Hardware and Software Considerations for Improving the Throughput of Scientific Computation Computers

In this paper, hardware and software techniques are presented for improving the Throughput (defined as computations per dollar) of computing systems which are oriented towards high-precision floating point computations. The various improvements are referenced to a baseline of the PDP 11/20, the NOVA 1200, and the TI 960A, all 16 bit minicomputers. The most beneficial hardware improvement is the inclusion of a Floating Point Processor, which yields up to 200X Throughput increase over a software floating point package. The inclusion of a cache high speed local memory and the availability of Polish Notation format instructions are shown to provide less than a 5X increase each. The use of 48 bit data paths, numerous registers devoted to various processor functions, instruction look ahead, a system I/O controller which frees the processor from I/O work, and partitioned main memory, result in a combined Throughput increase of 5.9X