Modern graphics processing units offer enticing speedups for atmospheric modeling. However, accelerating large and complex models using GPUs is not a straightforward task. To explore the potential for using GPUs to accelerate an atmospheric model, I ported an expensive portion of the CAM radiation code to an Nvidia GPU. I will talk about performance that can be achieved for this test case, what it means in the context of a full model, current limitations to using GPUs, and future trends and technologies that will help overcome these limitations.
Rory C. Kelly is a Software Engineer in the Consulting Services Group. While studying physics at the University of Colorado, he accidentally learned Fortran, which eventually landed him a job at NCAR in 2001. He specializes in benchmarking and performance testing, and in his spare time he likes to fiddle with weird computing hardware, including GPU, FPGAs, and other computational accelerators.