Hybrid Single/Double Precision Floating-Point Computation on GPU Accelerators for 2-D FDTD
Graduate School of Information Sciences, Tohoku University, Japan
International Conference on Parallel and Distributed Processing Techniques and Applications (PDPTA), pp.1001-1002, 2012
@article{waidyasooriya2012hybrid,
title={Hybrid Single/Double Precision Floating-Point Computation on GPU Accelerators for 2-D FDTD},
author={Waidyasooriya, H.M. and Takei, Y. and Hariyama, M. and Kameyama, M.},
year={2012}
}
Acceleration of FDTD (Finite-Difference TimeDomain) is very important in computational electromagnetic. We propose a hybrid single/double precision floating-point computation to accelerate FDTD on GPUs. We apply single-precision when the dynamic range of the electromagnetic field is low and double-precision when the dynamic range is high. According to the experimental results, we achieved over 35 times of speed-up compared to the CPU implementation and over 1.79 times speed-up compared to the conventional GPU acceleration.
January 24, 2013 by hgpu
