Reproduce “DSO: A GPU Energy Efficiency Optimizer by Fusing Dynamic and Static Information” 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
Qiang Wang, Laiyi Li, Weile Luo, Yijia Zhang, Bingqiang Wang
Researchers from Harbin Institute of Technology, HKUST, and Peng Cheng Laboratory developed DSO, a GPU energy efficiency optimizer that integrates lightweight dynamic runtime metrics and static code features. The system demonstrated up to 19% energy conservation with less than 5% performance loss on a Tesla V100 GPU for unseen applications.
Key insights
- Fusing lightweight dynamic runtime metrics with static code features provides a more accurate and efficient way to model GPU behavior under DVFS than using either source alone.
- A parameterized theoretical model, which explicitly considers the DVFS roofline phenomenon and fundamental performance-energy tradeoffs, can effectively guide GPU energy optimization.
- The optimization problem for GPU DVFS can be simplified to a one-variable problem (optimizing core voltage) for a given memory frequency by leveraging the cost function's monotonic properties.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.