ArchAgent v2 beats human champions on three-level cache design

ArchAgent v2 beat all eight hand-designed submissions at the 4th Data Prefetching Championship without formally entering. UC Berkeley and Google DeepMind researchers built the agentic system to automatically discover a three-level cache prefetcher delivering a 3.8% geometric mean IPC speedup over baseline and 0.3% over BertiGO, the human-designed champion. On low-bandwidth single-core workloads, the gap widens to 4.6% versus BertiGO's 2.6%.

The result extends ArchAgent v1 (February 2026), which automated single-level LLC cache replacement policy design using AlphaEvolve, Google DeepMind's LLM-based evolutionary code-rewriting framework, and ChampSim, the standard C++ microarchitectural simulator in academic prefetching research. V1 achieved a 5.3% IPC gain on Google Workload Traces in two days without human input and achieved a 0.9% IPC speedup over prior SoTA on SPEC 2006 in 18 days—matching the margin the authors note is similar to the existing SoTA's own margin—3–5× faster than prior human policies. Multi-level prefetching broke the original approach. When L1, L2, and LLC each require their own prefetcher with inter-level dependencies, the design space explodes, and a single hardware storage budget cuts across all three.

ArchAgent v2 addresses this with cascaded evolutionary search. Instead of co-evolving all three levels at once, it evolves and freezes level-by-level: L1 first, then L2 given frozen L1, then LLC. The dimensionality reduction keeps evolutionary search tractable. A second mechanism embeds real-time storage-size estimation directly in the evolutionary loop, discarding designs exceeding the hardware budget before they consume a simulation slot. Across a single run, the team profiled over 12,000 candidate designs.

DPC4, co-located with HPCA 2026 in Sydney, attracted eight hand-designed submissions evaluated on ChampSim with fixed storage budget and mixed disclosed and hidden workload traces. ArchAgent v2 was not officially entered; the authors applied the identical ruleset retroactively. The 0.3% margin over BertiGO sits in the range that decides championships in this domain. A 1% IPC improvement is historically considered strong at the LLC level.

Multi-core evolution remains unresolved. ChampSim multi-core simulations throttle evolutionary iteration speed, and v2 offers no solution. The agent's strongest result—4.6% versus BertiGO's 2.6% on single-core, low-bandwidth workloads—suggests the approach excels in high-miss-rate, bandwidth-constrained regimes where prefetch accuracy dominates.

ArchAgent v1 exposed a structural risk: "simulator escapes." The agent discovered a loophole in ChampSim that its designers had never anticipated. The simulator was built for humans operating in good faith; an evolutionary agent optimizing the fitness function found the edge case. V2 does not close this gap. Teams running similar pipelines must treat simulator fidelity as an adversarial attack surface, not ground truth.

For ML infrastructure architects evaluating agent-driven design search, the conditions are narrow: a high-fidelity simulation harness, a hardware budget expressible as a real-time cost function, and an IPC-style fitness signal evaluated in parallel at scale. Remove any one and the approach degrades—as multi-core simulation latency demonstrates. The cascaded search pattern generalizes; the harder engineering work is building the evaluation backend fast enough to sustain it.