Workshop: EESP 2026: 3rd International Workshop on Energy Efficiency with Sustainable Performance: Techniques, Tools, and Best Practices
Authors: Matthew Dearing and Yiheng Tao (University of Illinois at Chicago), Xingfu Wu (Argonne National Laboratory), Zhiling Lan (University of Illinois at Chicago), and Valerie Taylor (Argonne National Laboratory)
Abstract: Energy efficiency is a critical objective for high-performance computing, yet optimizing scientific applications for energy has traditionally required deep expertise and substantial manual effort. This paper investigates whether large language models (LLMs) can generate energy-efficient parallel codes when guided by empirical execution feedback, and whether their optimizations adapt to the target application and architecture. We propose LASSI-EE, an automated refactoring framework integrating runtime power profiling, energy-aware prompting, self-correcting feedback loops, and an LLM-as-a-Judge for screening generated code. To quantify expected savings under LLM non-determinism, we introduce energy-reduction@k, a metric that accounts for all generation attempts, including failures. Across twenty-two scientific benchmarks and applications on NVIDIA A100 and AMD MI100 GPUs, compiled at full optimization, LASSI-EE achieves expected energy savings of 29% for a single attempt and 48% when selecting the best of three. Average reductions reach 34%-36% among passing energy-reducing refactorings, largely preserving or improving runtime without precision-reducing transformations.
Back to EESP 2026: 3rd International Workshop on Energy Efficiency with Sustainable Performance: Techniques, Tools, and Best Practices Archive Listing
Back to Full Workshop Archive Listing