Running for the next two and a half years, the research project aims to define new industry standards for resource and performance monitoring to be widely adopted by embedded hardware GPU vendors. The consortium will define a methodology for accurate power estimations for embedded GPU and will try to enhance existing Dynamic Voltage and Frequency Scaling (DVFS) mechanisms for optimum power management with sustained performance.
Ideally, the result will be a unique power and performance visualization tool which informs application and GPU device driver developers of potential power and performance improvements.
“It is very hard for application developers to know what power the GPU will consume when running their code”, explained Philip Harmer, Senior Manager at the Samsung Research & Development Institute UK. “Typically, GPU vendors provide good practice guidance for performance optimisation and improvements, they aim for maximum performance. Although they have a very good knowledge of their GPU’s internals and a very good idea of code implications on power dissipation, they don’t make that information available to developers”, he continued.
GPU vendors are willing to share benchmarking results, but they don’t want to reveal more, although you could argue that they all run some form of power optimization routines at chip design level, and that they could infer some coding tips without compromising their IP.
“We hope to keep our work as open as possible and create a standardized API to enable tool vendors to estimate GPU power dissipation” he said, without saying yet if the resulting tools would be made open source.
The consortium will investigate and define GPU power usage, measuring power dissipation on existing silicon from different vendors. Then, the different project partners will attempt to define new software models that match existing benchmarks when running code. Ideally, this will translate into power-estimated code libraries for different GPUs, or a low-power GPU code exploration tool helping developers make coding decisions based on their GPU of