Share this story
Close X
Switch to Desktop Site

Massively Parallel Supercomputers Show Promise

THE Wavetracer supercomputer that sits in the company's Acton, Mass., headquarters is not exactly awe-in- spiring.

It stands waist-high, looking more like the scale-model of a skyscraper than a superfast machine. Its processors are ordinary - no more powerful than IBM's original personal computer.

About these ads

Welcome to the future - at least the possible future - of supercomputing.

Wavetracer Inc. and other companies are building powerful new computers using rather mundane parts. The idea is called massively parallel processing, or MPP. Though not new, MPP has become suddenly popular in the last two years. "Everyone is building a parallel computer," says Rich Fiorentino, Wavetracer president and chief executive.

"The next wave, at least in everyone's mind, is massively parallel computers," adds Christopher Willard, senior industry analyst with Dataquest Inc.

The MPP market has been growing about 40 percent a year, albeit from a small base. The industry reached over $200 million in factory revenue last year, and Dataquest expects the market to grow just over 20 percent a year over the next five years.

"It's better than any other part of the supercomputer market that we follow," Mr. Willard says. Cray working on hybrid

Even Cray Research Inc., the famed manufacturer of traditional supercomputers, has said it is speeding up its MPP program. By 1993, it plans to deliver a supercomputer hybrid that incorporates that architecture into its Y-MP line of machines.

MPP computers are much easier to explain than to build. Instead of using a handful of very fast, specially cooled processors to power a supercomputer, MPP machines link together thousands of more ordinary processors to achieve similar results.

About these ads

"What it does is make supercomputing much more cost-effective," says Danny Hillis, cofounder of Thinking Machines Corporation. "These days you can buy a supercomputer basically for what it used to cost to install the air conditioner" into a specially cooled supercomputer room. For example, Cray Research is selling its Y-MP C90 supercomputer for $30 million. A comparable Connection Machine from Thinking Machines would cost something like $5 million to $6 million, Hillis estimates.

Wavetracer's MPP machines run much slower than Cray's but at a cost of no more than $400,000.

Such comparisons are problematic, because MPP supercomputers can only handle certain kinds of problems and require a new kind of software. Writing that software can easily take a year and boost the price of the machine by a factor of three or even 10, industry officials say.

"There have not been any breakthroughs in the process of building software for massively parallel applications - both building the software and getting it to work," says John White, president of the Association of Computing Machinery. When to use MPP machines

So, if the software already exists to handle a company's problems, it might well buy a Cray or some more traditional machine. If the company needs to solve brand new problems, it should look seriously at an MPP machine, industry analysts suggest.

Traditionally, supercomputer architects used very fast single processors to handle computer programs line by line - rather like getting the world's best speed-reader to zip through War and Peace. More than a decade ago, they recognized the importance of parallel processing - using a team of "speed-readers" to break up the program into chunks. MPP goes a step further by using massively parallel processors - 1,000 or more - each reading a separate page of the novel.

Many MPP machines, including Wavetracer's, use simple processors to handle the data, says Jim Jackson, Wavetracer's chief scientist. The drawback is that the processor falls asleep if its particular "page" of data is blank. Other MPP machines use more sophisticated processors that avoid that problem but create communications bottlenecks.

If MPP succeeds, supercomputers will move increasingly into the business world, doing what Willard of Dataquest calls "supercomputing of the mundane designing a better golf ball or figuring out how to cook meatballs in their original can.