A comparison of computing architectures and parallelization frameworks based on a two-dimensional FDTD

The Finite-Difference in Time-Domain (FDTD) has been a very useful method for solving Maxwells Equations. Since simulations with large model spaces, or long non-sinusoidal waveforms, imply a huge amount of floating-point calculations, it is of practical interest to take advantage of current and emergent multicore architectures, such as Central Processing Units (CPUs) and Graphics Processing Units (GPUs) to accelerate this type of processing. The presented manuscript assess and compares the performance of different parallelization frameworks and different multicore architectures for exploiting parallelism when Maxwell´s Equations have to be solved. The simulation of a plane wave propagation was used to assess the parallelization frameworks and options by using different application programming interfaces (APIs) and compilers. Moreover, implementations using these different approaches were executed in different CPU and GPUs architectures. Experimental results show that different frameworks led to different performance and, as expected, the GPUs achieve higher speedups when compared to CPUs. However, results also show that the recent GPU architecture, Kepler, did not outperform the older Fermi architecture.