Reproduce “SyclKittens: A Tile Programming Model for Programmers and Coding Agents on Intel GPUs” 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
Yehong Jiang, Sheng Chen, Fangwen Fu, Yen-Kuang Chen, Xinmin Tian, Stuart H. Sul, Simran Arora
SyclKittens provides reusable tile operations encoding Intel GPU methods while leaving tile shape and prefetch schedules to programmers and agents. In a controlled single-GPU BF16 GEMM test, Opus 4.8 reached 82.1% of tuned oneDNN with SyclKittens, versus 57.5% with raw SYCL. On tested Llama-3.1-8B settings, composing SyclKittens with torch.compile achieved 1.59× geometric-mean full-request throughput over torch.compile alone across jointly fitting configurations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.