For embedded GPUs, Vatjus-Anttila et al. [1] built an energy model based on three render complexity characteristics: number of triangles, render batches and addressed texels. Th eir proposed energy model overestimates the power consumption of the graphic processor. To compensate the error, they empirically deducted 45% of the consumption based on the ad-hoc hypothesis that 50% of the 3D content could be left unacknowledged due to the back-face triangle culling, and 10% due to depth testing. Mochocki et al. [10] used three embedded processors to simulate different stages of the 3D pipeline. They analyzed how the factors (resolution, frame rate, level of detail, lighting model, and texture model) affect the 3D pipelines to result workload variations and imbalances.
Our approach differs from above studies by systematically design the testing benchmark to analyze the relationships between high-level graphics parameters and the graphics pipelines. Then, we build a high-level energy model for embedded GPU that only requires high- level parameters to estimate the energy for real-time graphic rendering.