New scientific programming tools developed at Duke University can help scientists tap the processing power of graphics processing units, or GPUs, from video gaming systems to handle complex calculations, such as better understanding the body’s response to experimental treatments for AIDS and other diseases.

Using GPUs, researchers now can complete complex statistical analyses on huge datasets orders-of-magnitude faster than on standard central processing units (CPUs). The speedup is a critical enabler for a range of biological studies, said Mike West, a statistical science professor at Duke.

For example, immunologists who take multiple blood samples from dozens of patients before and after experimental treatments need to analyze many proteins on millions of cells to understand the immune system response.

“Using standard CPU computing, the job would run on a single computer for two weeks," West said. "Now we can do it in a couple of hours."

With his Duke research group and UCLA colleague Marc Suchard, West has developed software to make it easier for scientists to harness the processing power of GPUs without having to learn a complicated programming language.

Unlike CPUs, GPUs have dozens or hundreds of small “cores” or engines capable of doing lots of little things at once, which makes them especially good for crunching away at what researchers call “embarrassingly parallel” problems.

Low-cost GPU hardware will bring high-performance data analysis back to the desktop for many scientists who previously relied on large-scale computing clusters, West said.

“Thousands of labs are generating this type of data, and experimental biologists are used to having their computer in the lab. They don’t care what a GPU is or how to use it, but they may need to customize it,” West said. "We need this as we’re looking to do lots of routine analyses with increasingly large datasets.”

West’s software aims at minimizing the bottlenecks between iterations of calculations. He compares the data flow to the stream of runners who compete in the Chicago marathon. At the start, thousands race side by side through the wide streets, but when the route narrows to a single passageway, runners must go through one at a time, which slows the entire process.

As researchers in more disciplines recognize the potential of GPU computing, West predicts new synergies between scientific research and the computer gaming industry.

"The next Xbox will be a scientific engine as well as a gaming engine,” said West, who now is experimenting with the potential for clustering multiple GPUs. “By embracing the technology to understand how it can help in research, we are also feeding back to the people who drive the technology changes.”

West’s collaborators at Duke include Cliburn Chan (biostatistics), Quanli Wang (Institute for Genome Sciences & Policy) and graduate students Andrew Cron (statistical science) and Jacob Frelinger (computational biology and bioinformatics).

The paper was published online in the Journal of Computational & Graphical Statistics, aAll code and additional resources are available at Duke University.

The research was supported in part by grants from the National Science Foundation and National Institutes of Health.