Picking out a supercomputer used to be easy. You'd find $15 million or so and call Cray Research Inc. Lately, though, Cray's near-monopoly has given way to a profusion of alternatives, each with a unique design and each claiming to excel in speed, cost, and ease of use--or all three.
Even for the savvy scientists and engineers who pay up to $30 million for such machines, the market is baffling. They could stick with traditional "vector" machines, such as those made by Cray and its Japanese rivals, Fujitsu and NEC. But performance improvements in such hardware are coming slowly as designers try to wring more speed from individual processors. Massively parallel processors (MPPs), on the other hand, hold out the promise of far greater speed at lower prices by combining dozens or even hundreds of ordinary microprocessors. Trouble is, there are about a dozen companies building such machines, and it's unclear which design will dominate--or survive. Finally, there are clusters of powerful workstations that can be programmed to work together on some supercomputing jobs, but not all.
The result? Customers sat on their hands, and 1992 turned out to be the worst year on record for supercomputer makers. Cray Research lost money in 1992. Through the first 11 months of its fiscal 1992, which ends this month, NEC Corp. sold just seven supers, half of 1991's total. Even the new MPP market was disappointing. At $260 million, sales of MPP gear were up 18%, according to researcher Gary Smaby. But that's about a third the growth he had forecast. In all, supercomputer sales last year declined 11%, to $2.1 billion, Smaby says--a sudden and painful drop from the steady 28%-plus annual growth rate for the past decade.
Against this backdrop, Supercomputer Systems Inc., a five-year-old startup founded by former Cray designer Steve S. Chen, folded last month after running out of money to build a superfast vector machine. Chen won't be the last casualty. Seymour Cray's Cray Computer Corp., now five years behind delivery schedule, is looking to Wall Street and Washington for emergency funding. And a shakeup in the crowded MPP field is inevitable.
An improving economy will help sales, but Smaby predicts only 7.6% growth overall--not enough to sustain so many players. "Everybody is eating everybody else's lunch," says Peter Gregory, senior vice-president of marketing at the Minnesota Supercomputer Center, which sells computing time on supers. "Nobody's getting enough to support the R&D that needs to be done."
This is especially true in the nascent MPP market. Two companies building MPP machines, Wavetracer and Alliant Computer Systems, went belly-up last year, and Thinking Machines just laid off 3% of its work force.
IN LIMBO. Until a shakeout makes clear which designs will dominate, customers seem determined to hold on to their wallets. William T. "Tilt" Thompkins, head of United Technologies Corp.'s corporate research center has no plans to replace his five-year-old super from Thinking Machines Corp., which is used to design helicopter rotors and environmentally safe coolants. Instead, he's writing generalized software that can run on almost any super that hits the market. That may mean sacrificing 20% in execution speed. But, he says: "It allows you take a wait-and-see attitude for the eventual winners."
For Jim Brown, waiting for the MPP market to take off is like waiting for Godot. "I keep thinking every year it's going to come," he says. As chief executive at Scientific & Engineering Software in Austin, Tex., he has kept a tiny MPP software project running, but "until there are lots of people using [MPPs], you can't make any money with it," he says. Indeed, beyond national laboratories and universities, MPP machines so far have only sold well to oil companies, for use in seismic modeling.
SOFTWARE SHORTAGE. For now, many customers who are buying are sticking with vector machines from Cray Research, NEC, and Fujitsu. They're simpler to program than MPPs, and there's proven software for them. But the cost of boosting the speed of vector designs is going through the roof as engineers pioneer esoteric microchip technologies.
Eventually, slowing gains in vector performance should tip the scales toward MPP. But now there's still trouble turning MPP theory into practice, as Los Alamos National Laboratory discovered last year. It benchmarked a Thinking Machines CM-2, a $10 million MPP rated at a peak of 28.6 billion calculations per second (28.6 gigaflops), against a similarly priced, 2.6-gigaflops Cray Research Y-MP/8 vector processor. The Cray won every test. Why? Software to unleash MPP power is still lagging. The trick remains breaking down programs to work across many processors.
While customers wait for better MPP software, there is yet another option: clusters of desktop workstations that collaborate on large problems. They're not as fast as the fastest supers, but for many calculations the cost can't be beat. Michael J. Ross, chief executive at Arris Pharmaceuticals in South San Francisco, uses such a workstation cluster to simulate molecular interactions in new drugs. "I can't see, without a major software breakthrough, that we'd buy a massively parallel computer," he says.
What will spur supercomputer sales? It would help if strong market leaders would emerge to establish standards in the nascent MPP field. Two likely candidates are Cray Research and IBM. Both plan to launch MPP machines this year. That would be a mixed blessing for many MPP pioneers. Their market may finally take off, but only because of the presence of two potent rivals.