Fine grain parallel processors and real-time applications: MIMD controller/SIMD array

Important issues related to the applications of massively parallel architectures for real-time cognition are discussed. One of these is real-time control of the array processor. Systems accommodating massive data parallelism are faced with the critical issue of how to incorporate adaptive programming decisions without losing machine efficiency. This involves reexamination of functions to be performed by the software versus those performed by the hardware. The discussion is based on a multiple-instruction multiple-data (MIMD) controller that has been constructed for a single-instruction multiple-data (SIMD) array. The efficiency of an SIMD architecture as a computational model is discussed with reference to several important functions in vision applications that have been benchmarked on the system. Applications that involve real-time control are described.<
>