Reproduce “Characterizing CUDA Locality Domains on Blackwell 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
Rocky Wang, Joyce Zhou, Keith Jiang, Jason Zhao
The authors assess CUDA Locality Domains on eight B300 GPUs, separating data placement from compute binding and testing both alone and together. Results depend on the accessor: memory placement helps a 1-MiB remote copy-engine read, compute binding helps a tested collective, and both raise hot FP8, M=8 grouped GEMM performance to 1.37× stock; streaming bandwidth and tuned attention limit these gains.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.