A Low-Overhead Approach to Dynamic Frequency Scaling for GPUs Using Multi-Dimensional Energy Efficiency Curves

Today’s society faces the realities of climate change while becoming increasingly reliant on computational technology. It is therefore vital to use our technology economically with respect to the energy needed to power it. Previous research has approached this problem using Dynamic Voltage and Frequency Scaling (DVFS) - real-time adjustment of the power drawn by a machine to better match current execution conditions. However, there remains a need for a low-overhead, generalizable method that is effective on Graphics Processing Units running as part of a server. This paper adapts previous work on CPUs to introduce a novel Dynamic Frequency Scaling method for GPUs, capable of predicting energy-optimal frequencies during an execution. Comparison with default execution shows consistent energy savings of up to 33%, with a temperature-aware extension displaying potential for even greater efficiency gains.