Asci Red king of supercomputers

Cray T3E systems from SGI took second, third, and fourth place on the semiannual list. Numbers two and three are at unspecified, classified U.S. government facilities. IBM-powered machines also figure prominently on the list.

The top 500 list is written by Jack Dongarra, Hans-Werner Meuer, and Erich Strohmaier, supercomputer experts who have updated it twice each year since the list was founded in 1993.

The top 500 list is a rarefied place. Just to get on, you have to have a machine that can perform 17 billion calculations per second. But if you want that performance on your own desktop, it's just a matter of time: a top-of-the-line desktop computer today could have made it onto 1993's top 500 list.

Although experts debate the validity of the rankings, high placement on the list typically comes with bragging rights for the companies and institutions that designed the machine. These computing behemoths are generally highly customized machines built through the efforts of both the computer vendor and the institution that eventually uses it.

Among other statistics from the list:

• SGI was a big winner in the top 500 list, with 183 of the systems. SGI machines are responsible for half the total computing power on the list. Sun Microsystems has the second-highest presence, with 126 systems, up from 111 in June.

• IBM gained systems on the list, moving from 75 systems in June to 105 now.

• The first Japanese computer is a Hitachi/Tsukuba at 14th place.

• Industry has more supercomputers than academia. While research computers on the list dropped from 94 to 90 machines since June, industry computers increased from 178 to 206.

Two contenders--Blue Mountain, Los Alamos National Laboratory's main supercomputer, and Blue Pacific, a machine from IBM and Lawrence Livermore National Laboratory--are designed to be about four times faster than Sandia's machine. Unfortunately for both, the performance data isn't in yet.

Lawrence Livermore spokesman David Schwoegler said his lab was unable to provide a performance measurement because its computer didn't run the measurement program while it was assembled at an IBM facility in Poughkeepsie, New York. The machine, based around IBM's Power3 processors, has been shipped to Livermore, but now is only two-thirds reassembled.

The Los Alamos Silicon Graphics machine is fully built, but LANL spokesman Jim Danneskiold refused to disclose any performance numbers until the SC98 supercomputing conference, which begins Saturday in Orlando, Florida.

Erich Strohmaier, who helps maintain the top 500 list, said he didn't know where the computers were, but added, "The National Security Agency is certainly one large consumer of computing power for the government, but most likely not the only one."

All three of the top machines are the result of an Energy Department project to develop computers that can simulate the effects of massive nuclear explosions. The computers have come to replace underground nuclear testing. Asci Red, meanwhile, is likely the last of its line. Intel dropped out of supercomputing development after the machine was completed.

The Avalon computer, now running 140 Alpha chips and the Linux operating system, moved up from 315th place to 114th, courtesy of a memory upgrade and additional processors. However, a machine at Sandia National Laboratories that is also made of a cluster of Alpha processors beat out Avalon by taking 98th place.

The list ranks computers according to how fast they perform on "benchmark" tests, a standard yardstick that allows different computers to be measured solving the same computing problem. The top 500 list uses the Linpack benchmark, originally developed in the 1970s, in which the computer has to solve a standard mathematical problem.

Some observers, including Dick Russell at Tera Computer, have questioned the validity of using Linpack to rank the systems. Russell said the benchmark doesn't actually reflect computer performance in real-world applications.

But Chris Willard of International Data Corporation said Linpack is "as good a definition as any" for comparing computers' relative performance. He did caution, however, "It's not a good measure to determine how fast your computer will run on a given application, but that's almost always true with benchmarks."

Strohmeier said one major reason the top 500 list uses Linpack is simply that it's the benchmark that has by far the most results available.

Intel is an investor in CNET: The Computer Network, publisher of News.com.