The TOP25 Supercomputer Sites

In this short note the Top25 supercomputer sites worldwide are introduced. This list of the Top25 sites has been compiled based on the information given in the Top500 report. First we briefly explain how this list was established and comment on recent changes. For most of the top twentyfive supercomputer centers we provide a short description of facilities, equipment, and mission.

The Top500 [1] provides an opportunity to rank the top supercomputer sites worldwide. In the Top500 report the 500 highest performing supercomputers are listed. The measurement of performance is based on the Linpack benchmark as reported in [2]. All the information here is based on the Top500 list of November 1996. We use results from this list because because it provides the most comprehensive list of supercomputer sites. Also real Linpack figures are available for all the machines on the list. Some of the limitations of using Linpack and relying on the Top500 list are discussed elsewhere [3].

An alternative ``List of the World's Most Powerful Computing Sites" is compiled by Ahrendt [4]. This list is based on the NAS Parallel Benchmarks code BT [5]. There is considerable overlap between these two lists. BT as benchmark generally does not perform as well on highly parallel machines as Linpack, when compared to vector machines. Also BT results are not always immediately available on new machines, so [4] resorts sometimes to estimates and extrapolations. On the other hand the list [4] is updated more frequently. In general, however, the results and the ranking are fairly consistent. The main point here is that we will give some additional information about the supercomputer sites, beyond just a listing of machines. This provides a general overview and pointers to further reading, on what the most powerful supercomputers on earth are actually used for.

List of the TOP25 Sites

The TOP25 list of supercomputer sites is given in Table 1. This list has been established by simply adding the Linpack Rmax\ performance in Gflop/s of all supercomputers installed at a given site. Generally under a ``site" we have combined supercomputers, which are installed in the same geographical location, and belong to the same organizational unit. Thus all machines belonging to a university on the same campus were added, even though they might be in different departments. The previous ranking from November 1995 is given in the second column (see [6]).

Table 1: TOP25 Supercomputer Sites

The list does not contain any of the vendor machines. Most of the supercomputer vendors have substantial compute capabilities, which would make the TOP25 centers list. However, the intent of this list is to give an indication where most compute power in terms of scientific and research applications is concentrated. Therefore we decided to list the vendors separately in Table 3.

In all tables the column ``machines" lists the machines whose performance have been added to reach the total performance for a site. The integers refer to the ranking of these supercomputers on the Top500 list. The performance column lists the aggregate performance of all the machines at the site in Linpack Rmax-Gflop/s. An overview of many of the supercomputers in use is [7].

There are several intriguing observations one can make from Table 1. In order to qualify as a top supercomputer site, and installation must have at least a machine with about 70 Gflop/s performance. This is almost twice the cutoff one year ago, which was about 35 Gflop/s. Three years ago the cutoff was only 13.7 Gflop/s, and 70 Gflop/s would have placed an institution on rank two. There has been a tremendous acceleration of available cycles at the top supercomputer centers. In 1996 again the number of machines at Top25 sites and their share of the total performance in Gflop/s increased slightly.

Another significant change is in the geographical distribution. In 1996 the most important change was that there were three European centers which entered the Top25 list. In 1995 there were no European sites among the Top25Table 2 shows the change in the geographical distribution of the centers.

Table 2: Geographical Distribution

The list also shows how much U.S. government spending dominates the supercomputing world. All 13 U.S. sites directly or indirectly are funded by the U.S. government. There are 9 U.S. government laboratories/centers (5 Department of Energy, 1 classified, 1 NASA, 2 Dept. of Defense), and the five U.S. universities receive their support for supercomputers from the NSF or DoD (Minnesota). However, also the foreign sites are also all falling into the same category, and are government institutions in their countries.

Vendor Sites

Most of the supercomputer vendors maintain substantial benchmarking capabilities. These are usually distributed worldwide. Since the vendor centers are geared towards benchmarking and internal software development, we in 1995 decided not to list them in the same list as the Top25 supercomputer centers, which are geared towards research. In Table 3, we list the all vendor sites. Only the first two, Cray and IBM, would have made it to the Top25\ list. However, we believe that the vendors no longer report benchmarking machines for the Top500 since there is a limit to the number of vendor machine whihc can be reported.

Table 3: TOP Vendor Sites

Background Information about some of the TOP25 sites

Wherever available some short summary of the mission and the environment of the Top25 sites can be found through the list below. If not noted otherwise, this information was gathered from home pages of these sites on the world wide web (WWW). Many of the supercomputer sites have created such home pages. In order to facilitate browsing the net, URLs for site home pages are given.

• University of Tsukuba, Center for Computational Physics - 408.0 Gflop/s
• University of Tokyo, Japan - 315.4 Gflop/s
• National Aerospace Laboratory Numerical Wind Tunnel, Tokyo, Japan - 239.7 Gflop/s
• Japan Atomic Energy Research Institute, Japan - 217.8 Gflop/s
• National Security Agency, Fort Meade, Maryland - 156.471 Gflop/s
• Los Alamos National Laboratory - 166.9 Gflop/s
• Pittsburgh Supercomputing Center - 157.7 Gflop/s
• Oak Ridge National Laboratory - 154.3 Gflop/s
• Sandia National Labs, Albuquerque - 143.4 Gflop/s
• University of Minnesota - 126.4 Gflop/s
• Lawrence Livermore National Laboratory - 123.7 Gflop/s
• ECWMF, Reading, UK - 120.8 Gflop/s
• Univ. Stuttgart, Germany - 117.3 Gflop/s
• CNRS/IDRIS, France - 112.9 Gflop/s
• DOD/CEWES, Vicksburg - 106.9 Gflop/s
• Natl. Lab. High Energy Physics, Japan - 98.9 Gflop/s
• Kyushu University - 94.3 Gflop/s
• Cornell Theory Center - 88.4 Gflop/s
• Tohuku University - 85.8 Gflop/s
• NCSA, Univ. of Illinois - 85.0 Gflop/s
• NERSC, Lawrence Berkeley Nat. Lab. - 81.5 Gflop/s
• Maui HPCC - 78.8 Gflop/s
• Atmospheric Env. Serv., Dorval, Canada - 73.9 Gflop/s
• Caltech/JPL - 69.6 Gflop/s
• National Laboratory for High Energy Physics (KEK), Japan - 98.900 Gflop/s
• Cornell University, Ithaca, NY - 96.024 Gflop/s
• University of Tokyo, Japan - 83.776 Gflop/s
• Maui High Performance Computing Center, Kihei, Hawaii - 83.700 Gflop/s
• Japan Atomic Energy Research Institute, Japan - 75.918 Gflop/s
• Tohuku University, Japan - 71.300 Gflop/s
• Atmospheric Env. Serv., Dorval, Canada - 58.670 Gflop/s
• Nagoya University, Japan - 58.509 Gflop/s

References

1

J. J. Dongarra, H. W. Meuer, and E. Strohmeier. TOP 500 supercomputer sites. Technical Report RUM 48/96, Computing Centre, Univ. of Mannheim, 68131 Mannheim, Germany, November 1996.

2

J. J. Dongarra. Performance of various computers using standard linear equations software. Technical Report CS-89-85, Department of Computer Science, Univ. of Tennessee, November 1993.

3

J. J. Dongarra, H. D. Simon. High Performance Computing in the U.S. in 1995. Supercomputer 63, Vol. 12, No. 1, pp. 16 - 22, January 1996.

4

Gunter Ahrendt. List of the world's most powerful computing sites. WWW URL http://www.interactive.net/ gunter/ November 1996.

5

D. H. Bailey, J. Barton, T. Lasinski, and H. (editors) Simon. The NAS parallel benchmarks. Technical Report 103863, NASA Ames Research Center, Moffett Field, CA 94035, July 1993.

6

H. D. Simon. The TOP25 Supercomputer Sites. Supercomputer 63, Vol. 12, No. 1, pp. 38 - 47, January 1996.

7

K. E. Gates and W. P. Petersen. A technical description of some parallel computers. Technical Report IPS 93-15, IPS, ETH Zuerich, Switzerland, IPS Zentrum, CH 8092 Zuerich, Switzerland, November 1993. (to appear in Int. J. High Speed Comp., 1994).

Horst D. Simon
Lawrence Berkeley National Laboratory
Mail Stop 50B 4230
University of California
Berkeley, CA 94720