At the end of 1994 the Numerical Wind Tunnel (NWT) from Fujitsu with 140 processors was at the top of the list again. Half a year earlier the Intel Paragon XPS/140 at the Sandial National Labs held this position. The NWT took over the lead again not because of changed hardware but only because of improved benchmark performance. Compared to the June 1994 list, the November list showed that the ranking of the first 40 systems were rearranged only due to changed LINPACK performances and there were no new en tries!
Now we first try to find out how fast the computing power of the indiviual machines present in the TOP500 and the total market size are growing. For this, we plot the performance of the systems at positions 1, 10, 100 and 500 in the list as well as the total accumulated performance. In Fig. 1 the curves of position 1, 100 and 500 show on the average an increase of a factor of two within one year. For a workstation processor an increase of the computing power of a factor of tw o within 18 month can be seen at present. This means that the increase in performance seen in the TOP500 can be explained only partially by the increase of the individual node performance. As an additional effect, we see an increase in the number of processors being the reason for the total factor of two in 12 month seen in the TOP500 .
Figure 1: The performance over time as it can be seen in the TOP500 .
In contrast to this, we find an increase factor of only 1.7 per year for the curves of position 10 and of the total accumulated performance. In general we see that during the last two years the individual performance of positions 100 to 500 in the TOP500 growed faster than the overall accumulated performance. As the first 50 systems account for more than half of the total accumulated performance, we take this as an indication that the performance of the very high-end systems did not increase as much as th e other systems. This is an indication that the number of processors for systems at the top of the list is presently just slightly changing. Looking at the different TOP500 lists, we see no systems with more than 1,000 processors in the last two years with one exception being the Intel Paragon at the Sandia National Labs. The market for the top of the end MPP systems with much more than 100 processors seems to be quite small and presently just slowly growing.
Now we look at the replacement rate at which new systems enter the list and at which systems at the end of the list fall out due to their performances being too small. We show in Table 1 the positions of the systems which were at positions 500, 100, and 10 in one of the last 4 lists. Looking at Table 1 we see that during the three revisions of the list, half of the entries were removed. This gives a very stable replacement rate of 20% for each ne w list, and thus only the first 80% of the list will be present in the successive issue of the list half a year later. A similar and sometimes even higher replacement rate can be seen for positions 100 and 10 in Table 1. The major exception is position 10 in the last two lists from 1994 as we had no new systems in the first 40 positions, but only rearrangments due to changed LINPACK performances. We now can roughly estimate how long the systems will be present in the TOP500 . In Table 2 we show the minimal position a system has to have to remain in the TOP500 for one to five years. A system that should remain in the TOP500 for 4 years should now be at least around position 79.
Table 1: In order to visualize the replacement rate, the positions over time for entries 10, 100 and 500 are shown for all lists. Brackets denote that the kind of system was not present in that list because it was not yet on the market.
Table 2: The estimate for the minimal position for a system, so that it remains n years in the TOP500 , based on a replacement rate of 20.5% each half year.