IBM exceeds expectations with supercomputer
Big Blue's supercomputer designed for simulating nuclear explosions turns out to be 23 percent faster than anticipated when the project began.
- Shankland covered the tech industry for more than 25 years and was a science writer for five years before that. He has deep expertise in microprocessors, digital photography, computer hardware and software, internet standards, web technology, and more.
The $110 million machine, called ASCI White, was assembled at IBM's test facility in Poughkeepsie, N.Y., and is being reassembled at its final home, Lawrence Livermore National Laboratory (LLNL) in Livermore, Calif. The machine takes up two basketball courts worth of floor space, weighs 106 tons, and has 8,192 CPUs.
It was clocked at 12.3 trillion calculations per second, nearly a quarter faster than the 10 "teraflop" figure specified in the contract, IBM said.
The results are more than academic. The massive machine is a scaled-up version of an upcoming RS/6000 SP IBM server code-named "Nighthawk 2," due to debut in July. The Nighthawk 2 will be the first commercially released computer to use IBM's new Power3-III processor.
As previously reported, ASCI White is one of a series of supercomputers commissioned by the Department of Energy to test nuclear weapons without explosions. The program, the Accelerated Strategic Computing Initiative (ASCI), aims to fund computer makers to create supercomputers out of large numbers of comparatively ordinary parts.
DOE cut costs for ASCI White by leasing it from IBM for two years instead of purchasing it outright, LLNL officials said earlier.
High-performance technical computing is a tough market to crack, with extremely high demands for performance and expertise. However, the ASCI program is considered a prime way to get ahead in the market because the U.S. government essentially subsidizes research and development costs in an effort to push technology ahead as fast as possible.
Other companies involved in the ASCI program include SGI, which built a machine called Blue Mountain at Los Alamos National Laboratory in Los Alamos, N.M., and Intel, which built ASCI Red at Sandia National Laboratories in Albuquerque, N.M. Like ASCI White, both machines contain thousands of processors.
In the future, SGI is in the running to land a contract to build a 30-teraflop machine at Los Alamos. However, Sun Microsystems, seeing SGI's financial weakness, also has staked a claim.
AMD, while not vying for the ASCI program, is sponsoring supercomputer research with a machine at the University of Kentucky.
Livermore researchers have been testing a smaller version of ASCI White called Baby Huey, which has just 16 nodes, or computing units.
ASCI White is much bigger, though. It has 512 nodes, each containing 16 Power3-III CPUs. The computer's memory size of 6.2 terabytes is about 97,000 times that of a 64MB PC. It has 160 terabytes of storage space, about 16,000 times the amount of a desktop computer with a 10GB hard disk. It draws 1.2 megawatts of power, as much as about 1,000 homes.
Designing a home for behemoths such as ASCI White isn't easy. In particular, getting thousands of processors to share information requires huge amounts of wiring--2,000 miles in the case of ASCI White--and a new interconnection scheme code-named "Colony." Cooling these systems also requires fans and systems that make a meat warehouse look toasty.
Even more difficult is making sure that software works well on the machines. The more readily a computing task can be broken into independent subproblems, the more amenable the task is to a computer with lots of processors.