Reproduce “Dissecting How Die Scaling Breaks GPU Fine-grained Scheduling” 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
Xiaoze Fan, Jianhao Wang, Weihao Cui, Han Zhao, Zhuobin Huang, Yangjie Zhou, Yuxian Qiu, Shixuan Sun and 3 more
The authors characterize GPU compute and memory layouts and use this physical topology to guide fine-grained scheduling. Their measurements show that equal resource counts can differ in capability, and topology-aware placement can outperform count-based or NUMA-only choices. In Qwen3-8B decode multiplexed with prefill on an H200, it raises throughput 14.3% over Cluster-Aware overlap; the result is prototype- and workload-specific.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.