Advertisement

Supercomputer Designers Are Shifting to Use of Chips : Technology: Other makers have jumped into the market once dominated by Cray Research. New players include Intel, Silicon Graphics and IBM.

Share
ASSOCIATED PRESS

The thing that makes a supercomputer super today is the same thing that drives a home computer. But more of it.

Supercomputers were once room-size machines and the stuff of technology legends, helping the government forecast weather or scientists analyze earthquakes, create new drugs and understand nuclear physics.

They’re still the most powerful and expensive computers being made but now they look like big refrigerators. And they’re different inside, using microprocessors, the chips that run desktop computers, by the dozen or sometimes hundreds.

Advertisement

And the market, once dominated by Cray Research, has opened up to include firms such as Intel Corp., the world’s leading chip maker, and Silicon Graphics Inc., known best for the computers used to make the film “Jurassic Park.”

Companies now use one of two approaches to building supercomputers: stringing chips together to share one central memory core, or stringing chips together and giving each its own memory.

The first approach is called symmetric multiprocessing, or SMP, and the second is known as massively parallel processing, or MPP.

The big difference is that SMP supercomputers can grow only so big. In theory, there is no limit to the size of massively parallel machines because each microprocessor carries its own memory.

“At some point, as you keep adding processors in the SMP design, efficiency declines because they’re all trying to get into the same memory,” said Wendy Vittori, operations director for the Intel scalable systems division, based near Portland.

“The main role of the supercomputer in the ‘big picture’ is primarily to do simulations of physical events, anywhere you have complex physical phenomena,” Vittori said.

Advertisement

They are the tasks that Cray Research began building supercomputers to handle when the main customers were universities, the military and the government. But things have changed since 1976, when Cray delivered its first supercomputer.

“It’s a market that’s shrinking mainly because of government contractors declining with the end of the Cold War and the ‘Evil Empire,’ ” said Steve Conway, chief spokesman for Cray, based in the Minneapolis suburb of Eagan, Minn.

Cray still dominates sales of supercomputers costing more than $5 million, with 71% of the 1994 market, said Gary Smaby, founder of the Smaby Group, a Minneapolis computer marketing research firm.

Cray also holds 35% of the $1-million to $5-million market, with Intel, IBM and the three main Japanese supercomputer makers, Fujitsu, Hitachi and NEC, all fighting for second place in the 5% to 15% range.

But in the rapidly growing market for small supercomputers in the $100,000 to $1-million range, the leader by far is Silicon Graphics, with more than a third of the market share. Cray has only a tiny slice of that pie.

Howard Richmond, analyst at the Gartner Group in Stamford, Conn., said Silicon Graphics grabbed a large share by offering relatively low prices for reliable SMP machines that run on proven software.

Advertisement

The combination not only appeals to researchers and engineering companies, but it also appeals to commercial customers, such as retail chains and banks.

“The issue is how you make money with technology,” Richmond said. “Follow the money. It’s in industry, never in universities and government.”

Willy Shih, marketing director for Silicon Graphics’ supercomputer division, says the company has exploited the supply of cheap computer chips and its own experience in computer graphics software to build SMP machines that operate more efficiently than more expensive competitors.

“Because it’s more affordable, you attract more mainstream users,” Shih said.

Sales of massively parallel machines have lagged behind symmetric multiprocessor computers partly because nobody has developed software that takes advantage of the design.

“Parallel processing is software intensive, and the key to getting it to work is the software,” said Irving Wladawsky-Berger, head of IBM Power Parallel Division in Somers, N.Y.

IBM and Intel are battling each other in the high- and mid-range of the supercomputer market, holding roughly equal market shares. But IBM is busily courting commercial customers while Intel remains focused on the scientific and university market.

Advertisement
Advertisement