Supercomputing in Japan
Sverre Jarp and Wolfgang Bez
The Japanese supercomputer market and the impact of Japanese supercomputers on
the world market has changed substantially since the last TOP500 list was
published. There are three important developments. First, the Japanese users
and manufacturers rely more heavily on vector processing for high performance.
Second, the Japanese manufacturers are integrating vector processing into novel
scalable parallel computer architectures. Third, the Japanese government and
private companies are investing considerable money into these very high
performance computer systems. A major government sponsered supercomputer
procurement program was completed in 1993 with 11 major procurements and a
volume of more than a quarter billion US dollars. Different types of
supercomputers were procured, traditional shared memory vector multiprocessors,
RISC based MPP systems, and also the newly introduced vector parallel systems.
The move towards RISC based parallel computers is not as strong in Japan as in
other parts of the world. Both Japanese users and manufacturers continue to
rely on vector instructions and large SRAM memories for computing power. At the
same time MPP systems are investigated very thoroughly in manufacturer's and
user's research labs.
Japanese organizations like powerful computers. Japan has
now only about 16% (82) of the systems in the list as opposed to 106 last year.
But these systems contribute 27% (709 Gflop/s) to the worldwide installed Rmax
performance. The difference is mainly caused by a large number of older
generation systems (NEC SX-2 and SX-1, Hitachi S-820/60 smaller, Fujitsu VP-2200
or smaller, and Cray) falling off the list. In spite of the reduced number of
systems, Japan has now more
Rmax per inhabitant (5.7 Kflop/s) than the US (5.6 Kflop/s) and far more
than Europe (1.3 Kflop/s). On the first page of the list, the top 20 computers,
12 are from Japanese and 8 are American manufacturers. For the first time, two
Japanese manufacturers are deviating substantially from their American
competitors in their basic supercomputer design. Scalable vector parallel
computers are being introduced by both Fujitsu and NEC, with the design goal of
combining the high single processor performance of vector processors with the
high scalability of CMOS technology.
Tue Nov 14 15:13:46 PST 1995