ArchAgent: Towards Agentic Architecture Discovery

Raghav Gupta, Akanksha Jain, Borivoje Nikolic, Parthasarathy Ranganathan, Sagar Karandikar

Keywords: Agile hardware design, computer architecture

Abstract:

Agile hardware design "ows are a critically needed force multiplier to meet the exploding demand for compute. We present ArchAgent, an agentic AI system that builds on AlphaEvolve [7] to automate computer architecture discovery. By designing and implementing logic directly within ChampSim [4], ArchAgent discovered novel "winning" cache
replacement policies that improve IPC speedup over prior
SoTA by 5.3% on multi-core Google Workload Traces [1] in two days and by 0.9% on single-core SPEC06 [6] work-
loads in 18 days, in a cache replacement championship-like setting. ArchAgent achieved these gains 3-5⇥ faster than human-led efforts in developing prior SoTA policies. We also introduce "post-silicon hyperspecialization," achieving a 2.4% IPC speedup improvement on single-core SPEC06 workloads via automated per-workload runtime parameter tuning. Finally, we discuss lessons learned from working on ArchAgent such as the phenomenon of "simulator escapes" and the need for performant and reliable simulation infrastructure to enable a future of agentic AI-driven computerarchitecure.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 17