Cray taps Nvidia chips for large supercomputer
Company will use Nvidia GPUs in its largest supercomputers, employing a back-to-the-future technology that played a role in the earliest supercomputers.
Cray will put Nvidia graphics processors in future large-scale supercomputers, the companies said today.
The announcement follows this summer's jump to the No. 2 spot in global rankings of China's Nvidia-equipped Nebulae supercomputer.
At Nvidia's 2010 GPU Technology Conference today in San Jose, Calif., supercomputer leader Cray announced that it is developing supercomputers that can use Nvidia Tesla 20-Series graphics processing units.
"We're putting this technology--the next generation of Nvidia Tesla--in our large XE6 systems," Barry Bolding, vice president of Cray's products division, said in a phone interview. Cray's largest system, Jaguar, at the Oak Ridge National Laboratory, based on processors from Advanced Micro Devices, is currently ranked No. 1 in the world, according to the Top 500 list.
Cray, which already uses Nvidia GPUs in its low-end desktop supercomputers, is targeting Nvidia's supercomputer-specific Tesla processors for accelerating "modeling code" for scientific applications in its Cray XE6 product line, according to Bolding.
"We're doing this because Nvidia is starting to produce accelerators (GPUs) that are useful to our customers. They've done some enhancements that go beyond a normal graphics accelerator. A couple of years ago, those (Nvidia) GPUs did not have functionality that was appealing to our customers."
Bolding continued. "What we'll see first is the data centers that run a few key applications on the accelerators. The data center that has to run 500 applications? That's the data center that won't move over to accelerators," he said.
An interesting sidelight to all of this is that Nvidia's chips, in supercomputer parlance, are classified as vector processors--the same type of processors Cray used in its earliest supercomputers, dating back decades.
"A GPU accelerator is really no different than a vector processor. And we know that vector processors in their heyday could not run every application efficiently. There's going to be a set of applications that will be phenomenal on GPU accelerators and a set of applications that are far better on standard x86 processors," Bolding said.
Cray and Nvidia are working together on future GPU accelerator technologies as a potential path towards exascale computing, which is many times faster than today's supercomputers. The two are partnering on a team that was recently awarded a $25 million research grant from the Defense Advanced Research Projects Agency, or DARPA, as part of its Ubiquitous High Performance Computing program.
Correction at 11:45 a.m. PDT: The Nvidia Tesla chips will be used initially in XE6 systems not Cray's Jaguar system, as originally stated.