Well, it takes more than raw computing power. That was the message delivered by one supercomputer user at the recent festival of raw computing power, SC2003.*

The World's 10 Speediest Supercomputers
	Facility	Number of Processors	Manufacturer
1	Earth Simulator Center's Earth-Simulator (Japan)	5,120 processors	NEC
2	Los Alamos National Laboratory's Asci Q	8,192 AlphaServer processors	HP
3	Virginia Tech's TerASCale Computing Facility	1,100 dual-processor Apple G5 machines	Homemade
4	NCSA's Tungsten	2,500 Xeon processors	Dell
5	Pacific Northwest National Laboratory's Mpp2	1,936 Itanium processors	HP
6	Los Alamos National Laboratory's Lightning	2,816 Opteron processors	Linux Networx
7	Lawrence Livermore National Laboratory's MCR Linux Cluster	2,304 Xeon processors	Linux Networx/Quadrics
8	Lawrence Livermore National Laboratory's ASCI White	8,192 Power3 processors	IBM
9	National energy Research Scientific Computing Center's Seaborg	6,656 Power3 processors	IBM
10	Lawrence Livermore National Laboratory's xSeries Cluster	1,920 Xeon processors	IBM
Source: Top500.org

While some organizations are using systems capable of teraFLOPS (trillion floating-point operations per second) to process terabytes of data, true insight requires creativity, said Donna Cox of the National Center for Supercomputing Applications (NCSA) in her keynote speech here. Cox, a senior research scientist and associate director of experimental technologies, said the real challenge is representing data in such a way as to make it useful for discovery.

Things were simpler when computing "was a small localized group activity," Cox said. Now, with multidisciplinary groups, dispersed internationally, collaborating virtually, new approaches are needed to visualize large amounts of incongruous data. One of the keys, she said, is creating visual metaphors that convey characteristics or concepts.

She used as an example an advertisement for a German beer. In the ad, a bottle of beer is shown in a champagne ice bucket. "[With] the beer engulfed in a champagne bucket, qualities of champagne are associated with the beer," Cox said. "We don't see the champagne, but the attributes are mapped onto the beer."


500 Approaches
Still, researchers love their floating-point operations. Much of the buzz at SC2003 was about the latest Top500 Supercomputer Sites list. Several innovative systems tailored for life sciences made the grade this time. Roaring in at number 3: Virginia Tech's Terascale Computing Facility, a homebuilt supercomputing cluster consisting of 1,100 Apple G5 computers, running Mac OS X, with each node having dual 2GHz 64-bit PowerPC processors, 4 GB of memory, and 160 GB of storage. Mellanox's InfiniBand technology supplies the primary communications fabric. The Virginia Tech system, just announced last month, has a peak measured performance of 10.2 teraFLOPS.

Virginia Tech will use the cluster to support work in computational chemistry, molecular statistics, and molecular modeling of proteins. The Virginia Tech system is only the third ever to be benchmarked with a peak performance that exceeds 10 teraFLOPS.

Another notable entry making the list for the first time is IBM's Blue Gene/L Prototype. The system is ranked 73rd on the list, with an official measured peak performance of 1.435 teraFLOPS. IBM's Blue Gene program is devoted to developing new hardware and new protein folding algorithms.

A number of university supercomputer systems, which will be dedicated to scientific research, also made the Top500 list for the first time. Among the new university entrants, all in the top 100, are the Chinese Academy of Science, the Korea Institute of Science and Technology, and the University of Liverpool.

Adoption of clusters continues to increase in high-performance computing environments. Seven of the top 10 computers in the new Top500 are clusters. On last November's edition of the semiannual list, there were only two. All told, 208 cluster systems made the most recent rankings.

For sheer performance, the total combined processing power of all the Top500 supercomputers would be 528 teraFLOPS. Six months ago, when the previous list was released, the total combined power was 375 teraFLOPS.

SC2003 had the slick booths associated with most trade shows, but attendees could be seen in intense discussions with vendors about high-performance computing issues. The conference network, Scinet, required installation of 55 miles of fiber, supporting a 40Gbps backbone. Next year, organizers hope to undertake a new initiative called StorCloud that would provide bandwidth of 1 terabyte per second and demonstrate innovative management and allocation technologies.

* SC2003, Nov 15-21, Phoenix