During the last few years several manufacturers promised a Teraflop Computer soon to come. When establishing our TOP500 list we questioned if all 500 computers put together would reach 1 Tflop/s ([tex2html_wrap1306] floating point operations per second). As it turned out with 1.47 Tflop/s this mark was even exceeded in our second TOP500. The entries in the first TOP500 just add up to 1.17 Tflop/s.
A special model by Fujitsu in cooperation with NAL, the so called `Numerical Wind Tunnel' with 140 processors and a Linpack performance [tex2html_wrap1308] of 124.5 Gflop/s made the top of the list, followed by four CM5 systems by Thinking Machines with 1024, 544 as well as two times 512 processors. The most powerful installed CM5 system in Los Alamos reached [tex2html_wrap1310] of almost 60 Gflop/s, the other three CM5-systems a bit more than 30 Gflop/s.
At the bottom of the list is the VP 200EX at the University of Dresden with 472 Mflop/s, so the list includes computer systems with two scales of different performance results. On the other hand the listed systems are much more powerful than all workstations available today. The entry level model from Cray Research, Convex or other smaller models of the MPP manufacturers like nCube, Parsytec, Kendall Square Research and others are not listed, which means that we have reached our goal to include only true supercomputers into our list.
Looking at the figure of systems installed per continent (Fig. 9.1) it stands out that Japan with 21%even falls behind Europe (27%) and has noticeable less supercomputers than the USA (48%). Our statistics of the last years (see chapter 
)-based on system counts-showed clearly different results but it must be taken into consideration that all small vector computer models by Fujitsu, NEC and Hitachi-which did not make it into the TOP500 this time-had been included in these old statistics.
Viewing the installed performance (Fig. 9.2) the dominance of the USA with 60%in comparison to Japan (24%) and Europe (13%) is even more evident. This is because of the larger amount of MPP systems in the USA where already 50%of all supercomputers belong to this computer category and surpass the vector computers in the USA with their accumulated performance by appr. 50%. Obviously this is a consequence of the HPCC program, prefering the selection of MPP systems at least at sites funded by the government. Even though 33%of the systems in Europe are MPP systems, their accumulated performance is still clearly smaller than the installed vector computer performance, according to their number these are 34%of all supercomputer performance installed. On the contrary almost only the traditional vector computers are installed in Japan, the MPP systems are rather under-represented with 12%altogether. This could change if the Japanese manufacturers would enter the market with their own MPP models.
[figure57]
[figure62]
Broken down into manufacturers (Fig. 9.3 and 9.4) Cray Research is up front as well in numbers of systems (40%) as in installed performance (30%). While the number of installed Cray systems is clearly ahead of MPP systems market leader Thinking Machines (14%), Fujitsu with Siemens Nixdorf (10%), Intel (9%), NEC (6%), Hitachi (5%) and KSR (5%) it looks totally different compared to the installed performance [tex2html_wrap1314]. Here TMC comes (27%) shortly behind Cray Research (30%), again because of the partly overproportionately powerful systems in the USA, ahead of Fujitsu with SNI (15%), Intel (10%) and NEC (9%). nCube and DEC as well as the `Europeans' Meiko and Parsytec could place less than ten systems into the TOP500. From IBM we expect an upswing in the future because of the parallel computer SP2.
[figure71]
[figure76]
After all 25%of all systems are used by the industry (Fig. 9.5, 9.6 and 9.7), a little less than in the university (31%) and research sector (29%). While the number of systems for classified operations are 4%, the installed performance is overproportionately high with 9%. The performance of supercomputers in the industry on the contrary comes to 13%, here smaller and middle sized vector computers are often used, the first industry user ranks on place 34 [l]#34 Ford: Y-MP C916/11256 in our TOP500 list. At the moment the industry is rather restrained using MPP systems, only 16%of these parallel computers are installed there and the accumulated performance is only 16%. From these figures can be concluded that the big MPP systems are not yet in use there. The only two industry segments, [l]Aerospace e.g.: #101, 102, 108 where MPP systems can already be found are aerospace and geophysics [l]Geophysics e.g.: #39, 41 in the USA.
Universities on the contrary have 39%of MPP systems in use with 33%of accumulated performance. The most powerful MPP parallel computers can be found in the research sector showing like the university sector 33%of the accumulated performance with only 26%of the installed MPP systems. The most powerful vector systems can also be found in the research sector, with the Numerical Wind Tunnel by Fujitsu up front.
[figure89]
[figure94]
[figure99]