US Supercomputing Jumps Ahead With Two New Parallel Processors
Intel, Thinking Machines shatter previous calculation speed records
THE United States is jumping out to an early lead in an emerging field of supercomputing that will be critical to producing some of the fastest machines on earth in the mid-1990s and beyond. While the Japanese can never be counted out of any high-speed computing race and the Europeans are also working on the technology, called massively parallel processing, US companies are now emerging with second-generation systems and other impressive work is going on in laboratories.
The technology is seen as central to the quest of creating a computer capable of doing at least 1 trillion calculations per second - a target the Bush administration has set as a goal by the mid-1990s as part of its high performance computing program.
At this speed, called teraflop, researchers believe they can begin to crack some of the big questions of science, such as weather prediction and simulation of the human nervous system, as well as mimic the effect of nuclear explosions and design supersonic aircraft.
Breaking the barrier is considered important to future national defense and industrial competitiveness.
"When you get to that point, it is the beginning of a whole new world," says Stephen Squires, head of high performance computing for the Defense Advanced Research Projects Agency.
One of the latest evolutions of the technology expected to lead to teraflop was unveiled last week at the Jet Propulsion Laboratory here, where Intel Corporation demonstrated its Touchstone Delta supercomputer.
The machine set a new performance record for supercomputing - 8.6 billion calculations per second. But the record won't last long. Thinking Machines of Cambridge, Mass., announced this week it would reclaim its title for the world's fastest computer by coming out with a machine capable of 9.03 billion calculation a second. It had previously set the record in March with a one capable of 5.2 billion calculations a second.
Such milestones need to be taken somewhat with a grain of silicon: These speed records are for a particular set of applications. It is difficult to compare results achieved by the new parallel computers with the more all-purpose traditional supercomputers.
Even so, the speed records underscore how quickly massively parallel processing is evolving and how well positioned the US is in this strategically important frontier of supercomputing.
"It proves massively parallel processing is viable," says Jack Dongarra, a supercomputer expert at the University of Tennessee.
Unlike conventional supercomputers, which work problems in sequence, those using parallel processing achieve their speed by yoking hundreds of microprocessor chips to tackle different parts of a problem at once.
Sales of massively parallel supercomputers, mainly to research labs, remain small: less than $200 million last year - far less than the $1.1 billion worldwide market for conventional supercomputers.
Yet the field is growing. Gary Smaby, a supercomputer analyst in Minneapolis, Minn., projects it will grow 40 percent per year between now and 1994. Perhaps more important is what the architecture means for the quest to produce the top-end computers of tomorrow.
Although there is no defined limit as to how far scientists want to push machines, the closest thing there is to a Holy Grail is the teraflop. Several companies, including Intel, believe they can reach teraflop speeds by the mid-1990s. The Pentagon has put more than $32 million into reaching in the past couple of years.
"It is a technological imperative to go parallel to achieve those those kinds of speeds," says DARPA's Squires.
Besides Thinking Machines and Intel, a Santa Clara-based company whose main business is producing computer chips, others include: Bolt, Beranek & Newman of Cambridge, Mass.; Ncubeei of Beaverton, Ore.; and Tera Computer of Seattle, Wash.
IBM has a research project under way, and Cray Research Inc. of Minneapolis, the dominant conventional supercomputer maker in the world, is moving into the area.
A couple of European and Japanese companies are also working on the technology.
Despite the US's leadership, analysts caution against hoisting the stars and stripes. They note how quickly the Japanese have risen in conventional supercomputing.
"The issue isn't who gets to teraflop first," says Jeffrey Canin, an analyst at Montgomery Securities in San Francisco. "It's who applies it first."
Dr. Dongarra says: "We shouldn't be complacent because Japan doesn't have a commercial product yet. Everybody is working on this."
The Touchstone Delta is a prototype that Intel expects to use in producing a future class of faster commercial machines which could come out within a year. It will be run by the Concurrent Supercomputing Consortium - 13 research institutes, government labs, and agencies.
Scientists here expect the Intel machine to advance knowledge in dozens of fields. They will use it for such things as monitoring global climate change, probing chemical reactions, and mapping the surface of planets.
"It lets us do research at the very frontiers of our fields," says Thomas Prince, a Caltech physicist.