The penetration of the supercomputer market by MPPs is another often debated issue. The TOP500 again delivers hard data, which come somewhat as a surprise. In Table 10.5 the share in %in number of installations for MPP machines among the TOP500 is given. Here the PVP category includes all machines manufactured by Cray Research (with the exception of the Cray Superserver and T3D machines), all machines manufactured by IBM (with the exception of SP-1 machines), as well as all machines from Convex, Fujitsu, Hitachi, NEC, and Siemens-Nixdorf (Fujitsu). All others are counted as MPPs. The MPP category thus includes a number of machines, which should rather be labeled SMPs, but their inclusion does not change the overall picture significantly.
Overall MPPs account for 37%of all installed supercomputers. The installed base of MPPs in the U.S. is significantly higher. In the U.S. MPPs are with 51%already in the majority, in Japan on contrast MPPs are with 13%considerably underrepresented. Table 10.6 shows that irrespective of the country of installation MPPs are generally more powerful than PVP in terms of [tex2html_wrap1340]-Gflop/s.
The considerable higher number of MPPs in the U.S. compared to the world average is clearly a direct consequence of the U.S. HPCC program. How many additional machines in the U.S. have been installed because of the HPCC support, and other subsidies for parallel machines, e.g. (D)ARPA funding? Again this is difficult to estimate, but as before we can assume that Europe is an objective test market. This is probably a fair assumption, since at the time of this study the European HPCN Initiative probably had no effect on the selection of MPP machines versus PVPs. Under this assumption then the share of MPP machines in the U.S. would be only 32%, which is about 77 machines. Following this analysis one therefore could credit the installation of about 45 machines in the U.S. to the strong government support for massive parallelism. At about $5Mper installation this amounts to approximately $225Mof direct support for parallel machines in the U.S. from 1990 to 1993. This figure appears to be in the right order of magnitude. For example the General Accounting Office report on ARPA [8] lists a total of $55Mspent in connection with procurements of Intel and TMC supercomputers alone. The list of ARPA supported machines ends with installations in late 1992, and does not include any of the more recent large scale Paragon or CM-5 machines, which have made the TOP500, and which are also part of the HPCC program. Thus as a result of direct government support, the U.S. is clear world leader in the adaptation and use of MPP technology.
[table254]
[table260]