Accelerating Electromagnetic Scattering Simulations with CUDA : Parallelising the Finite Difference Time Domain method for GPUs

Simulating the RADAR cross section of fighter aircraft is an exceptionally intensive computational task. However, the finite difference time domain method is a highly parallelisable scheme which in this thesis has been implemented in the C++ programming interface of CUDA to be able to run on GPUs. To simulate the RADAR cross section of a scatterer, some key aspects need to be implemented. The important ones are to parallelise the core update equations in the domain, to generate a plane wave towards a scatterer, and the reflected waves need to be absorbed at the boundaries. The fields are also transformed to see what they would look like in the far-field to simulate problems when the observer is not in the near vicinity of the scatterer. The main update equations of the finite difference time domain method were benchmarked and achieved a speed-up of 40 for larger problem sizes with double precision compared to a parallelised CPU implementation. The GPU program demonstrated a memory throughput of around 90% compared to the theoretical maximum of the hardware. The RADAR cross section of two spheres were then computed. One consisted of a perfectly electric conducting material, whereas the other had a coating of a dispersive medium with a thickness of 1 cm. The program presented good agreement with the CPU implementation for frequencies up to around 2.5 GHz, but the plane wave experienced leakage for higher frequency content most likely due to numerical dispersion. The boundaries were also not absorbing as strongly as desired which caused reflections. Therefore, abnormally thick layers were utilised. Still, acceptable results for lower frequencies achieved a speed-up of 16.3 for the complete spherical conductor and 43.2 for the dispersive sphere. The larger speed-up for the dispersive case is most likely due to a larger problem domain, which favours the parallelism of the GPU.