Power-Gated Arithmetic Circuits for Energy-Precision Tradeoffs in Mobile Graphics Processing Units

摘要

In mobile devices, limiting the Graphics Processing Unit's (GPU's) energy usage is of great importance to extending battery life. This work shows that significant energy savings can be obtained by reducing the precision of graphics computations, yet maintaining acceptable quality of the final rendered image. In particular, we focus on a portion of a typical graphics processor pipeline - the vertex transformation stage - and evaluate the tradeoff between energy efficiency and image fidelity. We first develop circuit-level designs of arithmetic components whose precision can be varied dynamically with fine-grained power gating techniques. Spice simulation is used to characterize each component's energy consumption, based on which a system-level energy model for the entire vertex stage is developed. We then use this energy model in conjunction with a graphics hardware simulator to determine the energy savings for real workloads. Results show that, solely by changing the precision of the arithmetic in the vertex shaders' ALUs, we can save up to 10% of the overall energy in the GPU. As promising as these savings are, we expect that our circuits and approach will enable energy-precision tradeoffs in other areas of the graphics pipeline, such as the pixel shader, which consume even more energy than the vertex shader.