一种基于GPU的高精度体系结构级功耗模型

High-precision Architecture-level Power Model Based on GPU

随着硬件功能的不断丰富和软件开发环境的逐渐成熟,GPU开始被应用于通用计算领域,协助CPU加速程序运行。为了追求高性能,GPU往往包含成百上千个核心运算单元,高密度的计算资源使得其性能远高于CPU的同时功耗也高于CPU,功耗问题已经成为制约GPU发展的重要问题之一。在深入研究Fermi GPU架构的基础上,提出一种高精度的体系结构级功耗模型,该模型首先计算不同native指令及每次访问存储器消耗的功耗;然后根据应用在硬件上的执行指令和采样工具获得采样结果,分析预测其功耗;最后通过13个基准测试应用对实际测试与功耗模型测试结果进行对比分析,该模型的预测精度可达90%左右。

As hardware functions are constantly developing,and software development environments gradually mature,the graphics processing unit（GPU）has been applied to general purpose computation to help the central processing unit（CPU）accelerate a program.To obtain high performance,a GPU generally contain hundreds of core arithmetic units.Owing to the existence of high-density computing resources,the performance of the GPU is much superior to that of the CPU,while its power consumption is larger than that of the CPU.Power consumption has become one of the important issues restricting the development of GPU.Based on the study of the Fermi GPU architecture,a high-precision architecture-level power model was proposed in this research.In this model,the power consumed by different native instructions,and each memory access,were first calculated,then the power consumption was analysed and predicted according to the execution instructions as applied to the hardware,and the sampling results were acquired using sampling instruments.Finally,the results obtained from practical testing and the power model were compared by using 13 benchmark applications.It is demonstrated that the prediction accuracy of the model can reach approximately 90%.