The Hole Of A New MachineRussell Mitchell
Cloak the creation of an exotic new product in mystery, and excitement and anticipation will build for its eventual debut. But stretch out the mystery too long, and anticipation can sour into skepticism--or worse, apathy. Just ask Steve S. Chen.
In 1987, the celebrated Chinese-born supercomputer designer left Cray Research Inc. to pursue his dream of building a new generation of ultrahigh-performance computers. At Cray, Chen had helped create the world's fastest computers. But the market leader could not afford to fund his most ambitious plans. IBM, seeing an opportunity to improve its standing in scientific computing, stepped forward to help out. Five years and an estimated $100 million-plus in IBM funding later, though, and Chen's Supercomputer Systems Inc. (SSI) has yet to show a working prototype. And now, a newer class of supercomputer threatens to make Chen's machine obsolete before it even gets out the door.
How so? Chen has been working on a so-called vector architecture, a computer design nearly 20 years old that most experts think is approaching fundamental speed limits. Vector computers such as Cray Research's top-of-the-line C90 rely on a relatively few highly specialized processors to perform calculations at lightning speed. Their exotic chips are costly to build and run so hot that they need elaborate liquid cooling schemes.
At this point, wringing more speed from vector technology costs tens of millions of dollars in research and development. Supercomputer legend Seymour S. Cray is trying by using a chip material called gallium arsenide. He is several years behind schedule, though, and only recently got a prototype to run at full speed for more than a few hours at a time.
`ZERO SUM.' The alternative, which is drawing increasing support among key academic, military, and industrial customers, is a scheme called massively parallel processing (MPP). It spreads big computing problems over hundreds or even thousands of standard, low-cost, PC-class microprocessors. Such machines already top vectors in theoretical peak performance, and new programming techniques should help them eclipse vector machines for most "real-world" applications as well. So the big growth will be in MPP (chart), and even vector champ Cray Research plans to start shipping MPP machines next year.
Even if he gets to market in 1993, Chen will find the going tough. Cray Research continues to dominate. Fully 58% of the world's vector supercomputers are Crays. And determined Japanese makers NEC, Hitachi, and Fujitsu are coming on strong in vector machines. "It's a mature market, a zero-sum game," says Gary Smaby of The Smaby Group, a market researcher.
Still, Chen displayed no sense of urgency when addressing a supercomputer conference in Minneapolis on Nov. 19. Instead of describing his creation, Chen gave a general talk on supercomputing. Asked by his audience for technical details, he said, smiling: "I don't think I'm supposed to do that." O.K., but when might he announce a working prototype? "My answer is very simple: soon." How soon? "Very soon." Said Paul Messina, director of the Concurrent Supercomputing Consortium at California Institute of Technology: "I was expecting a bland speech, but maybe not this bland."
PEAK PIQUE. Chen defends his continuing reticence, saying: "You promise too much, you get killed." But he also may be trying to lower expectations. For example, he now criticizes competitors for overemphasizing theoretical performance levels. And he let slip that his machine now bears little resemblance to what he had proposed building at Cray--a system embodying many technical breakthroughs, including circuits interconnected by light beams.
Will IBM pull the plug on Chen? Not likely. The two companies have set up a small "marketing agency" at SSI--a handful of people led by an IBM executive. "We do count on SSI," says Irving Wladawsky-Berger, IBM's assistant general manager for supercomputing. But IBM has other irons in the fire: At the Minneapolis convention it showed a prototype of a forthcoming MPP machine. A return on its huge SSI investment would be nice, but IBM's foray into supercomputing now seems to depend only a little on mystery man Steve Chen.