Performance Comparison of GPUs with a Genetic Algorithm based on CUDA

Abstract

Generally genetic algorithm (GA) has disadvantage of taking a lot of computation time, and it is worth reducing the execution time while keeping good quality and result. Comparative experiments are conducted with one CPU and four GPUs using CUDA (Compute Unified Device Architecture) and generational GA. We implement the fitness functions of the GA which are suitable for the each environment of the CPU and the GPUs for performance comparison. When experimenting with the CPU, we handle the individual one by one. On the other hand, when experimenting with the GPU, we handle all individuals concurrently. And then we compare and analyze each result of our GA and each time required to process fitness function. There was not a huge difference between the results of the CPU experiment and the GPU ones. In the case of the analysis of computation time, the memory bandwidth of the GPU affects the computation time of fitness evaluation. The numbers of genes that can be processed at the same time are increased by the growth of the clock rate and the number of cores of GPU.