Performance Optimization of Massively Parallel FDTD Computations

Abstract

The advancements of General Purpose Graphic Processing Units GPGPUs, have paved the way for computationally intensive scientific calculations to be done on an off the shelf massively parallel graphic processors GPUs, rather than the use of expensive solutions such as High Performance Clusters HPCs or Supercomputers. In this paper the NVIDIA’s Compute Unified Device Architecture CUDA on an NVIDIA GeForce processor will be used to solve a Finite Difference Time Domain FDTD computation. An FDTD computation has unfeasible running time using a single processor, and attempts to use multi-core CPUs have been made, but with the high overhead of network traffic in HPCs or synchronizations among cores. Multiple attempts have been made to utilize GPU for solving FDTD computations; but in this paper the focus will be on optimizing the efficiency of the algorithm by maximizing the throughput through affective use of the fast on-chip shared memory, and avoid using the slow off-chip global memory.