TECHNOLOGY FOR CHANGE

Knowing How to Play the Chemical Analysis Game

Stan Gibson | Date: 07-21-09 | Comments

Powerful graphics processors score big in scientific computing.

Playtime�s over. The graphics processors that produce life-like effects for the latest video games have been harnessed for the serious work of chemical analysis. By rewriting algorithms originally crafted for conventional CPUs to run on the graphics processors (GPUs) of video game consoles, researchers at Stanford University and the University of Illinois have sped up processing by 650 times.

GPUs are suited to the task of scientific computation because of inherent parallelism of their architecture. For example, the Nvidia chips used in this work each have 240 computing cores. But putting each core to work simultaneously on the task of analyzing molecular structure was not easy.

�GPUs have a large number of cores, but they are simple, so you have to find out how you can program them. It took several months and a lot of experimentation to rewrite the software. There is still not a lot known about the tricks for doing this,� said Professor Thom H. Dunning Jr. at the National Center for Supercomputing Applications (NCSA) at the University of Illinois in Urbana, Ill., and lead chemist of the project.

Todd Mart�nez, a professor of chemistry at Stanford University and Stanford graduate student Ivan S. Ufimtsev carried out much of the work on the project, rewriting the molecular design program called GAMESS to calculate the structures of molecules ranging from the 24-atom caffeine molecule to the 453-atom olestra molecule.

A catalyst that enabled the work was Nvidia�s release of Cuda (Compute Unified Device Architecture), an extension to the C programming language that lets the CPU and GPU work together on compute-intensive tasks. The trial ran on three Nvidia GeForce GTX 280 GPUs, but the aim is to run on many more, piling parallelism upon parallelism for far higher performance, said Dunning.

With the chemical analysis success on the books, many other tasks can be tackled far more efficiently than is now possible. Mathematical libraries, for example, will be adapted to run on GPUs. �This is going to become much more common in the future than it is now,� said Dunning.

Funding for this research is provided by the NSF Divisions of Chemistry and Materials Research.