Performance of Schur´s algorithm on an optically connected multiprocessor

This paper extends earlier concepts for autoregressive computation on an optical crossbar interconnected parallel processor. The dataflow architecture uses optical coupling to provide reconfigurable interconnections. Processing elements are reconfigured to perform different functions by downloading microcode. Schur´s algorithm is modeled for N = 5, in both rolled and unrolled implementations using a dataflow simulator. The unrolled version uses 100 processing elements to solve one set of inputs per clock after a latency of 44 clocks. Each solution processes only 11 input values to produce the 6 output values, thus minimizing I/O bandwidth.