计算机工程

• 先进计算与数据处理 • 上一篇    下一篇

通用图形处理器功耗估算模型

王吉军1,程华2   

  1. (1.江南计算技术研究所,江苏 无锡 214083; 2.中国科学院计算技术研究所,北京 100080)
  • 收稿日期:2015-12-31 出版日期:2017-02-15 发布日期:2017-02-15
  • 作者简介:王吉军(1990—),男,硕士研究生,主研方向为绿色计算;程华,高级工程师。

Power Estimation Model of General Purpose Graphics Processing Unit

WANG Jijun  1,CHENG Hua  2   

  1. (1.Jiangnan Institute of Computing Technology,Wuxi,Jiangsu 214083,China; 2.Institute of Computing Technology,Chinese Academy of Sciences,Beijing 100080,China)
  • Received:2015-12-31 Online:2017-02-15 Published:2017-02-15

摘要: 为精准快速地获得GPU功耗数据,提出一种基于硬件性能计数事件的通用图形处理器(GPGPU)功耗估算方法。通过分析GPGPU程序运行时的功耗分布情况,选择一组与应用程序运行功耗密切相关的硬件性能计数事件集合,使用反向传播人工神经网络分析硬件性能计数事件与实时功耗间的关系,最终建立GPGPU功耗估算模型。实验结果表明,与多元线性回归的功耗估算模型相比,该模型具有更高的估算准确性和通用性。

关键词: 通用图形处理器, 硬件性能计数事件, 反向传播人工神经网络, 交叉验证, 功耗估算

Abstract: In order to get the GPU power data quickly and accurately,this paper proposes a General Purpose Graphics Processing Unit(GPGPU) power estimation model based on hardware performance counting events.Through analysing power distribution during GPGPU program running,it selects a set of performance events which are closely related to application program running power.Then it figures out the relationship between hardware performance counting events and realtime power using Back Propagation Atificial Neural Network(BP-ANN).Finally,it builds a GPGPU power estimation model.Experimental results indicate that compared with the Multiple Linear Regression(MLR) power estimation model,the proposed model has higher estimation accuracy and versatility.

Key words: General Purpose Graphics Processing Unit(GPGPU), hardware performance counting event, Back Propagation Atificial Neural Network(BP-ANN), cross-validation, power estimation

中图分类号: