Optoelectronic system for 2-D nested crossbar architectures

Abstract

The nested crossbar class of optoelectronic matrix vector architectures allows a varying degree of detectors, transmitters, optical fanout, area and execution time.1 The 2-D dedicated link nested crossbar is especially attractive since it performs an N×N matrix vector multiplication in O(log N) time while requiring only O(N3/2) transmitters (with O(1) fanout) and O(N3/2) detectors, thereby reducing demands on transmitter technology. The 2-D nested crossbar decomposes the matrix vector multiplication into the preliminary calculation on N3/2 partial sums, with these partial sums then being added to yield the output vector. The optoelectronic system consists of three planes of processing elements (PEs) with N PEs per plane. The first plane consists of vector input PEs each having N1/2 transmitters. The second plane consists of N PEs, each having N1/2 detectors, N1/2 transmitters, and an N1/2 × N1/2 electronic crossbar. Each second plane PE calculates N1/2 partial sums to calculate one output vector element.