The "Big Mac" Supercomputer Biz
Nobody was more surprised than executives at Apple (AAPL ) on June 23, 2003. That's when scientists at Virginia Polytechnic Institute & State University called to see if they could buy the first 1,100 PowerMac G5 desktop computers. Even though the G5 had been unveiled just hours earlier, the Virginia Tech team already knew it wanted the G5s, each fitted with two of IBM's (IBM ) new PowerPC 970 microprocessors, to build a world-class supercomputer.
Philip Schiller, Apple's senior vice-president of worldwide marketing, admits he was stunned by the audacity of Virginia Tech's concept. Until then, Apple hadn't been a serious player in big-time supercomputing, even though so-called cluster supercomputers -- clumps of wired-together personal computers built with chips from Advanced Micro Devices (AMD ), Intel (INTC ), IBM, and Sun Microsystems (SUNW ) -- had been the hottest breed of "super" for more than a decade.
Up to that point, Apple's cluster entries had been mainly low-end systems, such as the 56-computer PowerMac G4 cluster assembled in 2002 by University of California at Los Angles physicists, and UCLA's earlier, smaller Macintosh G3 clusters. What impressed the UCLA researchers was how easy a bunch of Macs could be roped together. They developed software to do the setup in hours, not the days or even weeks typically needed to create a Linux cluster of PCs.
UCLA's software is a major reason why there are dozens, maybe hundreds, of smallish Apple clusters around the world. But they had been much too slow to earn a spot on the semiannual Top500 List of Supercomputing Sites.
After talking with the Virginia Tech folks, Apple execs quickly became enthusiastic about building a Top500-class machine. They arranged for Srinidhi Varadarajan, an assistant professor of computer science, and Patricia Arvin, Virginia Tech's associate vice-president for information technology, to fly to California. By week's end, Apple and Virginia Tech had decided to collaborate.
Just five months later, when the Top500 List was published in November, 2003, Virginia Tech's system ranked as the world's third-fastest super, notching a peak speed of 17.6 teraflops. That's geek-speak for 17.6 trillion (tera) floating-point operations per second (flops). Flops are the calculations at which supercomputers excel.
The Virgina cluster -- called System X in honor of its being the first academic super to surpass 10 teraflops -- isn't ranked on the latest Top500 list because it has already being upgraded. The PowerMacs have been replaced with Apple's newer Xserve G5 servers, and Varadarajan and his crew at Virginia Tech's Terascale Computing Facility are still putting the revamped cluster through its initial shakedown cruise.
The tight-knit supercomputing community was surprised by what Virginia Tech and Apple pulled out of their hat -- and impressed. It didn't take long for others to jump on the Big Mac wagon. The newest system aims to one-up Virginia Tech by linking 1,566 G5 Xserve units. Dubbed Mach 5 and being built by Colsa Corp. in Huntsville, Ala., this cluster will have a theoretical peak speed of more than 25 teraflops.
That would earn the No. 2 spot on the Top500 list that was issued in June. Only Japan's Earth Simulator, built by NEC (NIPNY ) and rated at 41 peak teraflops, is faster -- although it and Mach 5 may soon be overtaken by clusters now a-building by longtime super suppliers IBM and Cray (CRAY ).
Mach 5 is slated to go into service this fall at the Army Research & Development Command's Aviation & Missile Research, Development & Engineering Center in Huntsville. It will do nothing but run one job: simulating hypersonic flight. Currently, the center uses an aging IBM supercomputer with 284 processors, and each simulation chews up a solid month of computing time. With Mach 5, says Anthony C. DiRienzo, executive vice-president at Colsa, "they'll be able to do a new run every day, overnight."
DiRienzo figures Apple can count on lots more supercomputer business. None of the other vendors that bid on the Army job, he explains, came close to matching Apple's price/performance ratio. The hardware was only $5.8 million -- 60% of the speed of Earth Simulator for 1.5% of its $400 million cost. "But what's really nice about the G5 pizza-box-size servers," he says, "is how well they dissipate heat." Apple's engineering stood head and shoulders above the competition on this score.
In fact, says DiRienzo, heat problems in the other systems prevented them from running their chips at maximum speed. "They had to slow their chips down. That meant they had to [submit bids for] installations with more processors. That drove costs up. So Apple was definitely the best value proposition," he notes.
DIGITAL JOHNNY APPLESEEDS.
A few other customers are picking up on that message. The University of Maine plans a cluster of 256 G5 servers. UCLA's Plasma Physics Group is pumping up its Apple resources again, this time with 256 G5s. And Princeton University's Center for the Study of Brain, Mind & Behavior has installed a 64-node Xserve G5 cluster.
Meanwhile, Virginia Tech is dreaming of still bigger Roman-numeral systems. First will be System L, offering 50 teraflops or more. Then comes System C, with 100 teraflops or 100 trillion calculations every second. Varadarajan and his collaborators might just become modern-day digital Johnny Appleseeds.
By Otis Port, a senior writer for BusinessWeek in New York
Edited by Beth Belton