Climate Lab Logo

Publications Portal

START flash tag cloud requires Flash Player 9 or better.

Tag Cloud created by Roy Tanck

DEVELOPMENT OF NUMERICAL MODELS: Atmospheric numerical models have traditionally improved in resolution and sophistication because processors were speeding up exponentially. Over the past decade this growth has slowed, and we now rely on massively parallel machines for compute power. These machines have hundreds of thousands of processing cores, and this number is growing quickly. We have finally breached the petascale barrier (one quadrillion operations every second), but current operational algorithms are insufficient to access this power by orders of magnitude. The field of computational climate is, therefore, investing significant effort into algorithms that scale massively parallel machines.

Current Research

Our lab is currently addressing the increasing complexities of efficiency from two fronts. First, we are developing a numerical algorithm for atmospheric dynamical cores with the following properties:

We believe this method will be competitive on modern architectures, and low memory requirements means lower energy consumption for a "greener" climate simulation.

Second, we are pursuing alternative hardware by coding this algorithm for Graphics Processing Units (GPUs). Results so far are very encouraging. For the same algorithm, a high-end Nvidia graphics card was roughly 20 times faster than our lab's best dual-core CPU. This can be improved even more by running in single precision and tuning the memory usage. This is economically competitive because the ratio of performance compared to the ratio of cost is greatly in favor of the GPU.

::: Webmaster : Robert Mera :::