Optimizing Performance in Power Bounded GPU Computing

Student: Tyler Allen (Clemson University)
Supervisor: Rong Ge (Clemson University)

Abstract: Power is consistently considered as a top challenge in HPC. To address the power challenge, power efficient GPU accelerators have become a standard component in mainstream systems and contribute the majority of compute capacity. Nevertheless, GPUs are still constrained by limited permissible power and required to sustain performance growth. In many situations, there is a need to set a power cap on GPUs. In our research we find the default hardware power capping on GPUs loses performance by up to 35%, in comparison to the maximum achievable performance. We demonstrate that performance can be improved by a large margin for the applications studied, compared to the default power capping and allocation.

ACM-SRC Semi-Finalist: no

Poster: PDF
Poster Summary: PDF


Back to Poster Archive Listing