GPU accelerated FDTD based Open-source SAR Simulator

Specific Absorption Rate (SAR) is a widely used metric to measure the amount of non-ionizing EM radiation power absorbed by a unit mass of biological tissue, thus it helps in determining the safety standard of any EM applications. The 3D Finite Difference Time Domain (FDTD) is a commonly used computational technique for accurate SAR computation. However, it is computationally as well as memory intensive and requires high-performance computing(HPC) clusters to handle problems of large sizes encountered during high-frequency applications. This paper describes the implementation of a fast, optimized Open Source GPU accelerated FDTD based SAR calculator (available at Github)using CUDA (Code Unified Device Architecture) as an alternative that can be developed locally at a relatively small cost. On a testbed comprising a serial implementation on Intel i7-4790 CPU and parallel implementations on NVIDIA Tesla K40 graphics card, we have achieved a speedup of up to 45x in double precision mode. We have investigated the effect of increasing the frequency (larger problem sizes) and threads per block on speedup. Finally, we have compared the performance of our parallel implementation on two different GPUs (Tesla K40 and GTX 670) and analyzed the effect of hardware architecture on the performance of the SAR calculator.