An implementation of analytical power model on integrated GPU

GPU has become an important component of the high performance computing system and its principal duty is parallel computing rather than graphical display. Determining the power and energy consumption is necessary to the scaling of GPU. This paper presents a statistic model to evaluate the power and energy consumption of AMD's integrated GPU (iGPU). By collecting the data of performance counters from real hardware measurements, we apply linear regression method to estimate the energy consumed by iGPU. Our results show that the median absolute error is less than 3%. Due to the limits of profiling tool CodeXL, power sampling period is much longer than the kernel execution time. We propose a kernel extension method to lengthen the kernel execution time so that we can deal with this problem. Furthermore, we conduct a study on the importance of performance counters and explore the possibility to simplify our statistic model. The results suggest that the accuracy and stability is still acceptable when there are only 12 performance counters in the simplified model.