Reproduce “ArchAgent v2: A Case Study with the Data Prefetching Championship” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Abraham Gonzalez, Raghav Gupta, Akanksha Jain, Hanna Alam, Alexander Novikov, Po-Sen Huang, Matej Balog, Marvin Eisenberger and 7 more

ArchAgent v2 searches for a three-level cache prefetcher by evolving each cache level in stages, checking candidates against storage budgets before simulation and then jointly optimizing the levels. It combines stride, delta-history, and reinforcement-learning mechanisms with bandwidth-aware controls. In DPC4, its normalized geometric-mean IPC score was 1.0382, narrowly above BertiGO’s 1.0352, but below BertiGO in multi-core.

Key insights

  • Agentic AI can effectively design complex multi-level microarchitectural components, achieving performance competitive with, and in some aspects surpassing, human-engineered solutions.
  • Embedding hardware-realizability constraints directly into the evolutionary loop is crucial for generating physically viable designs and efficiently guiding the search.
  • A cascaded search strategy effectively manages the complexity and simulation latency of large, interacting design spaces, particularly when transitioning from single-level to multi-level optimization.
  • LLM-based agents excel at synthesizing and combining existing architectural principles to create high-impact optimizations, rather than generating fundamentally novel concepts.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.