Improved Performance of FDTD Computation Using a Thread Block Constructed as a Two-Dimensional Array with CUDA

摘要

In a previous study, the authors proposed an finite-difference time-domain (FDTD) implementation for a compute unified device architecture (CUDA) compatible graphics processing unit (GPU) using a thread block constructed as a two-dimensional (2-D) array. However, it was found that the larger the computational domain of the 2-D FDTD simulation using the GPU, the slower the computational speed. In the present paper, the authors investigated the computational performance with respect to the size of a thread block constructed as a 2-D array, and improved the performance of the implementation. Finally, regardless of the size of computational domain, the computational speed using a single GPU (NVIDIA GeForce GTX 280) achieved approximately 30.0 Gflops, which was approximately 20 times faster than that of a single core of a central processing unit (Intel 3.0-GHz Core 2 Duo). The improved performance was approximately 65% of the theoretical peak performance (47.23 Gflops) obtained by the theoretical memory bandwidth (141.7 GB/s).