SCIENCE

In what has become a closely watched event in the world of high-performance computing, the 24th edition of the TOP500 list of the world’s fastest supercomputers was released today at the SC2004 Conference in Pittsburgh, Pa. The new TOP500 list, as well as the former lists, can be found on the Web at http://www.top500.org/. After a close race to the finish line, the DOE/IBM BlueGene/L beta-System was able to claim the No. 1 position on the new TOP500 list with its record Linpack benchmark performance of 70.72 Tflop/s (“teraflops” or trillions of calculations per second). This system, once completed, will be moved to the DOE’s Lawrence Livermore National Laboratory in Livermore, Calif. It is closely followed by the Columbia system built by SGI and installed at the NASA Ames Research Center in Mountain View. Calif. Columbia clocked in at 51.87 Tflop/s. Both systems represent fairly new and exciting technologies, which enabled U.S. manufacturers and research institutions to regain the top spots on the TOP500 list from the Earth Simulator supercomputer, built by NEC and installed in 2002 at the Earth Simulator Center in Yokohama, Japan. The Earth Simulator, with its Linpack benchmark performance of 35.86 Tflop/s, had held the No. 1 position for five consecutive editions of the listing and is now shown as No. 3. The other positions in the top 10 also showed significant changes, such as the IBM-built MareNostrum cluster installed at the Barcelona Supercomputer Center, at No. 4 with 20.53 Tflop/s. The list also sees the return of the Virginia Tech X system. This system was often referred to as ‘SuperMac’ as it is built using Apple’s Xserve servers. It had temporarily disappeared from the listing 6 months ago, due to its unavailability during a major hardware upgrade, and is now listed at No. 7 with its new Linpack performance of 12.25 Tflop/s. Here are some highlights from the newest Top 500: The entry level for the TOP10 approaches 10 Tflop/s - only one system with less than 10-TFlop/s Linpack performance is in the TOP10. The entry point for the top 100 moved from 1.922 Gflop/s to 2.026 Tflop/s. The number of systems exceeding the 1 Tflop/s mark on the Linpack benchmark jumped from 242 to 398 and we expect that the next list in 6 months will only list systems exceeding 1 Tflop/s. Entry level is now 850.6 Gflop/s, compared to 624.3 Gflop/s six months ago. The last system on the list would have been listed at position 310 in the last TOP500 just six months ago. This exemplifies the continuous rapid turnover of the TOP500. Total combined performance of all 500 systems on the list has for the first time exceeded the 1 Petaflop/s mark. It is now 1.127 Pflop/s, compared to 813 Tflop/s six months ago. Other trends of interest: A total of 320 systems are now using Intel processors. Six months ago there were 287 Intel-based systems on the list and one year ago only 189. The second most commonly used processors are the IBM Power processors (54 systems), ahead of Hewlett-Packard’s PA Risc processors (48) and AMD processors (31). There are296 systems now labeled as clusters, making this the most common architecture in the TOP500. At present, IBM and Hewlett-Packard sell the bulk of systems at all performance levels of the TOP500. IBM remains the clear leader in the TOP500 list with 43.2% of systems and 49.3% of installed performance. HP is second with 34.6% of systems and 21% of performance. No other manufacturer is able to capture more than 7% in any category. A new geographical trend, which started during the last few years, emerges more clearly. The number of system in Asian countries other than Japan is rising quite steadily. In this list Japan is listed with 30 systems and all other Asian countries accumulated an additional 57 systems. However, Europe is still ahead of Asia with 127 systems installed. Seventeen of the systems in Asia are installed in China - up from 9 systems one year ago. The number of systems installed in the U.S. has also increased to 267 -- up from 247 one year ago. 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.