Depending on how the term supercomputer is defined, there are between 10 and 20 supercomputer companies currently in existence. What follows is a rundown of some of the current crop of high-performance machines being offered by these companies.
Convex
Convex makes two lines of supercomputers, the SPP and the C4. The former is a parallel machine based on Hewlett-Packard's PA-RISC processors; the latter is a traditional vector supercomputer.
Cray Research
It's simply not possible to talk about the history of supercomputing without prominently featuring the name of Seymour Cray. For decades, the machines that he designed were the fastest in the world. In a field as volatile and com
petitive as the supercomputer field, he compiled an amazing record. He has since left Cray Research to form a new company, Cray Computer, that hasn't yet shipped any machines.
Cray Research still makes one of the fastest vector supercomputers, the 16-processor Cray C90, which is discussed in more detail in ``A Closer Look at Two Supercomputers.''
While massively parallel machines offer much-higher theoretical throughput, vector-based systems like the C90 are the workhorses of the supercomputer world. Getting peak performance from a vector machine is not easy, but scientific programmers have been working with these machines since the 1960s and are familiar with them.
In addition to offering its vector supercomputers, Cray has recently branched out into massively parallel architectures. Its T3D contains up to 256 DEC Alpha processors.
IBM
IBM recently announced the SP2, a parallel architecture based on Power2 RISC processors. The Power2 is a successor to the RS/60
00 chip, and the SP2 can contain as many as 128 of them. The SP2 machines supplant IBM's older ES-9000 line of traditional vector supercomputers.
Intel
In addition to offering its well-known CPUs for personal computers, Intel has a supercomputer division that has been selling parallel computers for 10 years. The current offering, which is called the Paragon, is discussed in more detail in the text box ``A Closer Look at Two Supercomputers.''
Meiko
Meiko is not as well known outside the supercomputer community as the other companies that are listed here. It sells a massively parallel machine called the CS-2, which is based on SuperSparc chips optionally connected to a pair of high-performance vector units (Fujitsu VP vector processors).
Silicon Graphics
Best known as a graphics company, Silicon Graphics, Inc., has long had a significant presence at most supercomputing sites. Its workstations have excellent graphics performance and
are often used to visualize the results generated by big supercomputers. With its new Power Challenge system, SGI is trying to expand its market to include high-performance computation as well as graphics.
Japanese Companies
Although supercomputing has been dominated by U.S. companies, the Japanese have also been active in the field. Their machines have received somewhat less attention than they deserve, partly due to the reluctance of U.S. government agencies to purchase them.
The three major suppliers of Japanese supercomputers are Fujitsu, Hitachi, and NEC. All are household names in the U.S., but they are not widely known in the U.S. for their computers. According to some benchmarks, the Hitachi S-3800 contains the fastest single processor currently available. NEC's line of supercomputers is called the SX-3, and Fujitsu offers a line called the VP2000. All three use vector processing and are well represented in the upper ranks of floating-point benchmark results.
Recent Casualties
Trying to market a supercomputer is a dangerous business; the field is littered with the skeletons of dead companies that have failed in the attempt. It's extremely difficult to design and deliver a supercomputer; it's harder still to make money doing it. And even when a company manages to do both, few have been able to maintain a successful effort for an extended period of time.
For example, Kendall Square Research (Cambridge, MA) announced a massive layoff in September 1994 and stopped producing its line of supercomputers. The end was accompanied by a series of ugly revelations about financial shenanigans; corporate officers apparently tried to hide the shakiness of the company by manipulating sales and inventory figures.
Thinking Machines Corp. was once the golden child of supercomputing; it had the newest and most radical approach and was stocked with the best and brightest minds from MIT and other top universities. The company's first two machines suppo
rted only the data-parallel model, which demanded that programmers rewrite the computationally demanding parts of an existing application. Although scientific programmers were originally skeptical (and many stayed that way), many applications achieved impressive performance on the CM-1 and the CM-2.
However, the company ran into difficulties on its next generation, the CM-5. These were much more conventional machines than their predecessors, but TMC ran into serious problems in designing the vector processors that provided the machines' raw performance. The company also had difficulty building the compiler support to handle existing data-parallel applications efficiently and bringing in new users. In the high-performance-computing arena, companies don't generally get second chances.
Cray T3D
photo_link (18 Kbytes)
The Supercomputer Makers
