Price-performance aspects of accelerating the FDTD method using the vector processing programming paradigm on GPU and multi-core clusters

The parallelization of the FDTD on GPUs has become popular due to the low cost, low power and high compute performance achieved with these devices. In recent years, manufacturers of multi-core processors have enhanced the vector processing capability inherent in conventional processing cores, to the extent that these are now contributing considerably to the acceleration of the FDTD and competing with GPUs. This paper will compare the power consumption and purchase cost versus the performance benefits of several parallel FDTD implementations, in order to quantify the effect of parallelizing the FDTD using various processing paradigms. The purchase cost of hardware, computational performance and power consumption are used to compare the parallel FDTD deployments on the BlueGene/P, GPU clusters and the multi-core clusters using SSE. It is shown that the deployment of the parallel FDTD using a hybrid programming paradigm achieves the best computational performance for the lowest purchase cost and power consumption