Current Prediction Model of GPU Oriented to General Purpose Computing

In order to reduce the complexity of energy consumption analysis in GPU general computing and solve the dependence of energy consumption analysis on hardware tools, a GPU general-purpose computing current prediction model was proposed based on the analysis of different GPU architectures and CUDA program execution rules. Firstly, the GPU program was decomposed and the features of the source program were extracted. Then, the relationship model between program features and running current consumption was figured out by using multiple linear regression and BP neural network. Finally, two prediction models were trained by typical sample data. The experimental results showed that the single-program prediction error was less than 10% and the average prediction error was less than 6%. The proposed two GPU current prediction models had good prediction accuracy and good universal property for different GPU architectures. The current prediction model laid the foundation for further analysis of GPU general-purpose computing energy complexity.