High performance , high accuracy FDTD implementation on GPU architectures

1. Introduction Simulations have become an integral part of a product design workflow, as devices are becoming more complex and development cycles are shrinking to be more competitive and meet market demand. In the last two decades, the capabilities of electromagnetic simulation techniques have vastly expanded to simulate increasingly more realistic and multifunctional devices for improved design, testing and production efficiency. Most simulation software relies on the Finite Difference Time Domain (FDTD) technique [1] for solving electromagnetic problems in a wide range of areas, including microwave and antenna engineering, bioelectromagnetics, electromagnetic compatibility, photonics, biomedicine and ultra low frequency applications. Simulations now require significant amounts of memory and processing power to generate results in a reasonable time frame. The emerging methodology applies the simulated model as the basis for an iterative design process, whereas, simulations were previously used for verification only. In the past, computing power and long processing times ranging from hours to weeks had always been limiting factors for engineers; however, the advent of cheap supercomputing has vastly increased the computational power, which in turn has increased accuracy and speed, and allowed solutions to complex problems that previously would have been impossible. In particular, the specialized microprocessor Graphics Processing Unit (GPU) has prodigious computational capabilities optimized for 2D and 3D technology, tailor-made for the computer game and video industries. The GPU technology for High Performance Computing (HPC) architectures features massive parallel processing cores with abundant onboard memory, packaged into a footprint that fits inside a workstation. Although other platforms, including the Cell processor and the field programmable gate array (FPGA), have attempted to bring HPC technologies to the EM market, they were largely unsuccessful due to the high deployment