Can Commodity Chips Meet Ambitious Goals?

Industry aims to develop a petaflop system, wonders if off-the-shelf chips will work
If you traveled the 300 desert miles between Phoenix and Las Vegas last week--and a few hardy souls in the IT field did--you might get a picture of an industry with a mild case of schizophrenia.

At Comdex, Sun Microsystems CEO Scott McNealy finally acknowledged that Sun can't resist the forces of commoditization roiling the IT market, even at the high end, revealing a deal with Advanced Micro Devices Inc. that will put Sun's Solaris operating system on low-priced, 64-bit AMD chips. "Here's a processor architecture that takes advantage of what Solaris has to offer," McNealy said of AMD's Opteron--a far cry from Sun's years-long assertion that only its Sparc chips could deliver the full wallop of Solaris. Sun sells a version of Solaris for 32-bit Intel chips but plans to deliver larger systems based on Opteron.

"Here's a processor architecture that takes advantage of what Solaris has to offer," says Sun CEO McNealy, with AMD CEO Hector Ruiz.

Meanwhile, at an annual supercomputing conference in Phoenix, attendees questioned if commodity processors could yield systems fast enough to meet an industry goal of producing a petaflop computer--capable of one-quadrillion math operations per second--by the end of the decade. True, the number of Intel-based machines on the list of the world's 500 fastest supercomputers has more than tripled this year to 189. But two of the most-talked-about machines, Cray Inc.'s X1, and IBM's experimental Blue Gene/L, forgo off-the-shelf chips in favor of architectures that emphasize moving data around quickly.

At stake: Whether delivering the best performance for the buck lies in building inexpensive systems with fast CPU clock speeds or in developing more-specialized systems that perform well on other attributes important to emerging computing jobs in engineering, weather prediction, and the life sciences. "We've been using Web-server and video-game processors to build our high-performance computers," says Jack Dongarra, a University of Tennessee computer-science professor who maintains the Top 500 supercomputer list. "They fall down on data movement and interconnect speed."

With support from the Pentagon's Defense Advanced Research Projects Agency, Dongarra last week unveiled a suite of benchmarks, called HPC Challenge, which measures memory bandwidth, latency, and processing speed. As the number of transistors that can fit on a chip increases, Dongarra says, vendors are putting more data-communications functions on chips, in addition to multiple CPUs.

IBM is trying to make hay from the trend. Blue Gene/L is the world's 73rd-fastest supercomputer, though it's constructed of IBM Power 4 processors running at just 500 MHz each. By putting two CPUs plus 4 Mbytes of memory on each of its 512 chips, the system, packaged in a small cabinet, can reach high speeds while running on relatively little power. Over the next several years, IBM will bring technology from the project to blade servers and other products. Last week, it shipped its first blade servers based on its Power 4 chip, which the company also plans to supply to Microsoft for its next-generation Xbox game console. On the high end, a Blue Gene/L with more than 65,500 Power chips could be the world's fastest computer in two years, IBM says.

According to Burton Smith, chief scientist at Cray, its X1's specialized processors and high data-transfer rate can yield computing speeds 15 to 25 times as fast as IBM's Power chips. But, he says, "it's been very difficult to get this message out because everyone's measuring processor speed."

Editor's Choice
Brian T. Horowitz, Contributing Reporter
Samuel Greengard, Contributing Reporter
Nathan Eddy, Freelance Writer
Brandon Taylor, Digital Editorial Program Manager
Jessica Davis, Senior Editor
Cynthia Harvey, Freelance Journalist, InformationWeek
Sara Peters, Editor-in-Chief, InformationWeek / Network Computing