GPU Power Modeling And Optimization

With more powerful Graphic Processing Unit (GPU), especiallylaunch of NVIDIA’s CUDA (Compute Unified Device Architecture)platform, general purpose GPU (GPGPU) computing gradually becomes amajor trend in the field of high performance computing. However, powerconsumption is always a big problem for GPGPU computing. Althoughgenerally speaking, the Performance/Watt ratio of GPU is higher thantraditional CPU (Central Processing Unit), its high absolute power will stillcauses problems such as rising chip packaging and cooling cost ordeclining device lifetime and reliability. And high power consumption isalso against the trend of Green Computing nowadays. In the academiccircle, GPU power related research is still at the beginning level, and thereare a lot of key technologies to discover.The two major contributions in this paper are:1) By investigating intoCUDA programming and memory models and analyzing how its PTX(Parallel Thread eXecution) and native instructions executed, we built anative instruction level power model based on GPU calculation andmemory access activities.2) Using dynamic computing capability andmemory bandwidth scaling, we proposed a method to obtain minimalpower consumption for four different workload categories withoutperformance loss. And we applied the model to estimate runtime power ofdifferent workloads under different GPU core clock and memory clock,thus verified our power optimization method.Detailed evaluation on7typical workloads verified that our powermodel’s prediction accuracy was over94%. Meanwhile, our optimizationstrategy increased average performance per watt of4different workloadcategories by14.5%. This paper is partially sponsored by the National High-TechnologyResearch and Development Program of China (863Program)(No.2009AA012201).