Multiprocessor architecture.

摘要

A numeric processing subsystem, where at least three processors concurrently run separate processes. One of the three is a control processor (110), one is a data transfer processor (120), which controls the external interface, and one or several are numeric processors (130). Preferably this subsystem also includes at least a megabyte of cache memory (140), which is linked to the floating-point processor by a very high-bandwidth bus. The interface to the numeric processor(s) is so constrained that a variety of integrated circuits can be readily substituted. To compensate for the different pin-out and routing requirements of different chips, exchangeable subboards are preferably used to mount the arithmetic chips. To achieve this chip-independence without sacrificing speed, an extremely flexible and highly multiported register file is used at the interface from the cache memory (140) to the numeric processing subsystem. To achieve algorithm flexibility, parallel writes from data cache memory to the control storage of the numeric processing subsystem are preferably possible. At least one of the numeric processors is preferably a floating-point processors, and one or more others may be application-customized processors. The application-customized processor is especially well adapted for some particular class of operations, such as discrete Fourier transform operations, and the numeric processor provides acceptably high speed on the general range of numeric computations. The data transfer processor (120) controls several external interfaces, including a port, for configuring multiple subsystems together, which includes two high-bandwidth input ports and a high-bandwidth output port. This permits a wide variety of highly pipelined and/or parallel architectures to be readily implemented.