Firm Builds SuperComputers for One-Fifth the Price.

THE average supercomputer starts at about $5 million and is typically found at only the wealthiest of universities and Fortune 500 corporations.

But one Westlake Village company believes that it can supply computers powerful enough to do what these supercomputers do that are both affordable and easy to maintain.

Supercomputers are generally used for projects that require complex mathematical computations, including the design of cars, aircraft and space vehicles. The recent mapping of the human genome would have been impossible without it, said Steve Conway, communications director for supercomputer maker Cray Inc.

Using existing technology, Dolphin Interconnect LLC and server-builder RackSaver Inc. of San Diego have linked clusters of computer processors together, called it "Beowulf" and claimed it could be the next step in the evolution of supercomputers.

Dolphin built the components critical in connecting all of the computer's processors and developed the software to run the unit. RackSaver built the hardware and assembled the cluster components into the Beowulf using processors already assembled by Sunnyvale-based Advanced Micro Devices Inc.

So far, the Dolphin/RackSaver partnership has sold two Beowulf clusters, both to universities. Dolphin officials believe there is even more of a market than that for their product.

Dolphin was founded in 1992 as a developer of high-speed interconnecting equipment, along with computer hardware and software for computer clusters. Last year, it became a major supplier to computer manufacturer Sun Microsystems, boosting its sales and revenue figures.

Dolphin posted $15.9 million in net income in 2000 on X in sales, compared to net loss of $91,000 on $10.3 million in sales in 1999, enabling the company to offer its first dividend ever to shareholders.

While a unit costs between $200,000 and $1 million, depending on the size of the cluster, it remains well below the millions it costs to purchase a traditional supercomputer built by International Business Machines Corp. or Seattle-based independent supercomputer maker Cray Inc.

Parts come off the shelf, as opposed to being built from scratch for traditional supercomputers, driving down the ultimate cost of the product.

But the biggest reason to get the Beowulf is the convenience of having a supercomputer in the lab, rather than sharing an institution's mainframe with other users, said Erik Asphaug, a principal investigator for the UC Santa Cruz's earth sciences, astronomy and physics Department.

Last month, the school purchased a Boewulf supercomputer to study asteroids and other bodies in deep space. Asphaugh said the computer will also help with research into the origin of the moon, the formation of impact craters and the fragmentation of asteroids due to impacts and explosions in space.

UC Santa Cruz' Beowulf, funded by the National Aeronautics and Space Administration, provides the university with a computing power nearly equal to most traditional supercomputers at a fraction of the cost, said Keith Murphy, Dolphin's vice president of sales and marketing.

As Murphy sees it, traditional supercomputers are on their way out, to be eventually replaced by cluster supercomputers like Beowulf and its next generation of similar units. He says faster processors and improved technology in so-called "off the shelf" parts will make them more convenient to use and maintain.

"People can get parts off a shelf instead of going through a costly maintenance project every few years," Murphy said.

But Cray's Steve Conway said clusters are still not the equal of traditional supercomputers.

"When you talk about a cluster, you're really talking about 500 separate computers that are linked together," Conway said.

"What you end up with is a computer where, the very fastest part is the microprocessor, but its other parts -- the interconnection -- are quite abit slower," he said of the links connecting the computers to each other.

Cray pioneered the technology in .1995 by linking several computer processors together. The theory worked, but the final product was much slower than the larger supercomputers it was already marketing, Conway said.

"We found that cluster computers are good when you have a small problem, but when you have big problem with lots of different computations, it's much, much slower than a traditional supercomputers and customers wanted speed," Conway said.

While IBM and NEC Corp. of Japan have marketed cluster computers, their focus remains on marketing the traditional supercomputer.