Cray's XK6 Supercomputer Promises 50-Petaflop Performance

Slideshow: Government's 10 Most Powerful Supercomputers

Cray this week released its fastest and most advanced supercomputer, which the company said can achieve a staggering 50 petaflops, or 50 quadrillion floating point operations per second.

The Cray XK6 Supercomputer handles it with a combination of AMD's 16-core, x86 Opteron 6200 Series server processors with Nvidia's Tesla 20-Series graphics processors and Cray's Linux-based operating system called Gemini interconnect. It is the first time Cray incorporated GPUs into a high-performance supercomputer system.

Cray said it could have released the XK6 earlier, but that it took about a year and a half to create the software that could integrate the GPUs. The software of the XK6 consists of a Linux OS based on SUSE enterprise OS 11, layered with Cray's own programming, Barry Bolding, vice president of Cray's product division, said in an interview.

"The whole package of hardware and software becomes the supercomputer," Bolding said. "Cray has a very strong belief that it isn't just about putting together a bunch of pieces of hardware. If you look at Chinese GPU systems over the last few months, we haven't seen a lot of scientific results coming out of those systems. That is probably because those systems, while they're put together with a bunch of very powerful hardware, don't have the software to run scientific applications at scale. That's really where our advantage is."

Earlier systems, such as the Cray XT4, Cray XT5, Cray XT6, or Cray XE6 systems, can be upgraded to a Cray XK6. It can be configured in a single cabinet with tens of compute nodes and expanded to a multi-cabinet system with tens of thousands of compute nodes. This is precisely how Cray has laid the claim of 50 petaflops--if enough cabinets are installed this level can be reached.

Cray's first customer for the new Cray XK6 system is the Swiss National Supercomputing Centre (CSCS) in Manno, Switzerland, which is upgrading its current Cray XE6m system, nicknamed "Piz Palu." The CSCS uses the supercomputer to support scientific research in fields such as weather forecasting, climatology, chemistry, physics, material sciences, geology, biology, genetics, experimental medicine, astronomy, mathematics, and computer sciences.

"Given the remarkable interest in GPU technology from the Swiss computational science community, it is essential that CSCS adopt this technology into its high-end production systems soon," Thomas Schulthess, director of CSCS, said in a statement. "However, we are not looking for another GPU based stunt to place high on any Top 500 lists. The Cray XK6 promises to be the first general-purpose supercomputer based on GPU technology, and we are very much looking forward to exploring its performance and productivity on real applications relevant to our scientists."

The Top 500 list Schulthess mentioned is usually revealed at the International Supercomputing Conference, which takes place in June in Hamburg, Germany. Last year, China took the top prize for fastest supercomputing with its Tianhe-I, beating out Cray's XT5 supercomputer at the Oakridge National Lab. The XK6 will not be competing at this year's conference however, since it will become available in the second half of 2011, Bolding said.

However, Bolding made a distinction between the Linpack benchmark used to assess processing speed and actual applications. He said that while the Chinese supercomputer has claimed the prize as the highest processing speed, it has yet to produce a single application.

Analyst Earl Joseph, vice president for high performance computing at IDC, said the new Cray supercomputer further establishes the company as one of the top producers of supercomputers in the world.

"We think Cray is historically very good at designing machines and this continues the same design points Cray has historically had," he said in an interview. "It has a lot of good software and is much easier to use with high performance computing. For everything we've seen its another very successful machine for Cray."

Still, next year's Top 500 list of the world's most powerful supercomputer will list systems capable of performing calculations as much as eight times faster than the fastest computers available today. IBM will have its own 20-petaflop system at the Lawrence Livermore National Laboratory in Livermore, Calif., in 2012.

Vendors are fighting it out in the market for integrated network, computer, and storage systems. In the new all-digital issue of Network Computing, we go ringside to help you pick a winner. Download the issue now. (Free with registration.)