Towards Accurate GPU Power Modeling for Smartphones

With the increasingly high power consumption of smartphone GPUs, accurate GPU power modeling is desirable for mobile game developers to optimize the power performance of their game code. However, existing GPU power models for smartphones simply use only GPU utilization to estimate GPU power consumption. In this paper, we observe that GPU utilization fails to capture real usage of modern mobile GPU hardware and thus has a high estimation error on modern smartphones. We discover that the root cause is that different types of GPU operations may consume very different amount of power even they have the same GPU utilization. To improve the accuracy of GPU power modeling, we propose to consider more fine-grained predicators, including vertex-processing load and pixel-processing load, in modeling GPU power consumption. We report how to build such a new GPU model for commercial smartphones and evaluate it using various benchmarks and mobile games. Experimental results show that compared to existing utilization-based model, our new model is able to significantly reduce the maximum modeling error from 14.8% to 6.5%.