Sunday, February 24, 2008

1,000 trillion floating-point calculations (flops) per second

Scientists have unveiled a new initiative, dubbed the Institute for Advanced Architecture, to lay the groundwork for a supercomputer that would be more than 1,000 times faster than any current offering.
Commercial supercomputer makers have recently begun to flirt with petaflop performance, meaning computers capable of completing 1,000 trillion floating-point calculations (flops) per second. The Sandia and Oak Ridge national lab scientists aim to leapfrog that benchmark by several orders of magnitude and are targeting one million trillion calculations per second, known as exascale computing.
(Exa is the metric prefix for quintillion, or 1018.)
"Both the [Department of Energy's] Office of Science and the National Nuclear Security Administration have identified exascale computing as a critical need in roughly the 2018 timeframe," said Sudip Dosanjh, the project's head. "We certainly think that there is a national competitiveness issue."
Ultrafast computers are integral to simulating complex systems, like the Earth's climate, nuclear warhead explosions or the protein interactions inside cells. They continue to progress, thanks to the well-known -- though often questioned -- Moore's Law, which has allowed chip makers to pack twice as much power into the same amount of space about every two years. More power has meant more so-called flops, a common measurement of computing speed. Ten years ago, Sandia's ASCI Red became the first teraflop computer, and in December 2000, Wired called 100-teraflop performance "unheard of."
Now, though, new challenges have presented themselves. The researchers say that moving data from the supercomputer's thousands of processors into its memory will require them to design new architectures that reduce the need to move data around.
"Some people say that flops are almost free, that really what you are paying for is moving the data," Dosanjh said.
In addition, power and reliability require new solutions when you've got thousands or millions of processors.
"The power budget for all computers seems to be going up rapidly. We need a machine you can afford to run," Dosanjh said, and one that actually works. With a million computing nodes working together, the odds are high that one of them will break, over the course of even a small calculation.
With current technologies, "an exascale computer might only stay running for a few minutes," said Dosanjh.
The Sandia-Oak Ridge collaboration has $7.4 million in fiscal year 2008 funding from the National Nuclear Security Administration and the Department of Energy, but it's not just nuclear weapons research that is driving the push for faster supercomputers. Researchers of many stripes have come to depend on the inexorable upward scaling of computing power.
Gavin Schmidt, a climate modeler at NASA Goddard, said that he's built the regularity of computational upgrades into the way he designs his climate simulations, which are so computing-intensive they can take several months of processing to complete.
"Generally speaking we don't do experiments that last more than three months," Schmidt said. "If you want to do an experiment that would last for six months, it's best to just wait a few months, and then [with faster computers] it only takes two months to run."
According to a semiannual list of the world's top 500 supercomputers, compiled in November 2007, IBM's BlueGene/L System is the fastest computer in the world, with benchmark performance of about 480 teraflops per second, or almost half a petaflop. That rig is a joint development of IBM and the National Nuclear Security Administration, and is housed in California's Lawrence Livermore National Laboratory.
With the research team trying to vault several orders of magnitude over any current system, Dosanjh said the new institute would need $20 to $30 million a year to accomplish its goals.
Even as individual supercomputers have grown in speed, distributed-computing initiatives, like the Folding@Home program, have enabled researchers to tap into thousands of users' computers and PS3s to solve some types of scientific problems.

No comments: