by Chris HillMassachusetts Institute of Technology
Performance of geoscience numerical computations is often constrained by memory bandwidth. This makes programmable GPUs attractive, as they have the potential to deliver several times more memory bandwidth than a conventional CPU, for algorithms that contain enough inherent data parallelism.
This talk highlights two geoscience projects, one from hydrology and one from oceanography for which we are deploying GPUs. The ocean application employs a hybrid, multi-scale algorithm in which concurrently executing GPU and CPU computations, solving fine-grain local and coarse-grain global equations respectively, are coupled together. The hydrology application uses GPUs to accelerate the solution of watershed area computations at high resolution. Both examples embrace algorithmic as well as implementation innovations to better exploit GPU characteristics. In both cases, reasonable comparisons with other approaches, show GPUs providing considerable speed-up.