How China may spur U.S. supercomputing

Computerworld - NEW ORLEANS -- U.S. supercomputing dominance is being challenged in ways not seen before, and that may be the best thing to ever happen to this field, particularly in Washington's climate of cost-cutting.

Of the top four systems on the semi-annual Top500 list of the world's leading supercomputers, which was released this week, two are in China: the top-ranked Tianhe-1A, at 2.5 petaflops, and the No. 3 Nebulae system. Japan has the No. 4 system. The U.S. is in second place with the Cray XT5 Jaguar system at 1.75 petaflops. The announcement came as experts in supercomputing were gathering in New Orleans for the SC10 conference.

Addison Snell, CEO of InterSect360, a high-performance computing research group, said Asia's showing in supercomputers will get the attention of political leaders.

"When it's all over the popular press that three of the top four supercomputers in the world, according to how they measure it, are in Asia, there is no way there is not a response in Congress to that," Snell said.

Earl Joseph, a high-performance computing analyst at IDC, said "global competitiveness" will drive HPC but it isn't about building the most powerful system in the world.

"The Chinese are not doing the old traditional supercomputing war," Joseph said. Instead, China's government is building 14 different petascale computing centers "because they recognize the competitive value of that."

Even Russia has realized that its products won't be as competitive without high-performance computing, Joseph said, a reference to Russian President Dmitry Medvedev's scalding criticism last year of his country's progress in supercomputing.

"I don't know what the budget situation will turn out to be in Congress," said Andy Keane, the general manager of Tesla business at Nvidia Corp., but supercomputing technology "is directly related to the progress we make on the economy."

Supercomputing allows research to simulate environments; the more powerful the computer, the larger and more detailed the simulation. For example, with the right capacity, a computer could simulate the working of a human cell at the atomic level. Supercomputers can also help manufacturers speed product development by allowing engineers to design, change and test products in virtual environments before producing physical prototypes. The largest systems are typically built by governments.

Nvidia's work on high-performance computing is supported by the U.S. government through a research contract with the Defense Advanced Research Projects Agency (DARPA), Keane said.

DARPA, in a request for proposals from vendors earlier this year, said that current processing systems are "grossly power-inefficient" and that technology trends "have reached a performance wall."

One of the means for solving this has been to turn to graphics processing units, which are roughly about 10 times more efficient than CPUs, said Keane. The DARPA contract is helping Nvidia to devote more research to such problems, he said.

Keane said it is possible to improve GPU efficiency by "another factor of about 100" to deliver exascale performance, which would make it possible to build a system that is roughly 1,000 times more powerful than a petaflop system.

"The computer that will be the fastest will be the computer that has the lowest energy per operation," he said.

Keane said the goal of developing an exascale system will be achieved by 2018, if not before, and he predicted that by 2014 or 2015 there will be systems that are one-third to halfway there, meaning in the range of 500 petaflops.

Nivida's GPU technology was used in the top-ranked Chinese and Japanese systems. The U.S. system, Jaguar, built entirely of CPUs, uses about 7 megawatts. In comparison, the GPU-based Chinese Tianhe-1A system uses about 4 megawatts.

Systems that incorporate accelerator technology are spreading rapidly. Silicon Graphics International this week announced a new system that was built from the ground up with "optimized accelerator architecture," according to Bill Mannel, vice president of product marketing at SGI.

Mannel said the approach used to build that system, along with its relatively small size, puts it in range of an exaflop.

The SGI Prism XL, code-named Mojo, can deliver approximately 1 petaflop from a cabinet that is about the size of three standard 19-inch racks. It can support accelerator cards from Nvidia, Advanced Micro Devices and Tilera. The price range isn't being disclosed.

Despite the change in the top supercomputing rankings, the U.S. maintains a sizable lead in supercomputing. Of the 500 top supercomputers in the world, 274 are in the U.S.

Patrick Thibodeau covers SaaS and enterprise applications, outsourcing, government IT policies, data centers and IT workforce issues for Computerworld. Follow Patrick on Twitter at @DCgov, or subscribe to Patrick's RSS feed . His e-mail address is pthibodeau@computerworld.com.

Read more about High Performance Computing in Computerworld's High Performance Computing Topic Center.