著者
三好 甫 吉岡 義朗 池田 正幸 高村 守幸 Miyoshi Hajime Yoshioka Yoshiro Ikeda Masayuki Takamura Moriyuki
出版者
航空宇宙技術研究所
雑誌
航空宇宙技術研究所特別資料 = Special Publication of National Aerospace Laboratory SP-16 (ISSN:0289260X)
巻号頁・発行日
vol.16, pp.99-106, 1991-12

In a few years, a computer which processes CFD programs over 100 times faster than the Fujitsu VP400 and has a main memory capacity of more than 32G bytes will be required for CFD technology to play an important role in aerospace research and development. A distributed main memory parallel processor is free from the memory throughput bottleneck which prevents the implementation of shared memory parallel processors with the necessary speed. In the light of regular characteristics of CFD codes, a distributed memory parallel processor is likely to deliver the above-mentioned processing speed. Its characteristics include a physically distributed main memory which logically provides programmers with global and local memory views, processing elements with high speed RISC scalar units and high speed vector units with large capacity vector registers, and a crossbar network which interconnects a large number of processing elements. Such a processor can be suitably called the "Numerical Wind Tunnel". This paper describes the basic main memory structure, system configuration, processing element, and interconnection network and communication mechanism of the Numerical Wind Tunnel.