ArchAgent v2 auto-designs a three-level cache prefetcher that beats the human championship winner

arXiv·medium signalXBlueskyLinkedInCopy link

A Google and UC Berkeley team (including DeepMind's AlphaEvolve authors) evolved over 12,000 candidate designs under the exact rules of the 4th Data Prefetching Championship, using cascaded evolutionary search that freezes prefetchers level by level plus a hardware-realizability feedback loop that folds real-time size estimation into the evolution. The result: 3.8% geomean IPC speedup over baseline, edging the hand-designed champion BertiGO by 0.3% overall and pulling clearly ahead on low-bandwidth single-core configs (4.6% vs 2.6%). This is one of the sharper existence proofs that agentic search can produce artifacts that beat expert humans in a domain with hard physical constraints, not just in software.