The GF11 parallel computer from Yorktown [5] is a SIMD architecture with 576 processors and an unusual interconnect. The processors are connected through a three-stage Benes network, capable of supporting any permutation of the processors, and reconfigurable among 1024 distinct configurations in a single cycle. Its main application is the numerical evaluation of some of the predictions of quantum chromodynamics.
The architecture has two key similarities to a reprogrammable scheduled router: it supports a notion of multiple phases of communication at the router level, and it allows for rapid transitions among phases. A COP backend for the GF11 would capitalize on the fast context-switch time as well as the large number of phases that could be held in hardware.
Each stage of the network consists of 24 24-input crossbar switches, with the middle stage connected to the outer stages with a perfect-shuffle fixed interconnection. By modifying the state of the crossbars any permutation can be supported, as well as other interesting communication operators such as broadcast. Configurations can be loaded into each crossbar for use on demand; a 4-neighbor torus, for example, would require four configurations, one for each direction. Configurations typically take some time to compute, and are loaded as part of the program startup; once loaded, the frontend processor can switch configurations on every cycle.
Their architecture, reported in 1985, was able to outperform a Cray 1 by a factor of approximately 100 on their target application. They suggest that similar performance might be obtainable for a variety of scientific and engineering applications other than chromodynamics.