SCIENCE

MANNHEIM, Germany; KNOXVILLE, Tenn.; & BERKELEY, Calif. - In what has become a much-anticipated event in the world of high-performance computing, the 21st edition of the "TOP500" list of the world's fastest supercomputers was released today. The Earth Simulator supercomputer built by NEC and installed last year at the Earth Simulator Center in Yokohama, Japan, with its Linpack benchmark performance of 35.86 Tflop/s ("teraflops" or trillions of calculations per second), retains the number one position. The number 2 position is held by the re-measured ASCI Q system at Los Alamos National Laboratory. With 13.88 Tflop/s, it is the second system ever to exceed the 10 Tflop/s mark. ASCI Q was built by Hewlett-Packard and is based on the AlphaServer SC computer system. The new TOP500 list, as well as the former lists, can be found on the Web at http://www.top500.org/. The number of systems in the TOP500 list using Intel processors grew in the last six months from 56 to 119. This is a major shift in this marketplace. With this increase, the Intel processor family joins IBM's Power architecture and Hewlett-Packard's PA- RISC chips as one of the dominant processors used in HPC systems. At position 5, the upgraded IBM SP system with 6,656 processors at the U.S. Department of Energy's National Energy Research Scientific Computing Center (NERSC) at Lawrence Berkeley National Laboratory recorded a new IBM Linpack performance record of 7.3 Tflop/s. A new Linpack implementation was responsible for this result, which even exceeds the old measurement of the larger sister system, ASCI White at Lawrence Livermore National Laboratory (at position 4), which has 8,192 processors. Both systems have otherwise identical architectures. With this new Linpack implementation, ASCI White could also achieve a better result, but the system was unavailable for Linpack testing due to its large workload. Therefore, the Intel Xeon-based MCR cluster at Lawrence Livermore National Laboratory outranked ASCI White and achieved the highest position ever for any cluster (position 3). It uses a Quadrics interconnect and was manufactured by Linux Networx. Two notable newcomers among the top 10 are Fujitsu's PrimePower HPC2500 system at the National Aerospace Laboratory of Japan, the largest new Japanese system at number 7, and at number 8 the highest ranked Itanium-based system, produced by Hewlett-Packard and installed at Pacific Northwest National Laboratory. Gaining entry into the top 10 positions on the new list now requires achieving a Linpack performance just shy of 4 Tflop/s. Already 59 systems report a Linpack performance exceeding 1 Tflop/s, and 157 systems report a theoretical peak performance above 1 Tflop/s. Performance levels at all positions of the TOP500 list keep growing impressively, driven by Moore's Law. The total combined performance of all 500 computers on the list is 375 Tflop/s, compared to 293 Tflop/s just six months ago. The slowest system on the newest listing is now 245.1 Gflop/s (gigaflops or billions of calculations per second), compared to 195.8 Gflop/s six months ago. In terms of total performance of all the installed systems, the latest TOP500 edition still shows IBM as the clear leader with 34.9 percent, ahead of HP with 24.1 percent and NEC with 11.7 percent. Hewlett-Packard was able to hold on to its lead in terms of the number of systems on the list by the slightest possible margin - one system! A total of 159 systems on the TOP500 list were installed by Hewlett-Packard, compared to 158 systems by IBM. This time SGI is third in this category with 54 systems. No other company was able to gain more than 6 percent in either of these two categories. The number of clusters in the TOP500 grew again, now totaling 149 systems. Of these, 23 clusters are labeled as self-made, as they are designed and assembled by the end users themselves. The fact that the remaining 126 clusters were assembled and sold by companies reflects the importance this class of systems has gained for the manufacturers active in this market segment. The new Cray X1 made also its debut with 10 systems listed. It is the only computer system currently being built in the U.S. that uses vector processors, as the Earth simulator does. Therefore, the Cray X1 is sometimes seen as a possible U.S. answer to the Earth Simulator. The highest ranked Cray X1 appears on rank 112. Seven of the TOP10 systems, 51% of all 500 systems, and 55% of the total performance are installed in the United States. Also, 92% of all 500 systems are produced in the United States. These market shares demonstrate the continued and clear leadership of the U.S. in the usage and production of high performance computing. The TOP500 list is compiled by Hans Meuer of the University of Mannheim, Germany; Erich Strohmaier and Horst Simon of NERSC/Lawrence Berkeley National Laboratory; and Jack Dongarra of the University of Tennessee. The list was released today and will be presented in detail at the ISC2003 Conference in Heidelberg, Germany, which starts Wednesday, June 25, and continues through Friday, June 27.