適用於嵌入式圖形處理器之高階耗能模型

Abstract

圖形處理器在近年來已逐漸成為在嵌入式系統中不可缺少的元件，尤其是被大量地使用在休閒娛樂用的平板電腦上。圖形處理器可以提升圖形程式的渲染速度，但同時也帶來可觀的耗能，所以如何提升能源效率已經成為目前嵌入式圖形處理器的重要課題。目前已存在許多用於桌上型電腦的圖形處理器的耗能模型，由於硬體的限制以及不同的硬體架構，這些耗能模型無法使用在嵌入式圖形處理器上。因此本篇論文針對嵌入式圖形處理器建立一套完整的量測環境，並且從軟體角度提出了一個耗能模型。實驗結果顯示，耗能模型與實際量測的耗能相比，其誤差率小於3%。

Recently, graphic processing unit (GPU) has become an indispensable component for embedded systems, especially for tablets. GPU can speed up rendering of graphic applications, but, on the other hand, it consumes a large amount of energy consumption. Hence, energy-efficiency of GPU becomes an urgent issue for embedded systems. Few power models for GPU are established on desktop computers. With limited hardware support and different architectures, they cannot be directly applied to embedded GPUs. Therefore, this thesis builds an experimental environment for fine-grained power measurement of an embedded GPU, and presents a software-level energy model for energy estimation. The experimental results demonstrate that the error rate of the estimated results is under 3%.