News & Analysis
IBM outperforms NEC in supercomputer speed test
Chappell Brown
10/11/2004 9:00 AM EDT
The test, conducted last month at IBM's production facility in Rochester, Minn., clocked the latest version of the Blue Gene/L series at 36.01 teraflops on the benchmark, unseating NEC Corp.'s Earth Simulator, which has held the top position for two years, at 35.86 Tflops.
The new performance figure, one more point on the development curve of the scalable Blue Gene/L project, represents a "seminal event" for the supremacy of U.S. supercomputing, said Dave Turek, vice president of deep computing for IBM. "We're hopeful that this helps everyone to understand in policy circles and other areas of industry that from our perspective, the U.S. supercomputer industry is alive and well," he said.
His remarks address the continuing concern in government and academic circles that U.S. proprietary computer technology is falling behind that of other countries. Supercomputers are pivotal in maintaining a lead in military, scientific and industrial areas of technology.
NEC's Earth Simulator was the result of a Japanese government-sponsored effort and is dedicated to the specific task of simulating the earth's environment. Development of the mammoth machine, which occupies an area of 32,500 square feet, proceeded under the radar until the system suddenly appeared in 2002 as the fastest supercomputer by a wide margin. As recently as June, at the International Supercomputer Conference in Heidelberg, Germany, the next contender on the list, Lawrence Livermore Laboratory's Thunder supercomputer, clocked in at 19.94 Tflops, and Blue Gene took fifth place, at 11.68 Tflops.
Turek said that the target speed for the machine that IBM will deliver to Lawrence Livermore next May is 360 Tflops. An important aspect of the architecture is its scalability, along with power efficiency, he said. Blue Gene/L only consumes one-twenty-eighth the power of the Earth Simulator at about the same performance level.
"The Earth Simulator was part of a government-funded project that focused on a narrow class of problems rather than a commercial set of technologies," said Turek, who called IBM's machine more suited to a wide variety of problems. "What we don't do is engage in science fair projects to establish benchmarks. We let the benchmarks come to us, in the sense that what we try to do is build systems that resonate in a wide segment of the marketplace."
At Lawrence Livermore the IBM machine will run physical simulations of materials, Monte Carlo-based modeling and environmental simulations. The architecture uses many small processing nodes connected by a torus network. Two PowerPC microprocessors are put on a chip along with some memory, two such chips are put on a small card and the cards are mounted on boards that are subsequently housed in racks. Lawrence Livermore's machine will have 64 racks carrying 65,000 chips, for a total of 130,000 microprocessors.
The architecture, resembling cluster-type supercomputers, is ideal for molecular biology, bioinformatics, materials simulations and applications that fuse large amounts of sensor data, Turek said. IBM opted for the Linux operating system to make it easier for users to port applications.
"Blue Gene focuses on applications that can be scaled to many computing elements, but it is not good for large shared-memory-type applications," said Eric Kronstadt, director of exploratory server systems at IBM's T.J. Watson Research Center and project leader for BlueGene/L. "We have a list now of two or three dozen applications codes that seem to scale very well. We are evolving programming tools to take advantage of that."
