Within the last couple of years, Graphical Processing Units (GPUs) have become one of the main drivers of the increase of the compute power of the TOP500 list. Next to High Performance Computing (HPC) use cases GPUs were also the main driver enabling the wide spread use of neural networks for machine learning. However, this increase of compute power comes at the price of an ever-increasing amount of energy consumption, with all the negative effects of associated environment and operational

more »

... . At the moment, there are 2 main vendors for HPC GPUs: Advanced Micro Devices (AMD) and NVIDIA. Both offer different software stack. This paper will discuss the impact of the choice of framework on the power draw and energy usage of the running kernels compared to the "native" software stack (i.e. CUDA and HIP for NVIDIA and AMD respectively). The goal is to gain an understanding whether the choice has a large enough impact on energy consumption to be significant for the choice of framework when aiming for green computing. As matrix multiplication is a common operation for a fast array of work loads including machine learning a simple matrix multiplication kernel is used and rudimentarily optimized for this paper. The comparison explores the impact of single and double precision, explicit synchronisation, and shared memory. The analysis has shown that whilst the choice of framework can have an impact of up to 15 % on the energy consumption, that difference often vanishes when compared to the savings caused by other factors that could easily reach 60 % e.g. for utilizing shared memory.