Fig. 2

Memory bandwidth growth rate of CPU and GPU, and performance comparison between software (SW) implemented with parallel computing and others SW. (Left) The finite difference method applied to AF simulation is particularly suitable for parallel computing due to its iterative nature and can solve partial differential equations at high speed through GPUs with high bandwidth and many cores compared to CPU. (Right) CUVIA SW, which is implemented based on CUDA-based parallel architecture, shows significant performance compared to CARP (Johns Hopkins & Bordeaux Univ.) and MATLAB-based SW designed with serial architecture