Hypercomputers Are More Than Hype - Star Bridge Systems' HAL300GrW and IBM's Pacific Blue - Company Business and Marketing

Computer Technology Review,  July, 1999  by Mark Brownstein

• E-mail
• Print
• Link

June 6, 1944, D-Day, was undoubtedly an important date in history. History may record June 8, 1999 as potentially even more earthshaking.

On June 8, 1999 a small, Salt Lake City, Utah, company named Star Bridge Systems announced that it has developed a technology that may "revolutionize the world of computers and electronics." If it works, their prediction may be even more far reaching than their somewhat immodest-sounding announcement suggests.

Star Bridge Systems (SBS) announced that it had developed what it called Hypercomputers--reconfigurable computers that are orders of magnitude faster than any previous computers. Within 18 months, for example, the company may be offering a $1000 desktop computer with performance that matches many of today's supercomputers. This low-priced computer will deliver 1000bips performance. The computer will deliver performance that is three orders of magnitude higher than today's PC, according to Kent Gilson, the company's chief technology officer and the brain behind the hypercomputer.

The architecture behind the reconfigurable computing technology uses one or more circuit boards with FPGAs, Field Programmable Gate Arrays, that can be reprogrammed at thousands of times a second. In effect, an FPGA can become an optimally tuned computer that is ideally programmed for each required computing task. Materials provided by the company compare reconfigurable computing to the process that the human brain performs, programming itself to solve new problems and changing the programming as often as necessary.

The massively parallel architecture of the hypercomputer can create new links between FPGAs, just as cells in the brain can create or change links, in order to optimize performance. With the addition of each FPGA, significant performance boosts are accomplished.

The consumer PC that is being developed is expected to be easily upgradable, with upgrades possibly being as easy as downloading improved software. Further, the consumer model will run the most popular software with operating system emulators, functioning in a manner similar to the way Alpha computers currently run Windows NT.

SBS is initially offering hypercomputing systems to large customers who are currently using supercomputers. Shell Oil, Raytheon, Dow Chemical, and others were among those companies interested in the HAL computers, Gilson said. Further, NASA and the National Weather Service also expressed interest in the computer.

The HAL-300GrW1, a $26 million hypercomputer, delivers 12.8TeraOps, as compared to the IBM Pacific Blue, which cost $94 million, and delivers 1.2TeraOps sustained and 3.9TeraOps peak. In fairness, the tests used to determine the actual number of operations per second (Ops) were different for the two computers, but the magnitude of the differences (price and performance) make it clear that the HAL is an interesting little machine. Other comparisons further spotlight the differences between the IBM's supercomputer, which consists of 5,856 PowerPC 604 processors and requires 8,000 square feet of floor space, and the HAL, which uses 280 FPGAs, takes up 3.78 cubic feet, and can fit on a (large) desktop or in a standard 19 inch rack. Power consumption is another important factor--the HAL burns about 1600 watts of 110 volt power and can be plugged into a standard wall jack, while the Pacific Blue requires 3.9 megawatts. The HAL uses one standard extension cord and 12 feet of internal cabling, while the Pacific Blue uses 5 miles of number 4 power cable, and 50 miles of cable internally. The HAL is cooled with 10 internal mini-fans, while 280 tons of air conditioning are required to keep Pacific Blue in the pink.

When I interviewed Gilson, he was calling from a Kinko's the night before a meeting with a major research facility. Because he was traveling, he took along what he called "Hal Jr.," a demonstration unit that fits into a suitcase, and delivered a paltry 640 billion instructions per second.

The company's initial strategy is to sell the hypercomputers--high ticket items that deliver world class performance. Although sales volumes will be low (Gilson expects the company to sell about 100 per year), the numbers are large enough to build a company with enough strength to venture into such high volume, lower cost areas as supercomputer PCs.

Having arrays of FPGAs running, potentially reconfiguring at as much as a thousand times a second, sounds like a recipe for chaos. Bringing the FPGAs under control, and designing an operating system and programming language that can tame the beast is a task that Gilson has been working towards for more than a decade. In the past two years, Gilson worked on a programming language that he calls VIVA. An Active-X front end on a HAL that will be hooked to the Internet will let programmers design and run test applications using HAL.

"With VIVA, the processor is one of the elements in the synthesizer flow that takes an algorithm and builds a computer out of it," Gilson said. "One of our primary breakthroughs was coming out with compiler tools that could take a high level algorithmic description (of a computer) and build a computer around (the algorithm)," Gilson said.