The graphics processor units (GPUs) used in playstations for games like Mortal Kombat are engineered to stream fast calculations in parallel to display pixels on TVs or computers. In fact, for certain applications, they run (per GPU) at least 15-20 times as fast as the CPUs that drive our laptops and parallel computing machines!  Why not use them for running science applications? This idea has been out there for awhile and has been highly developed by the research group of Vijay Pande of Stanford University. He developed “Folding at Home,” using hundreds of thousands of idle home computers to carry out simulations of the folding of proteins from random conformations to well defined structures. Now his group is using hundreds of thousands of idle playstations.

Following this lead, three graduate students at the University of California – Davis, Jonathan Lawton, Robert Hayre, and Jesse Singh, supervised by ICAM Co-Director Daniel Cox, have set up a small (starting with 12 GPUs) but growable cluster of GPU-based computers which will soon be available for scientific applications by ICAM scientists. They tested the cluster on various molecular dynamics simulations—applying Newton’s laws of motion to model forces derived from quantum mechanics to study the stability of biological molecules in various conformations. Lawton, Hayre, and Singh and their mentor, Daniel Cox, anticipate that the cluster will also be useful for quantum Monte Carlo simulations of cold atom problems or for work on magnetic domain/interaction problems in magnetic recording. The OpenMM (Open-source Molecular Mechanics) software and Zephyr graphic user interface molecular dynamics from the Stanford group mentioned above will be installed and available for users.

Preliminary runs have been encouraging;   Jonathan Lawton carried out “implicit solvent” runs for a hypothesized conformation of the 108-amino-acid protein URe2 from yeast. This protein is known to act in a non-impactful manner as a “prion” protein, like those in mad cow disease. Implicit solvent means that the water is not treated as a collection of molecules, but rather as a continuum which can respond to the electrical forces from the protein.  Jonathan found a 15 fold increase in running speed on the GPU when compared to runs on a state-of-the art 64-bit Opteron CPU. The results were comparable in quality to those from far slower all atom explicit solvent runs. A key result, which should excite graduate students in theoretical biological physics: 10 nanoseconds of simulation time on one GPU took 1.5 hours. Ordinarily, with explicit molecular treatment of water, this can take days. This opens up the prospects of ‘’high- throughput in silico proteomics’’ to test models of protein structure and folding. Graduate student Robert Hayre is rewriting a “homegrown” molecular dynamics code in the CUDA framework (C++) for GPUs and writing a blog about his experiences.

The group has started a wiki/blog page at the ICAM web pages. You can look up the specs for the system at http://icamconferences.org/icam_dev/index.php/icam_infrastructure/Hardware/

We will be opening up pages for applications to use the cluster beginning Dec. 1, 2009. Watch the page
http://icamconferences.org/icam_dev/index.php/icam_infrastructure/ for details.

We hope to report in future issues of ICAMNews on the projects undertaken using this frontier approach to scientific computing.

By Daniel Cox and Karie Friedman, ICAMNews, October 2009