IE 11 is not supported. For an optimal experience visit our site on another browser.

Desktop supercomputing

If you can off-load the heavy processing for the purpose of graphics and gaming, then why not apply the same concept to create a supercomputer in a desktop?
/ Source: Forbes

The chips inside personal computers work pretty hard, but they're not very good at handling graphics.

That's why gamers make a big fuss about graphics acceleration cards from such companies as Nvidia and ATI. The add-in cards from those companies go into a PC to take over the computing tasks associated with graphics, leaving the CPU chip in the computer to do other things.

If you can off-load the heavy processing for the purpose of graphics and gaming, then why not apply the same concept to create a supercomputer in a desktop?

That's the operative question behind a startup called ClearSpeed, based in the United Kingdom and in Los Gato, Calif. The company has designed a chip it says can be added to almost any existing computer, giving it the power of a serious supercomputer.

Remember if you will that FLOP is short for floating point operations per second. A gigaflop is a billion of those operations. ClearSpeed proposes to give desktop computers the ability to do 50 gigaflops per second.

Now with the most powerful supercomputer in the world delivering 35.8 teraflops--that is, 35.8 trillion calculations per second--a mere 50 billion of these operations per second may not sound like much. But as recently as June of 2000, a computer capable of 50 gigaflops would have qualified for the list of the Top 500 supercomputing sites, maintained by the University of Mannheim in Germany. And not the bottom of the list either: A 50-gigaflop system came in at No. 358 of the list released that month.

"There's a fierce appetite for computational off-load," says Mike Calise, ClearSpeed's U.S. president. And there's generally two ways of feeding that appetite. One is to buy or build a massive--and expensive--supercomputer; the other is to build a cluster using standard computers. "But they require lots of sophisticated cooling, middleware and other things. And they're basically unreliable. We're going to bring extremely high-performance computing to the mainstream at a low cost."

At the Fall Processor Forum earlier this month, ClearSpeed announced the design of the chip that will make it happen. The chip is expected to be available in the first quarter of 2005.

"If you look back 18 months, to get a teraflop worth of computing power would have cost you a million dollars, says Calise. "A year ago, you might have gotten it for half a million. Right now, you can do it for something approaching a quarter-million. Within 18 months, we see the potential for a teraflop in a desktop for under $50,000."

It's ClearSpeed's second chip design. The first had 64 "processor elements" incorporated into the chip. The new chip has 96 processor elements, and so can bring more power to bear on intensive computing problems since one chip can act like 96.

ClearSpeed designs the chips, which IBM then manufactures. ClearSpeed's plan is to sell add-in cards that can be installed into standard PCs with open PCI-X slots, including those with microprocessors from Intel or Advanced Micro Devices, and even high-end Apple Computer PowerMac G5 computers.

That will create demand for more substantial systems using large clusters of ClearSpeed's chips. Calise says target customers are big computer manufacturers, which might sell standard computers with ClearSpeed add-in cards as an option.

Target end users would be in the financial sector, where credit analysis is starting to demand a good deal of computing power, life sciences, and oil and gas exploration. ClearSpeed has already started shipping evaluation boards using its older chips.

"We can help build a petaflop machine for some of these extremely difficult computational tasks, but also build a desktop underneath some researcher's desk that can do 250 gigaflops at a relatively low cost," Calise says.