The LINPACK benchmark on the Fujitsu FAP 1000

The author describes an implementation of the LINPACK benchmark on the Fujitsu AP 1000. Design considerations include communication primitives, data distribution, use of blocking to reduce memory references, and effective use of the cache. The LINPACK benchmark results show that the AP 1000 is a good machine for numerical linear algebra, and that one can consistently achieve close to 80 percent of its theoretical peak performance on moderate to large problems. The main reason for this is the high ratio of communication speed to floating-point speed compared to machines such as the Intel Delta and nCUBE 2. The high-bandwidth hardware row/column broadcast capability of the T-net (xbrd, ybrd) and the low latency of the synchronous communication routines are significant