high performance computing on graphics processing units: hgpu.org

The Finite-Difference Time-Domain method (FDTD) is a modelling technique for electromagnetic waves propagation. There is a great range of domains of application, for example geophysics, defence, microwaves like radar, or biomedicine. However, FDTD is a computationally intensive method, but has potential for parallelisation. The use of General-Purpose computing on Graphics Processing Units (GPGPU) is examined here to enhance the performance of the FDTD method. GPUs are good candidates thanks to their massively parallel architecture. We expose three principal methods to implement the FDTD method on GPUs. Two of them have been implemented with CUDA and executed on a Tesla architecture. The results show that one approach take the most of the capability of GPU’s parallelism. This approach is at least six times faster than the others in single precision, and at least three times faster in double precision. We searched the reason of such a difference. The best method presents a high capacity of latency hiding, coupled with a good occupancy of the multiprocessors.