Prediction of the Running Time for Multiplying Two 2048x2048 Matrices Using a GPU OpenCL SGEMM Kernel

Predicting execution time helps optimize performance, allocate resources, and compare hardware efficiently. For a fundamental operation like matrix multiplication, the applications span across various fields like high-performance computing and cryptography, and it would greatly benefit from accurate predictions. This paper presents the implementation of various models that predict the computation time of a system consisting of a CPU (Central Processing Unit) as well as a GPU (Graphics Processing Unit) as it performs the multiplication of two 2048 x 2048 matrices. The prediction is based on GPU specifications, which consist of 14 different configurations. The prediction models were able to properly learn the data and generalize exceptionally on new unseen data with a prediction accuracy of 99.94% exceeding the results of previous literature.