Performing of the FDTD method calculations on "Sergey Korolev" GPUs via MATLAB language

The paper deals with the FDTD-method block algorithm efficiency and with its implementation in MATLAB language on the high performance computer called Sergey Korolev. Block algorithms with one-dimensional and two-dimensional decomposition are considered. As a result of the study, it was found that the one-dimensional decomposition efficiency is 3 times higher than the two-dimensional decomposition efficiency. The average performance of the FDTD-method block algorithm implementation with one-dimensional decomposition is 190 million nodes per second, which corresponds to 15 times speedup on the considered system. The maximum possible domain discretization in GPU computing has been expanded from 50 million to 410 million nodes. Also, it exceeded the maximum discretization in CPU computing by 60 million nodes.