How The Numbers Crunched Cray ComputerJohn W. Verity
It was difficult, even for his rivals, not to wish Seymour Cray the best with his latest supercomputer. In a streak unmatched in computing, the brilliant computer architect has been stretching the limits of performance since the days when computers ran on red-hot vacuum tubes. One of his early machines, the CDC 6600, gave IBM fits by executing 1 million calculations a second. In 1976, his odd-looking Cray-1 kicked off what has since mushroomed into today's $2.5 billion supercomputer business.
Lately, in a computer industry known for its self-promoting millionaires, the 68-year-old recluse has become all the more of a hero to fellow engineers. "Seymour Cray is in a class of his own," says Irving Wladawsky-Berger, general manager of IBM's Power Computing Div., which makes and sells supercomputers.
NEW ERA. Ultimately, though, no amount of admiration could make up for product delays and repeated cash shortages. On Mar. 24, having spent an estimated $200 million without selling a single machine since its founding in 1989, Cray Computer Corp. filed for bankruptcy protection.
Although it was just one of a string of U.S. supercomputing makers to bite the dust in recent years, the event marks the end of an era. Gone are the lush cold-war budgets of the government's scientific and code-breaking agencies, and with them the expensive luxury of building computers from the ground up, designing everything from scratch.
Indeed, the high-performance computing market has shifted dramatically. Sales of high-end supercomputers, which cost as much as $30 million, have declined from about $1.5 billion in 1990 to $1.2 billion this year, according to Smaby Group, a market researcher. Cray Research Inc., Seymour Cray's old company, has steadily improved its share there from 60% in 1990 to 75% today as U.S. and Japanese competitors have fallen by the wayside. The real excitement, though, is in machines that sell for $100,000 to $2 million and often harness hundreds of microprocessor chips to attack big problems in parallel. Sales of such computers may hit $1.35 billion this year, up from $1 billion last year.
It's precisely this sense of historical passing that gives Seymour Cray's troubles a certain poignancy. "People around here feel badly for Seymour," says Robert H. Ewald, president and COO of Cray Research. "You hate to see something turn out like this for him." Especially, it seems, because his new machine was his most ambitious venture. So ambitious, in fact, that Cray Research decided it couldn't afford to fund it, forcing Cray to spin off his own company and find other backers.
GaAs GUZZLER. As in previous designs, his main quest has been to cram more tiny electronic circuits into a tighter space. This yields higher computing speeds, but it also makes it more difficult to manufacture circuits and keep them cool so they won't burn up. It didn't help that the inventor opted for a difficult-to-use semiconductor material called gallium arsenide (GaAs).
Seymour Cray wouldn't comment. But it's clear that by pushing on so many different fronts at once, he ran into a thicket of design and manufacturing problems. His company had to invent its own miniature robots to help with manufacturing and build new equipment for testing its high-speed circuits. And it ended up having to make its own GaAs chips instead of buying them outside.
All this added up to delay after delay and a string of cash crises. The original Cray-3 computer slipped past several delivery dates, until the decision was made to scrap it in favor of the still more advanced Cray-4, which would be twice as fast. Yet all the while, Cray's intended market was in the process of drying up. Government budgets were thinning, and the one order Cray did win--from the Lawrence Livermore National Laboratory--was canceled when the company failed to deliver on time.
To the end, Cray's only substantial revenue came from a development project undertaken for the National Security Agency. But even if the Cray-4 shipped today, "the technology is late to market," says Debra Goldfarb of International Data Corp., a market research firm. And a late supercomputer, in the end, is still late.
THE SUPER LINEUP
CRAY RESEARCH The leader, with a 75% share of high-end market
IBM Has shipped more than 370 parallel systems
INTEL Unprofitable but committed to parallel processing
CONVEX COMPUTER Rushing to tap commercial markets
FUJITSU, NEC Still not making money in supercomputers
NICHE PLAYERS Meiko, nCube, and Silicon Graphics