Adaptive GPU Power Capping: Balancing Energy Efficiency, Thermal Control and Performance

Abstract

GPU power caps are usually blunt. This work uses machine learning to choose caps that save energy and reduce temperature while keeping performance close to unconstrained runs.

Key contribution

Developed a machine-learning model that predicts effective GPU power caps, achieving about 12.87% energy savings and 11.38% lower temperature with minimal performance overhead.

Lessons learned

Systems research is most useful when the objective is multi-dimensional—energy, heat, and performance together—and when the intervention is something operators can actually apply. Small, measurable gains compound at cluster scale.