Scientists seek revamped federal supercomputing effort

In April, Japan rocked the U.S. supercomputer sector, which traditionally has led the world in producing the fastest machinery, with the unveiling of the Earth Simulator, a $400 million product that outpaces the speed of U.S. supercomputers.

Now U.S. scientists are hoping that their government will boost spending in fiscal 2004 to keep pace with Japan, whose government paid the Japanese computer firm NEC to create the supercomputer. The Earth Simulator-so named because it aims to create a "virtual earth" that could simulate various climate conditions-matches the raw computing power of 20 of the fastest American computers combined.

"There is a fear that our technology leadership could go overseas" unless the United States invests in keeping its lead in high-end computing, said Jim Rottsolk, chairman and CEO of the Cray supercomputer company.

The United States has increased significantly its investment in non-defense, high-end computing research and development over the past five years, with an estimated $846.5 million expected to be spent in fiscal 2003, up from $462.2 million in fiscal 1998. The increase is spread throughout agencies such as the Energy Department and Defense Advanced Research Projects Agency.

The White House Office of Science and Technology Policy declined to comment on whether the Bush administration will continue to increase its investment in high-end computing because the formal fiscal 2004 budget will not be released until February, but Ray Orbach, director of Energy's Office of Science, has testified to Congress that keeping the United States' lead in supercomputing is his top priority.

The nation has fallen behind Japan in supercomputing, some scientists argue, in part because over the past 10 years the United States has focused on "distributed cluster computing," which involves a collection of interconnected computers that comprise a unified resource. Some government agencies considered this type of computing less expensive than investing in singular, large supercomputers.

Chris Jehn, vice president of government programs at Cray, said the problem with clustered computers is that the high-speed connections achieve less than 10 percent of the potential speed of a networked machine. Japan invested in building faster connections that enable the Earth Simulator to reach 50 percent to 70 percent of the machine's processing power, he said, and that is why Japan has taken the lead in producing the fastest supercomputer.

"Ten years ago, many thought clustered commodity computing was the future of supercomputing," Jehn said. "But as it turns out, challenging scientific and technical problems ... are not handled well by cluster computing." He cited cryptanalysis, the theory and art of cracking computer codes, as one such problem.

Cray customized its processors and connections within machines for higher speeds and hence is on track to eventually match the Earth Simulator's computing speeds, Jehn said.

On another front, Jehn noted that there is no U.S. supercomputing policy, a point that got Congress's attention last year. The fiscal 2002 Defense appropriations law required the National Security Agency to study whether a supercomputing strategy should be crafted. That study has been completed and is being reviewed within the Pentagon, Jehn said.