Research On The Main Factors Influencing Power Consumption Of CUDA Program

Since GPU itself having plenty of arithmetic units, GPU has been widely used in many general purpose computing fields.However, higher computing performance on a chip brings high power consumption than CPU chip. High power consumption obviously brings some problems including the reliability decline, stability of the system decline and the calculation of cost increasing,can not be ignored.This paper describes the CUDA programming model, CUDA software architecture and memory model, as the subject of research basics. By analyzing the impact of CUDA applications of thread organizational structure, the number of thread blocks and storage location(global memory and shared memory) of CUDA program variables of these three main factors,we simulate examples such as vector addition and matrix multiplication in CUDA SDK.At the end of paper, the CUDA program of vector addition and matrix multiplication are estimated on power simulator GPUWattch and analysis of the experimental results.Experimental analysis of results, thread structure, thread block number and CUDA program variable storage location have an effect on cuda power consumption.Thread of organization structure influencing the thread of power consumption.Initially, with the number of thread blocks increasing, power consumption becomes smaller.Then, when power consumption reaches a certain minimum, as the number of thread blocks increases,power consumption becomes larger. The factor of variables of CUDA program stored in the shared memory and global memory obtained, rational use of shared memory, power consumption could be reduced.