The quest to build an exascale system that's 1,000 times more powerful than the petaflop systems being deployed today may be the biggest challenge yet in HPC. It requires new programming models and methods to manage data and memory, along with improved system resiliency.
HPC researchers are cooperating internationally, to various degrees, on developing exascale system software.
William Harrod, research division director in the advanced scientific computing in the Department of Energy's Office of Science, told attendees at the SC12 conference that this international collaboration is needed. "I personally believe there is no way to achieve these goals [of building an exascale system] by any one government, one country - it far exceeds what people are going to invest and also exceeds the technical talent, so collaboration -- that's easy," said Harrod.
"The competition is not on the computer systems," said Harrod. "The competition is on the science that you perform on the systems and what you do with that."
The U.S. has not yet funded its Exascale Computing Initiative nor has it put a price tag on it, although it's expected to cost billions of dollars. Congress is expected to get a budget request for exascale system development in the 2014 fiscal year budget, which begins next October.
Addison Snell, CEO of Intersect360 Research, said the development of petaflop systems points to what may happen with regard to exascale.
The first petaflop system, IBM's Roadrunner at Los Alamos National Lab, was a custom, hybrid design. It was soon surpassed by China's Tianhe-1A, which relied heavily on accelerators to achieve its high Linpack benchmark score, said Snell.
Although some could argue that Cray's Jaguar at Oak Ridge National Lab was more efficient and more productive at the time, "the public perception was that the U.S. had lost its lead in supercomputing. This was exacerbated when Japan's K Computer became the first to break the 10 petaflop barrier," he said.
The newest top-dog system, the 20-petaflop Titan at Oak Ridge, "brings the U.S. back to the top, as the first U.S. system in that top echelon that relies heavily on accelerators to hit its number," said Snell.
"I believe the Chinese can and may build an exaflop computer by the end of the decade," said Snell. "The U.S. may intend to wait for a more sophisticated design, but it will have to deal in the meantime with the public perception that China will have passed us by."