The MX/5000-SP Series, announced in 1989, was a series of super minicomputers which were the first Japanese-made machines equipped with automatic vectorizing FORTRAN and a built-in vector processor. This series was comprised of three models: the MX/5600-SP, MX/5700-SP and the MX/5800-SP (the MX/5800-SP could be equipped with two built-in vector processors). Compared to channel-connected external array processors,the built-in vector processors eliminated the overhead of transferring vector data and the overhead due to the mixture of scalar and vector computation.
The built-in vector processor called Scientific Processor, or SP was comprised of hardware and firmware attached to the basic processing unit of the MX/5000 Series. In terms of hardware, it had features like floating point arithmetic elements to execute vector operations, and 32 floating point register files. The floating point arithmetic elements were comprised of adders, subtracters and multipliers in a 4-staged pipeline, and the pipeline pitch of each arithmetic element was 100ns. Adder, subtracters and multipliers could operate in parallel, and thus overall peak performance of the SP was 20MFLOPS. SP instructions followed the same instruction system as scalar instructions of the MX Series, so it was possible to mix such instructions with scalar instructions in the same series of instructions.
By simply compiling an conventional program as-is, automatic vectorizing FORTRAN could analyze parallel arithmetic potential of DO loops in programs, select a type suitable for vectorization, and automatically generate efficient vector instructions. Mitsubishi also provided an execution performance measurement tool for analyzing dynamic characteristics of programs called FORTRAN profiler and 240 SP libraries, which made and it possible to use the SP explicitly.
The MELCOM70 MX MS/5000II Series, which was announced in September 1990, could be equipped with processors such as a high-speed sorting processor, a high-speed image processor, a facsimile conversion processor and a character recognition processor in addition to the built-in vector processor (SP) with peak performance of 20MFLOPS. The high-speed sorting processor could speed up data processing tasks like data search or sort by 30 to 50 times compared to software processing.
