Performance and Energy Aware Training of a Deep Neural Network in a Multi-GPU Environment with Power Capping.

In this paper we demonstrate that it is possible to obtain considerable improvement of performance and energy aware metrics for training of deep neural networks using a modern parallel multi-GPU system, by enforcing selected, non-default power caps on the GPUs. We measure the power and energy consumption of the whole node using a professional, certified hardware power meter. For a high performance workstation with 8 GPUs, we were able to find non-default GPU power cap settings within the range of 160–200 W to improve the difference between percentage energy gain and performance loss by over 15.0%, EDP (Abbreviations and terms used are described in main text.) by over 17.3%, EDS with k = 1.5 by over 2.2%, EDS with k = 2.0 by over 7.5% and pure energy by over 25%, compared to the default power cap setting of 260 W per GPU. These findings demonstrate the potential of today’s CPU+GPU systems for configuration improvement in the context of performance-energy consumption metrics.