Fujitsu Ltd is preparing to plunge into the massively parallel processing market with a machine that it claims delivers 300 GFLOPS peak using only a few hundred processors: on current plans, it will be the first major computer manufacturer to enter the market with its own machine rather than one bought in. The company has been familiarising itself with the technology with its own experimental highly parallel scalar machine called the AP1000, and it clearly does not trust the literature – it says that it was this machine that taught Fujitsu scientists that parallel supercomputing requires a radically different approach to programming. Having experimented with parallel scalar and parallel vector processing, and straight vector processing, the company concludes that while vector parallel processing is very efficient over a wide range of applications, conventional vector machines offer superior performance for many applications, so the company will continue to develop its VP and VPX machines. The company’s planned highly parallel machine will consist of an array of proprietary vector processors, each the size of a large briefcase and each delivering GFLOPS performance, to get TFLOPS performance out of hundreds rather than thousands of processors. The vector nodes, derived from the company’s existing vector processors, consist of a vector and a scalar unit in shared memory architecture that reallocates memory according to the work being done at each node. The company plans to launch its first machine before the end of the year: it will run Unix and be used with Fortran. The announcement was made at the ACM International Conference on Supercomputing in Washington yesterday. No word on any availability.
