The supercomputer gene pool has expanded.
Of the top 10 systems on a list of the 500 most powerful supercomputers announced on Sunday, three machines are new, one is upgraded, and two are based on processors that have never before appeared on the list: IBM's PowerPC 970 and Advanced Micro Devices' Opteron.
The IBM chips show up in a $5.2m system officially named X but informally dubbed "Big Mac" that Virginia Polytechnic Institute and State University researchers assembled over two months from 1,100 dual-processor Apple G5 computers. That system ranked No. 3 with a performance of 10.3 trillion calculations per second, or 10.3 teraflops.
The Opterons are used in a 2,816-processor cluster built by Linux Networx for Los Alamos National Laboratory. Its sustained speed is 8 teraflops.
The No. 1 and No. 2 systems -- NEC's Earth Simulator and LANL's ASCI Q, built by Hewlett-Packard, maintained their positions with performance of 35.6 teraflops and 13.8 teraflops, respectively. Earth Simulator has held the top spot for four consecutive lists.
The Top500 list is compiled by researchers at the University of Mannheim, the University of Tennessee and Lawrence Berkeley National Laboratory.
The list, released twice each year on the eve of a supercomputing conference, shows the waxing and waning power of various manufacturers, the arrival of new technologies and the constantly increasing power of the world's number-crunching elite.
The total performance of the top 500 systems has been increasing exponentially, growing by a factor of 10 about every four years. In 1993, the total performance was just over 1 teraflop; it reached 10 teraflops in 1997, 100 teraflops in 2001 and list compilers projecte it to cross the 1000 teraflop -- or 1 petaflop -- boundary in about 2005.
Also according to the projections, a system must have about 1 teraflop performance to make it on the list at all. Currently 131 systems surpass that level.
Machines on the system are ranked with a speed test called Linpack that involves solving linear equations.
Looking beyond the coveted top spot, either HP or IBM has the strongest presence in the list, depending on which measurement is used. HP had the most systems on the list, 165 to IBM's 159 -- not much change from the 159-156 gap in June.
But in terms of the fraction of the list's total computing power -- 530 teraflops -- IBM was the leader with 35 percent to HP's 23 percent. That gap is nearly unchanged from June, when the percentages were 35 and 24 for the two companies.
Intel's Xeon surpassed HP's PA-RISC to become the most often used processor, rising from 76 systems in June to 152 systems in November. PA-RISC, meanwhile declined from 128 to 114. Intel's high-end Itanium 2 processor -- with which HP is gradually its PA-RISC line -- grew from 19 systems to 32.
IBM's Power4, meanwhile, dropped from 64 systems to 50.
The US government, alarmed by Japan's top ranking, has begun a programme to spur new research. The government is funding IBM, Cray and Sun Microsystems to prototype new systems.
US officials who long have placed controls on supercomputer exports to China likely won't be encouraged by work in that country by Beijing-based Legend Group.
Six months ago, the most powerful Chinese system was Legend's DeepComp 1800 at the Academy of Mathematics and System Science in China, a system with 512 of Intel's 2GHz Xeon processors that ranked No. 51 with a sustained speed of 1 teraflop.
On the latest list, No. 14 is Legend's DeepComp 6800 at the Chinese Academy of Science, built from 1,024 of Intel's 1.3GHz Itanium 2 processors.
Among other changes on the list:
• The first module of IBM's Blue Gene/L supercomputer placed 73rd with a sustained performance of 1.4 teraflops. When the full Blue Gene/L is completed in 2005, a system with 128 times as many processors as the current prototype, IBM expects performance of 360 teraflops, enough to top the list by a wide margin.
• At least three systems on the list are clusters of lower-end computers connected by the high-speed InfiniBand networking technology, including the No. 3 Apple Computer machine.
• The turnover rate on the list is "well above average," its compilers said in a statement. The pokiest system on the list, at 403 gigaflops, was No. 290 in June.
• Opteron systems also show up at Nos. 93, 116, and 247.
Clusters of independent lower-end systems linked with a high-speed network continue to gain in importance. Such systems first appeared on the list in 1998 but now account for close to half its systems and total performance.
