Acceleration of Iterative Solver for Electromagnetic Analysis using GPU/MIC

The mixed precision variable preconditioning (VP) Krylov subspace method is implemented on Graphics Processing Unit (GPU) and Many Integrated Core (MIC) architecture, and the linear system obtained from an electromagnetic analysis is solved by the method. In recent year, high-performance multi/many core computer architecture can be cheap and easily available, and the simulation code must be parallelized by using parallelization API. Although the ordinal program code that developed on CPU can be implemented on MIC without transcribing, GPU programming cost using CUDA becomes very high. In the present study, the performances of the mixed precision VP Krylov subspace method on GPU and MIC are compared by solving the linear system obtained from electromagnetic analysis discretized by edge element. The results of computation show that the communication cost of MIC is much higher than that of GPU.