Performance and Scalability of the JCSDA Community Radiative Transfer Model (CRTM) on NVIDIA GPUs

An atmospheric radiative transfer model calculates radiative transfer of electromagnetic radiation through earth´s atmosphere. The community radiative transfer model (CRTM) is a fast radiative transfer model for calculating the satellite infrared (IR) and microwave (MW) radiances of a given state of the Earth´s atmosphere and its surface. The CRTM takes into account the radiance emission and absorption of various atmospheric gasses as well as the emission and the reflection of various surface types. Two different transmittance algorithms are currently available in the CRTM OPTRAN: optical depth in absorber space (ODAS) and optical depth in pressure space (ODPS). ODAS in the current CRTM allows two variable absorbers (water vapor and ozone). In this paper, we examine the feasibility of using graphics processing units (GPUs) to accelerate the CRTM with the ODAS transmittance model. Using commodity GPUs for accelerating CRTM means that the hardware costs of adding high-performance accelerators to computation hardware configuration are significantly reduced. Our results show that GPUs can provide significant speedup over conventional processors for the 8461-channel IASI sounder. In particular, a GPU on the dual-GPU NVIDIA GTX 590 card can provide a speedup 375x for the single-precision version of the CRTM ODAS compared to its single-threaded Fortran counterpart running on Intel i7 920 CPU, whereas the speedup for 1 CPU socket with respect to 1 CPU core is only 6.3x. Furthermore, two NVIDIA GTX 590s provided speedups of 201x and 1367x for double precision and single precision versions of ODAS compared to single threaded Fortran code.