GPU computing for atmospheric modeling: Experience with a small kernel and implications for a full model

Modern graphics processing units (GPUs) offer the potential for extremely high performance floating point computation. Many important kernels, as well as certain types of applications, have been greatly accelerated by being ported to and run on GPUs. Atmospheric modeling is extremely computationally expensive and model designers are eager to speedup atmospheric models with the power promised by GPUs. However, the path to accelerating a large model using GPUs is not straightforward. We explore the potential of GPUs for atmospheric modeling, first by porting an expensive routine selected from the Community Atmosphere Model (CAM) to and NVIDIA 9800 GX2 using CUDA, and then by applying the experience from this exercise to thinking about what it would take to accelerate the rest of the model.