SCD News > SCD story/photo of the week: December 27, 2004

NCAR's Scientific Computing Division researches ultimate algorithms, computer architecture

Sidebar:  The fastest machine in the world

With BlueGene/L, IBM has taken a different approach to computer architecture," says Scientific Computing Division computational scientist Steve Thomas. "They've pulled back on the clock speed and put dual processors on a chip. The processors are not as fast, but they can be more tightly packed, so they won't melt. It would be as if you could take all of blackforest, NCAR's IBM SP system [24 large frames containing a total of 1,236 processors], and put them into a single cabinet."

In point of fact, 1,024 BlueGene nodes (2,048 processors) can be configured on one 19-inch rack, delivering a peak speed of 5.6 teraflops and consuming about 25 kilowatts of electrical power—a small fraction of that required by other microprocessor-based systems.

More processors in a smaller space means a vast scalability. A beta version of BlueGene/L, now running at the IBM facility in Rochester, New York, with 32,768 processors, is already ranked by the Top500 Supercomputer Sites as the fastest machine in the world. The full machine, being built at Lawrence Livermore National Laboratory in California, will have 131,072 processors. With a peak speed of 367 teraflops, it will be roughly ten times faster than the Earth Simulator in Japan.

"This is the next step in the competition between vector processors and microprocessors," Steve notes. "The Earth Simulator is a vector processor. Its individual processors are powerful, but they're more expensive, involve special construction, and consume a lot of power by comparison.

"Scientists often feel that vector machines are more desirable; they say we need vector machines. Well, do we? BlueGene/L [which runs on relatively cheap, commodity microprocessors] is clearly already faster than the Earth Simulator. It's beating everything. It can integrate a realistic but simplified climate model at 1.5 teraflops, sustained—and that's without even doing all the optimizations possible, because we've just started using the machine.

"The question for us is, how could we possibly use tens of thousands of processors—is it even possible? How can we do the necessary science on these new machines?" Steve asks. "This is what is interesting to SCD, to explore this technology."

 Back to "NCAR's Scientific Computing Division researches new algorithms, computer architecture"