Georgia Tech researchers building an experimental new supercomputer say graphics processors may help pave the way toward future exascale machines, which would be 1,000 times faster than today's most powerful supercomputers.
How to build your own supercomputer
Georgia Tech was recently awarded a five-year, $12 million grant from the National Science Foundation to build a supercomputer that combines HP's Intel-based processors with graphics processing units from Nvidia.
Graphics processors are not typically used at large scale in the supercomputing world because they generally lack error correction and the mathematical precision found in x86 processors, explains Jeffrey Vetter, a Georgia Tech professor and computer scientist at the Department of Energy's Oak Ridge National Laboratory.
Nvidia is solving these problems in its next generation processors, code named Fermi, which come out early next year, he says. Fermi features so-called "double precision math," a level of accuracy required by NSF and Department of Energy computational science workloads, he says.
"The reliability and double precision make Fermi an important part of this and make it easier to build a scalable system," Vetter said during an interview with Network World at the SC09 supercomputing conference in Portland, Ore.
The world's fastest supercomputers have hit the petaflop scale, performing one thousand trillion calculations per second. Researchers believe that within a decade supercomputers will be capable of performing at exaflop speed, 1,000 times faster than a petaflop.
Vetter says future exaflop machines will likely contain a heterogeneous mix of processors, and he thinks graphics processors could be among the major components. Graphics processors may not work with as many types of applications as general purpose x86 processors, but they can accelerate many workloads and offer space and power savings that could be crucial in building larger systems, he says.
"There is a growing emphasis on power performance," Vetter said. "You can literally, for some applications, take out two, four or eight regular x86 processors and replace them with one GPU (graphics processing unit) and get the same amount of work done."
Georgia Tech's machine will be deployed at the National Institute for Computational Sciences at Oak Ridge. An initial system to be set up next year will have performance of about 200 teraflops, and a successor machine planned for 2012 will operate at speeds of 2 petaflops.
The computer will be used by the NSF but there will be opportunities for outside organizations to give it a test drive.
"We have time reserved for education and industry users to come in and try things," Vetter said. "So if you're from an oil company, finance company or a car company you can … see if graphics processors accelerate your workload."
Highly parallel workloads in fields such as molecular dynamics and biological simulations may be ideal for the Nvidia-based supercomputer, he says.
Just a year ago, graphics processors were considered a fringe part of the supercomputing world, but are starting to catch on, according to Vetter.
"What we want to demonstrate is we can continue to solve scientific problems with a much more power efficient architecture that will help us get to exascale and not encounter those power problems," he said.