Cool Designs: Low-Frequency Chips And Other TricksCool Designs: Low-Frequency Chips And Other Tricks

IBM's Blue Gene project may be a grand experiment in supercomputing, but results of the research could show up in more commonplace products down the road.

In the future, smaller versions of Blue Gene (the one it's delivering to Lawrence Livermore National Laboratory will weigh in at 131,092 processors) and derivatives of its technology in small-but-powerful "blade" servers could appeal to companies maxed out on building big systems from off-the-shelf chips. Blue Gene's design features thousands of low-powered processors each running at a relatively slow 700 MHz; integrated processing and memory for fast, on-chip data movement; and embedded microcontrollers that show up more often in MP3 players and dishwashers than high-end computers. The goal is to keep price and power down and performance up.

Embedded microcontrollers such as the ones that serve as the processor "cores" of Blue Gene/L chips can cost less than $200 each, compared with more than $1,000 for IBM's standard Power 4 chip, says Michel McCoy, deputy associate director for high-performance computing at Livermore Lab. "IBM is trying to figure out whether they can harness this technology in high-performance computing and data management," he says. "The advantage to this architecture in business is a lot bigger machine for the same amount of dollars."

Then there's the issue of keeping computers cool. The engine that's driven computer-industry growth for the last 30 years--the doubling of processing power every 18 months or so achieved by shrinking the size of on-chip components--has just about given way because designs are running too hot, hurting performance. That requires new approaches. "Power gets out of control when you get to these dimensions," says Bernard Meyerson, chief technologist in IBM's systems and technology group. "Devices run hot, and you get bizarre technical results. Any computing technology that attempts to drive beyond today's level of performance runs into a power cliff."

No. 1 chipmaker Intel is dealing with similar problems. Last month, it canceled a major Pentium product line because of a hot-running design and said it would concentrate on developing special "dual-core" chips instead of pushing the performance of its conventional ones.

But for a technology to break though, customers need to buy in. IBM competitor Cray Inc. learned its lesson about building an enthused customer base in the '90s when it altered one of its supercomputer lines, and most customers didn't follow with special coding changes. The company had to make an expensive reversal--arguably at the price of its independence. The once-dominant company was sold twice between 1996 and 2000.

This time around, Cray has a road map for combining the vector-processing capabilities that make its X1 system attractive to users who need to crunch large groups of numbers at one time, with the industry-standard Advanced Micro Devices Inc. chip technology inside its Red Storm supercomputer. That machine comes online at Sandia National Laboratory in New Mexico this summer. Cray's getting marketing savvy, too: Last year, it poached Peter Ungaro, who spent 13 years at IBM, to become its VP of sales and marketing. In February, it hired Ulla Thiel, IBM's technical computing sales director in Europe, to do the same job at Cray.

Return to main story, Petaflop Imperative