DARPA awards pacts to juice computing

The Defense Advanced Research Projects Agency last week gave a green light — and a total of more than $146 million — to three companies for the second phase of its High Productivity Computing Systems (HPCS) program. The program will build supercomputers that far outpace those in use today.

DARPA selected Cray Inc., IBM Corp. and Sun Microsystems Inc. as the three primary contractors.

Cray Inc., which is partnering with New Technology Endeavors Inc., calls its effort Cascade. DARPA awarded the team $43.1 million. IBM Corp. received $53.3 million for an approach it calls Productive, Easy-to-use, Reliable Computing Systems, or PERCS. Sun Microsystems Inc. got $49.7 million to continue work on its integrated systems approach, called Hero.

The three companies advanced from a field of five and will spend the next three years figuring out how to solve the myriad problems standing in the way of progression. DARPA officials want a working system built by 2009 or 2010.

"They're going to be fleshing out concepts, doing risk-reduction efforts," said DARPA spokeswoman Jan Walker. "The ultimate end of this phase will be a preliminary design."

The agency wants the system to bridge the gap between current technology, some of which is built on foundations that are decades old, and quantum computing, a technology still in its embryonic stages in laboratory experiments.

"Current systems are still too hard to program, they're too unreliable," said Keith Shields, Cascade project manager at Cray. "And they're getting more out of balance. Processors are getting more powerful while things such as bandwidth aren't growing as much."

With DARPA, "we were encouraged to start with an extremely fresh approach," said John Gustafson, senior scientist with Sun's high-performance computing division. "There's a lot in computing that goes back to the Eisenhower administration."

"We're focused on improving the whole process," Walker said. The agency wants to see a system that is from 10 to 40 times faster than current technologies, but she emphasized that the measure will be consistent performance, not theoretical peak output.

"I don't think people are interested in peak performance anymore. They are interested in solving problems," said Mootaz Elnozahy, program manager of IBM's PERCS. "We are going to see an evolution in the way customers are thinking."

Increasing computing speed is not as simple as stringing processors together, though. The connections between the processors slow things down. They generate heat that must be dissipated. DARPA wants to emphasize tighter interaction between processors than typical clustered configurations allow, said Gordon Haff, senior analyst and information technology adviser at Illuminata Inc., a technology analysis firm in Nashua, N.H.

"This is really the lunatic fringe, specialist edge," he said. "There are some types of problems that you really need to look at the whole problem at one time, because any change you make to one part of the problem typically affects everything else."