Accelerating the RTTOV-7 IASI and AMSU-A radiative transfer models on graphics processing units: evaluating central processing unit/graphics processing unit-hybrid and pure-graphics processing unit approaches

The radiative transfer for television operational vertical sounder (RTTOV) is a widely-used radiative transfer model (RTM) for calculation of radiances for satellite infrared and microwave sensors, including the 8461-channel infrared atmospheric sounding interferometer (IASI) and the 15-band Advanced Microwave Sounding Unit-A (AMSU-A). In the era of hyperspectral sounders with thousands of spectral channels, the computation of the RTM becomes more time-consuming. The RTM performance in operational numerical weather prediction systems still limits the number of used channels in hyperspectral sounders to only a few hundred. To take full advantage of such high-resolution infrared observations, a computationally efficient radiative transfer model is needed to facilitate satellite data assimilation. In this paper, we develop the parallel implementation of the RTTOV-7 IASI and AMSU-A RTMs to run the predictor module on CPUs in pipeline with the transmittance and radiance modules on NVIDIA many-core graphics processing units (GPUs). We show that concurrent execution of RTTOV-7 IASI RTM on CPU and GPU, in addition to asynchronous data transfer from CPU to GPU, allows the GPU accelerated code running on the 240-core NVIDIA Tesla C1060 to reach a speedup of 461x and 1793x for 1- and 4-GPU configurations, respectively. To compute one day's amount of 1,296,000 IASI spectra, the CPU code running on the host AMD Phenom II X4 940 CPU core with 3.0 GHz will take 2.8 days. Thus, GPU acceleration reduced running time to 8.75 and 2.25 min on 1- and 4-GPU configurations, respectively. Speedup for the RTTOV AMSU-A RTM varied from 29x to 75x for 1 and 4 GPUs, respectively. To further boost the speedup of a multispectral RTM, we developed a novel pure-GPU version of the RTTOV AMSU-A RTM where the predictor module also runs on GPUs to achieve a 96% reduction in the host-to-device data transfer. The speedups for the pure-GPU AMSU-A RTM are significantly increased to 56x and 125x for 1- and 4-GPU configurations, respectively. C (C) 2011 Society of Photo-Optical Instrumentation Engineers (SPIE).