high performance computing on graphics processing units: hgpu.org

Recently, many numerical methods that are developed for the solution of electromagnetic problems have greatly benefited from the hardware accelerated scientific computing capability provided by graphics processing units (GPUs) and orders of magnitude speed-up factors have been reported. Among these methods, the finite-difference frequency-domain (FDFD) method as well can be accelerated substantially by utilizing an efficient algorithm customized for GPU computing. In this contribution, an algorithm is presented that treats iterative solution of the FDFD linear equation system similar to solution of three-dimensional Finite-Difference Time-Domain (FDTD) method, which inherently yields itself to high level parallelization. The presented algorithm uses BICGSTAB iterative solver. Integrated with BICGSTAB, an efficient method of performing matrix-vector products for the linear system of FDFD equations is adapted and implemented in Compute Unified Device Architecture (CUDA). It is shown that FDFD can be solved with a speed-up factor of more than 20 on a GPU compared with the solution on a central processing unit (CPU), while memory usage as well can be reduced substantially with the presented algorithm.