing parameters), broadly within the confines of an established cache replacement policy design competition.

In two days and without human intervention, ArchAgent generated a policy achieving a 5.3% IPC speedup improvement over the prior SoTA on public multi-core Google Workload Traces. On the heavily-explored single-core SPEC 2006 workloads, in only 18 days ArchAgent generated a policy showing a 0.9% IPC speedup improvement over the existing SoTA (a similar "winning margin" as reported by the existing SoTA). Comparing against the effort involved in developing the prior SoTA policies, ArchAgent achieved these gains 3-5$\times$ faster than humans.

Agentic flows also create an opportunity once a hardware system is deployed, which we call ``post-silicon hyperspecialization''. This means having the agent tune runtime-configurable parameters exposed in hardware policies to further align the policies with a specific workload (mix). Exploiting this, we demonstrate a 2.4% IPC speedup improvement over prior SoTA on the SPEC 2006 workloads.

Since ArchAgent is a first-of-its-kind architectural discovery system, we also outline lessons learned and broader implications for computer architecture research in the era of agentic AI.
For example, we demonstrate the phenomenon of ``simulator escapes'', where the agentic AI flow discovered and exploited a loophole in a popular microarchitectural simulator---a consequence of the fact that these research tools were designed for a (now past) world where they were exclusively operated by humans acting in good faith.