05 Oct Cray Rides Again
With X1, supercomputing’s pioneer company looks to reclaim lost ground
CHIPPEWA FALLS, Wis. — How ironic that Cray Inc. should be operating out of offices on Lowater Road here in its hometown, just as the tide is lifting its ship to new and wonderful heights.
With its next-generation X1, on the market since December 2002, the company with the name in supercomputing is looking toward the future with a twinkle in its corporate eye that, just a few years ago, had nearly been snuffed out.
After prospering throughout the 1980s, the market for supercomputers dwindled and the company was purchased by Silicon Graphics Corp. (SGI) in 1996. A variety of factors led SGI to sell Cray in 2000, about three years after engineers began developing what was to become the X1. (SGI maintains its own huge presence in Chippewa Falls, largely because of Cray’s legacy and the talent the company has lured to northwest Wisconsin for 30 years.)
X1 is a fusion of the vector-based computing concept hatched by the company’s late genius founder, Seymour Cray, and Cray’s T3E massively parallel processing (MPP) system. Skipping the mind-bending technical details, in melding the two concepts, Cray has stuffed the computing power contained in a machine that was sized just north of your grandma’s chest freezer just one model generation ago into a black box (called a node module) about the size of three hardback novels laid end to end.
“Cray research was founded in 1972, and this is really the very first time with a proprietary processor that we deviated from that original architecture,” said 25-year Cray veteran David Kiefer, vice president of product development and manufacturing. “Our plan now is to take this architecture and to develop a number of machines based on this architecture for the next 10-plus years. It’s a whole new line of computer systems.”
Kiefer has seen more supercomputing projects than Michael Jackson has cosmetic surgeons—and with much more plausible results. He worked on the back end of the Cray-1 project, doing design verification during production. Now he’s overseeing the marriage of the old and the new, a marriage that has spawned a machine that can perform at 800 gigaflops (800 billion calculations per second) with 16 node modules in one chassis. Four X1s working in concert during WTN’s recent tour of the manufacturing facility were perfectly scaled, cranking at 3.2 teraflops (3.2 trillion calculations per second).
The X1 took five years to develop. But such was the government’s faith in Cray’s corporate abilities that it awarded the company a contract for the groundbreaking machine just one year into the project—strictly on the Cray name alone. Government agencies, particularly the departments of Defense and Energy, were the first buyers; now, the company is gearing up to reclaim customers in the commercial sector that Cray lost in years past.
“I think beyond [DOD and DOE], we’ll see a significant market, probably in the weather area—that’s a pretty significant market for us—as well as potentially aircraft and automotive,” Kiefer said. “We haven’t penetrated that market at this time, but these are traditional Cray markets we see as potentials for using the system.”
Industry analyst Alan Robinson said it may be difficult for Cray to break out of the government market for the X1—a 16-node, water-cooled chassis will put you back a cool $8 million—but Cray does have a couple things going for it: For one, Moore’s Law should drive that price tag lower and widen the X1’s market. Second, the X1 in the past 12 months or so has built a solid track record with climate modeling, particularly at charter customer Oak Ridge National Laboratories in Tennessee, where it has outperformed equivalent rival IBM Power4 processors, Robinson said.
“As government applications prove to be running a lot more quickly on the X1, and as the X1 now opens up new problems to be tractable and soluble, I think that may turn the heads of the corporate market a lot more,” Robinson said.
Kiefer noted that the company is in development with the next two generations to succeed the X1. Last month Cray announced a funding agreement with the federal government in which Cray and Uncle Sam will ante up $17.5 million for the development of one of those successors, code-named Black Widow, which Robinson said would be introduced in 2006.
Meantime, the X1 ought to continue to bring good things to the company that invented supercomputing itself.
“Vector computing can take advantage of doing many operations with the execution of one instruction,” Kiefer said. “You need to be able to find that parallelism in your code to take advantage of that. There are a lot of codes that have been written or optimized for exploiting that parallelism, and those are codes that we will continue to do very well on.”
Lincoln Brunner is a Stevens Point, Wisconsin-based freelance writer and a regular contributor to the Wisconsin Technology Network. He can be reached at firstname.lastname@example.org.