CNET también está disponible en español.

Ir a español

Don't show this again

Christmas Gift Guide
Tech Industry

IBM builds fastest Blue yet

IBM and a nuclear lab say they've built the fastest supercomputer yet, which can perform almost 4 trillion calculations per second.

IBM and Lawrence Livermore National Laboratory have stepped into the supercomputer lead with a machine that can perform almost 4 trillion calculations per second.

Vice President Al Gore boasted of the machine's prowess at a news conference today, saying it's the world's fastest computer, boasting a speed of 3.88 trillion computations per second, or about 15,000 times faster than the average desktop personal computer.

The machine, called Blue Pacific, is a mammoth IBM RS/6000 SP system with 1,464 "nodes," or individual processing units, each with four processors. A total of 5,856 processors are ganged together with a proprietary IBM interconnection hardware.

Blue Pacific has 2.6 terabytes of memory, hundreds of thousands of times as much RAM as the average desktop PC. And it has 75 terabytes of storage--enough to hold the entire Library of Congress.

The computer is big, too, taking up 8,000 square feet, weighing 105,000 pounds, using more than 4 miles of cables, and drawing 3,900 watts of power.

The RS/6000 is a long-standing proprietary workstation and server line from IBM. The same architecture is used in IBM's Deep Blue chess-playing computer.

The supercomputer's primary mission is simulating nuclear weapons explosions in an era when the real thing is banned. But Gore pointed to other possibilities the new computer will enable, including advances in medicine, manufacturing, aviation, and global climate change.

And IBM sees a wider market for the supercomputers as well--not just in research areas such as aerospace, pharmaceuticals, and petroleum, but also in more traditional business areas.

"We see high-performance computing as a growth area for IBM," said Mike Henesey, who runs IBM's scientific and technical computing program. Businesses will be interested in using the computers for numerically complicated analysis such as "data mining" or optimizing a portfolio on Wall Street, he said.

Smaller versions of the RS/6000 SP computers are running at companies such as Schwab, which uses a 100-node system a fourteenth the size of Blue Pacific to handle its Web-based stock trading, said Mike Borman, who's in charge of IBM's worldwide RS/6000 sales. And United Airlines has one for analyzing passenger traffic.

Blue Pacific is being assembled at the Lawrence Livermore lab in Livermore, California. It achieved the 3.88 "teraflops" (trillion floating point operations per second) rating at IBM's RS/6000 headquarters in Poughkeepsie, New York in September, said Borman, but the machine has been used to run actual Livermore code.

Nuclear weapons labs long have been one of the biggest customers for the supercomputers, but President Clinton's signing of the Comprehensive Test Ban Treaty has increased the Energy Department's appetite for big iron.

The new supercomputers are needed to run the vastly more complex simulation programs that the weapons labs use to certify that the nation's nuclear weapons will work as advertised as they grow older in the post-Cold War era. Earlier, computers were used for simulating weapons physics, but only using one- and two-dimensional simplifications. The new supercomputers will have to handle simulations in all three dimensions.

DOE is trying to increase supercomputer power faster than it otherwise would develop by paying companies to collaborate with the nation's three big nuclear weapons labs.

IBM got a $94 million contract in 1996 to build Blue Pacific with the Livermore lab, a machine designed to reach a speed of 4 teraflops. Silicon Graphics, a company with a huge presence in the supercomputing area with its Cray Research division, is working with Los Alamos National Laboratory in New Mexico on another 4-teraflop machine called Blue Mountain. And Intel already has its 1-teraflop machine up and running at Sandia National Laboratories in Albuquerque, New Mexico.

Wayne Pfeiffer of the San Diego Supercomputing Center said the Livermore/IBM computer is the world's fastest supercomputer as measured by peak performance, but that the effort by Los Alamos National Laboratory is likely to come in a close second.

Both the Lawrence Livermore and Los Alamos labs have been working hard to prepare their performance scores in time for the upcoming SC 98 supercomputer conference, which will start on November 7 in Orlando, Florida. When that conference starts, the top 500 supercomputer list will be updated.

More computers are on the way as part of DOE's Accelerated Strategic Computing Initiative. Last February, DOE announced that IBM won another contract, this one for a 10-teraflop machine called "Option White" at Livermore. Also on tap is a 30-teraflop machine slated for LANL, and ultimately, at the end of the 10-year, billion-dollar program, a 100-teraflop machine.

IBM isn't always the first company that comes to mind when speaking of supercomputers. But the company does have a significant presence on the June 1998 version of the top 500 supercomputers list. And IBM believes it has room to grow, Henesey said.

IBM's Scalable Parallel (the "SP" in the RS/6000 product name) architecture can extend all the way up to 1,000 teraflops--a petaflop, or quadrillion floating point operations per second.

IBM has 5,000 RS/6000 SP systems deployed worldwide, Henesey said.

This week, the Los Alamos lab--historically a good-natured competitor with its Livermore sister--fully assembled its computer, Blue Mountain, said LANL spokesman Jim Danneskiold. Blue Mountain, with 6,144 processors, has been running weapons code rewritten for the massively parallel machine as the machine was reaching its full size.

Using one sixth of its total computing power, Blue Mountain was able to run a simulated nuclear weapons test that analyzed physics interactions in an area divided into 30 million zones, he said. A similar simulation on the tried-and-true Cray Y-MP supercomputer was only able to run the simulation with 2.5 million zones.