Reproduce “The GPU Needs an Operating System” 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
Reza Nourai, Behnam Bastani
The authors propose an OS-like GPU layer that uses latency slack to run background work without missing urgent text-generation deadlines. Decode repeatedly fetches weights, leaving tensor cores idle. In Qwen3.5-397B, tensor cores averaged 7% active at a deadline-meeting load; a one-GPU Kimi K2 proxy ran real FP8 work at 0.09 GPU-equivalents and 4.7% median latency cost, not demonstrating general sharing.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.