Research Briefs

Modeling the atmosphere at video game speed

Sky with towering cumulus clouds

Software to simulate cloud microphysics may get a boost from speedy graphical processing units used in video games.(Image ©UCAR.)

NCAR researchers are looking at how new generations of super-fast Graphical Processing Units (GPUs) and multicore chips, such as the Cell Broadband Engine, can be used in atmospheric models. GPUs, which are very efficient at manipulating and displaying computer graphics, are used in gaming systems such as Microsoft Xbox, Sony PlayStation, and Nintendo Wii—as well as in the PC on your desk. Until now, they haven’t been applied to atmospheric modeling, although in the past few years researchers have been exploring their applications in other scientific fields that require high levels of computing power.

“We’re basically taking non-graphics applications and trying to exploit the computation power of graphics processors,” explains John Michalakes (ESSL/MMM).

John is working on rewriting parts of the Weather Research and Forecasting model (WRF) code to be compatible with GPUs, which have the potential to run much faster than standard chips, in addition to being cost-effective. A GPU-accelerated kernel of the code that John completed runs 40 times faster than the host model. Though still experimental, this microphysics kernel is available to the community, making it the first community atmospheric model with GPU acceleration as an option.

John plans to incrementally adapt more WRF modules for GPU acceleration. “There’s real potential for applying these to the more computationally intensive parts of our code,” he says.

In CISL, Rory Kelly and José Garcia are also recoding atmospheric models for GPUs. In addition, they’re developing general-purpose mathematical libraries that modelers can use on accelerators in the future. “One of our goals is to explore the technology and figure out where it’s useful for atmospheric computation science, and another is to accelerate some specific models,” Rory says.

When they recoded one computationally intensive part of HAO’s MERLIN model, it ran 10–20 times faster than the baseline code. They also worked on the radiation calculations in an older version of the Community Atmosphere Model (CAM), speeding them up by about 15–20 times.