Supercharging the Supercomputer Race
DECEMBER 26, 2002
NEWS ANALYSIS:TECHNOLOGY
Supercharging the Supercomputer Race
Japan has the speed and power lead right now, but IBM promises to put the U.S. back out front -- by far -- in 2005
In the last few years, the U.S. supercomputer industry hasn't had much to cheer about. Government funding had slowed
considerably, crimping a major catalyst and incubator for the innovative technologies behind powerful supercomputers. At the same time, the performance of desktop and server computers has improved dramatically, enabling those machines to increasingly take over many tasks once reserved for only the fastest machines.
Then came the big blow: Earlier in 2002, the Japanese took the mantle of leadership when NEC Corp. (NIPNY ) built a supercomputer clocked at a head-snapping speed of 36.5 teraflops, or trillions of floating-point operations a second, a measurement of how fast a computer does mathematical calculations. The advance sent shock waves throughout the U.S. supercomputer industry and put the Japanese on top with a machine four times the speed of the swiftest American creation.
Soon, the U.S. can start crowing again. The Energy Dept. in November announced a $290 million deal with IBM to build two supercomputers with a combined top speed of up to 460 teraflops. According to IBM, that's mightier than the combined power of all the machines on the latest ranking of the world's top 500 supercomputers released by the University of Tennessee and the University of Mannheim, Germany.
GOING MAINSTREAM. The new IBM machines will be installed at Lawrence Livermore National Laboratory in Livermore, Calif., and be used for simulating nuclear weapons tests, material properties, and high-explosives behavior. "They have very innovative technology," says Mark Seager, assistant director for Advanced Technologies for Lawrence Livermore.
What's so remarkable about the IBM supercomputers isn't what they'll do. It's the way they'll do it. The lightning-fast machines will harness tens of thousands of powerful microprocessors -- something akin to jamming all the PCs in a small city into one box, wiring them together, and programming them to cooperate on a single problem.
The economics of high-performance computing is undergoing radical change. Traditionally, supercomputers have been expensive, highly customized designs purchased by a select group of customers. Now, the industry is being overhauled by mainstream technologies such as Intel chips, standard high-speed connections, and inexpensive storage networks that have become fast enough to handle many tasks. "A lot of what has been used to build commercial computers is being leveraged by supercomputer makers," says Brad Day, an analyst with researcher Giga Information Group.
"MORE ACCESSIBLE." That's bringing supercomputing to the masses. Driven by an abundance of free open-source software and advances in programming that make supercomputers easier to use -- and consequently less expensive to maintain -- these muscle machines are finding greater use in Corporate America. "The technology is getting much more accessible," says Debra Goldfarb, an analyst with market researcher IDC Corp.
Increasingly, more scientific and engineering problems are being simulated on computers rather than in laboratories. In the past, supercomputers have been used for speeding research into drug development, airplane design, weather forecasting, automotive crash-test analysis and weapons research -- the mainstay applications of the supercomputing business.
That's changing. Now, Hollywood is using supercomputers for animation, financial services companies are calculating complex products such as derivatives, and even cell-phone makers are using powerful computers to design smaller, more powerful handsets. Today's generation of high-performance computers, says Nicholas Donofrio, IBM's senior vice-president for technology and manufacturing, "are becoming so capable that they're providing people with the opportunity to solve problems they could only dream about just a few years ago."
GOLDEN ERA. Consider carmaker Chrysler (DCX ). Besides the traditional crash tests, it's using powerful machines to do acoustical modeling to analyze vibrations and noise levels. "We're using supercomputers to make cars safer and quieter," says John Picklo, senior manager of high-performance computing at Chrysler. "We're simulating the heck out of noise levels and aerodynamics." Chrysler is using supercomputers built with Intel chips and running on the Linux operating system that are 20% faster and 40% less expensive than Chrysler's previous Unix supercomputers, says Picklo.
Credit Uncle Sam with spurring a new golden era in supercomputing. In the early 1990s, the government decided to push supercomputer makers to build future generations from off-the-shelf parts. The program, dubbed Accelerated Strategic Computing Initiative (ASCI), set in motion a wave of innovation surrounding industry standards and computing schemes that are equally at home in a nuclear lab or accounting department of a global corporation.
That has spurred computer makers to continue to fund their supercomputer businesses. "If the government's strategy was to build niche computers [using proprietary parts and designs], we wouldn't play," says Winston Prather, Hewlett-Packard's (HPQ ) vice-president for high-performance computing. HP, thanks to the Compaq acquisition, has approximately 40% of a $5 billion supercomputer market growing approximately 10% annually, according to IDC.
MAC INSIDE. The IBM machines that will be delivered to Lawrence Livermore by 2005 will use some of the same parts that go into a PC. The first supercomputer, ASCI Purple, will comprise 12,544 microprocessors using chip technology that IBM supplies to Apple Computer (APPL ) for the Macintosh. When these chips are lassoed together, the machine will do a top speed of 100 teraflops. The second system, Blue Gene/L, will consist of 130,000 processors running Linux and will hit about 360 teraflops running flat out.
The machines "reflect the industry's transition from the old Cray mainframe to killer microprocessor technology," says Seager. "That's what we've been riding." In 1992, Seager says, Livermore paid $16 million for a 50-gigaflop supercomputer. The new machines will provide a 10,000-fold improvement in power for about 20 times what the lab paid about a decade ago.
The economics are enticing. General Motors went from a stable of some 34 supercomputers including powerful Cray machines to systems from IBM and Silicon Graphics (SGI ) comprising 1,400 off-the-shelf processors. GM has slashed its supercomputing bill by $7 million, according to Kirk Gutmann, a product-information officer for GM. The machines help the auto giant take new cars from final design to production in just 18 months, down from the 42 months it took several years ago.
"EXCITING TIME." Mainstream technology doesn't mean middle-of-the-road performance. Computer scientists say technology trends are enabling the construction of even more powerful machines that could break the petaflop (a thousand teraflops) barrier by decade's end. "It's an exciting time," says Jack Dongarra, a University of Tennessee computer scientist who helps compile the list of the fastest computers. "Building a petaflop computer is the next major challenge."
Considering how fast the technology is evolving, it may not be too long before that challenge is met.
By Ira Sager in New York
Copyright 2000-2002, by The McGraw-Hill Companies Inc. All rights reserved.
Terms of Use Privacy Policy
Used with permission of businessweek.com
businessweek.com |
