Of the 82 supercomputers in the TOP500 list installed in Japan, 40 systems were installed or upgraded in 1993 or 1994. 12 of these procurements were done in a special government program to strengthen Japanese applied supercomputer research mainly in the areas of biology and medicine, computer science, communications, chemistry, and materials science. More than 250 million US-Dollars were spent for supercomputers by the Japanese government in this program. In addition to this special program, government and industry continued to fund 28 new installations or upgrades in the course of regular procurement cycles.
Japanese supercomputers tend to be very powerful systems. One reason for this is that the budget for each procurement is rather big, often 20 million US-Dollar and more. The other reason is that the Japanese users and manufacturers continue to rely on vector processing technologies with a very high single processor performance.
As a result of this, the number of systems installed in Japan is only 82 systems or 16% of the world market, but the installed performance as measured by Rmax in the Linpack benchmark is 709 Gflop/s or 27% of the 2621 Gflop/s Rmax performance installed worldwide.
As stated in the description of the Linpack benchmark, this benchmark does not reflect the overall performance of a given computer system. It does, however, give the performance of a computer system dedicated to the task of solving a dense system of equations which has been adjusted in size to give the best performance with the optimal algorithm chosen by the vendor. Taken these precautions of a somewhat ideal benchmark, it is still interesting to note that the 82 Japanese TOP500 computers reach 709 Gflop/s Rmax of 929 Gflop/s peak, or 76% efficiency, while the other 418 computers reach 1912 Gflop/s Rmax of 3264 Gflop/s peak, or only 58% efficiency. One would therefore have to speculate that the overall performance of the computers installed in Japan relative to the other computers might be higher than indicated by Linpack Rmax figures. Another indication for this suspicion is the lower N 1/2 figures of the Japanese systems, which says that these systems reach a much better performance at smaller problem sizes.
The reason for these observations is that Japanese users continue to depend on vector processing for performance. This gives higher efficiency and higher sustained performance per system. The reliance on vector processing does not necessarily mean that Japanese users are extremely conservative. Of the 40 TOP500 systems installed in 1993/94, more than half were new architectures, different from tradtional vector multiprocessors. Six systems of the new Fujitsu VPP500 series are listed, an architecture which is new and deviates substantially from existing vector processing or parallel processing concepts.
The six VPP500 systems account for 310 Gflop/s Rmax of the 709 Gflop/s in Japan. The emergence of a new computer architecture is therefore another reason, why the Japanese computers are more powerful.