Forget Terascale. Here Comes Exascale Computing.

Albuquerque (NM) - Ok, let's not exaggerate. With our everyday PCs, we are still stuck in gigascale and are still waiting for terascale to arrive. But at least scientists will soon have access to petascale capabilities as we learned yesterday and now we know that groundwork on exascale system is already being done.

Sandia and Oak Ridge national laboratories said they have begun developing a concept for the next generation of supercomputers - supercomputers that will be able to analyze an enormous amount of particles in real time to examine and predict real world conditions.

The raw computing capability of today's supercomputers is still described as terascale, an Era that has begun about 10 years ago. For example, The TACC announced its 504 TFlops Ranger supercomputer, the world's second most capable system of its kind, just yesterday. The world's fastest supercomputer is IBM's BlueGene/L with a peak performance of 596 TFlops.

To put this number crunching horsepower into perspective, typical quad-core PCs today are estimated to deliver about 50 GFlops, which means that Blue-Gene/L is about 12,000 times faster than what some of you may use when reading this article. High-end graphics cards crank out up to 500 GFlops today, but they lack the memory capacity and bandwidth of supercomputers and are not quite comparable.

A 1 PFlops computer, which may become a reality within the next year or two, will be 1000 times faster than a 1 TFlops system. Now scientists are talking about exascale systems, which will be 1000 times faster than Petascale computers. A 1 EFlops supercomputer will be able a million trillion calculations - every second.

To achieve this goal, scientist have begun laying the groundwork and laying out general requirements for such a system. The basic idea, you guessed, it are "novel and innovative computer architectures" that will be able to close critical gaps between theoretical peak performance and actual performance.

A key challenge will be the growing mismatch between data movement and processing speeds. "In an exascale computer, data might be tens of thousands of processors away from the processor that wants it," says Sandia computer architect Doug Doerfler. "But until that processor gets its data, it has nothing useful to do. One key to scalability is to make sure all processors have something to work on at all times."

Despite the fact that supercomputers always have worked with highly parallel software, multi-core processors are bringing new problems for scientists as well, especially when there are dozens of cores on one die.

"In order to continue to make progress in running scientific applications at these [very large] scales, we need to address our ability to maintain the balance between the hardware and the software," said Jeff Nichols, who heads the Oak Ridge branch of the institute. "There are huge software and programming challenges and our goal is to do the critical R&D to close some of the gaps."

A rather obvious problem is the power consumption such computers. Many organizations are already planning their supercomputers around available power needs and are even building these systems close to power plants as Petascale systems easily can consume 5 MWwatts and rake up millions of dollars in power bills. The researchers said that just because of the sheer cost they "want to bring that down." As of now, there was no information on how this could happen.

Sandia's ASCI Red was the first Teraflop system. Unveiled in 1997, it used 10,000 Pentium processors to deliver 1 TFlops - or one trillion mathematical operations per second. Exascale computers are beginning to take shape, but we clearly will have to wait several years until we will have a better idea how these systems will look like.

Background: NCSA: A look inside one of the world's most capable supercomputer facilities