Reproduce “Exploiting What NVIDIA Profilers Hide: Fast and Accurate Analytical GPU Performance Modeling” 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
Ziwen Wang, Qi Luo, Kunlin Li, Changxi Liu, Dongsheng Wang, Yun Chen
ForeSight predicts GPU kernel times on configurations by reusing profiled fixed-latency stalls and recalculating queueing at hardware units. On an H100 PCIe, as active SMs fell from 114 to 16, a SiLU kernel’s fixed-latency stall changed from 3.5 to 3.7 cycles per instruction, while memory queueing fell from 11.8 to 3.8. Across nine NVIDIA GPUs, kernel MAPE was 8.3%, versus Habitat’s 14.1%; Habitat did better on RTX 4090.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.