Standard microprocessors versus custom processing elements for massively parallel architectures

Abstract

Choosing a standard microprocessor or a custom processing element as CPU of a massively parallel architecture has major impact on the hardware and software costs. The standard microprocessors have a high performance/cost ratio, but the associated cache hierarchy leads to completely reconsider the programming of the applications that have been developed for vector supercomputers. Low hardware cost is associated to a high software development cost. New approaches, as multithreaded architectures, have low software costs, but expensive hardware costs.