Measuring GPU Power with the K20 Built-in Sensor

Abstract

GPU-accelerated programs are becoming increasingly common in HPC, personal computers, and even handheld devices, making it important to optimize their energy efficiency. However, accurately profiling the power consumption of GPU code is not straightforward. In fact, we have identified multiple anomalies when using the on-board power sensor of K20 GPUs. For example, we have found that doubling a kernel's runtime more than doubles its energy usage, that kernels consume energy after they have stopped executing, and that running two kernels in close temporal proximity inflates the energy consumption of the later kernel. Moreover, we have observed that the power sampling frequency varies greatly and that the GPU sensor only performs power readings once in a while. We present a methodology to accurately compute the instant power and the energy consumption despite these issues.