Research On GPU Performance Prediction Based On AMD Tonga Architecture

As a special processor chip for graphics modeling and general computing,GPU has been a hot spot in chip research.However,in the field of performance verification of GPU chips,as the size of GPUs becomes larger and the systems become more complex,traditional verification methods are difficult to solve the problem that the performance of the chips does not reach the theoretical expectations.Therefore,a better performance prediction model is needed to help locate GPU internal module bottlenecks and evaluate GPU architecture performance.Based on AMD's Tonga GPU chip,this thesis presents an improved GPU performance prediction model and completes the verification based on the model.Firstly,in the framework establishment of the model,the overall design of the model framework is studied according to the engineering requirements of multiple test case input,multiple hardware configuration environment,and data result summary output.Secondly,on the specific problem of model design,the time interface information of each module of GPU rendering pipeline is studied,and the time cost function of the module is calculated,which provides theoretical basis for performance analysis and prediction of the model.Then,the program design of the prediction model is completed,and the implementation methods of data interception,performance prediction analysis and data summary output are mainly solved.Finally,build a hardware verification platform,through the graphical rendering and general computing two rendering pipelines,verify the performance changes caused by the GPU engine frequency changes,memory frequency changes,and the number of calculation units to achieve the function of the model of the prediction model.Completeness and predictive accuracy verification.The experimental results show that the predictive model software is complete in function,the bottleneck module is accurate in positioning,and the prediction accuracy of the bottleneck module blocking time ratio is less than 10%,which significantly shortens the time-consuming time for GPU architecture problem positioning.