Abstract

The programmable optoelectronic multiprocessors (POEMs) machine has a highly parallel architecture based on wafer scale integration of optoelectronic processing elements (PEs) and reconfigurable free-space optical interconnects. The POEMs machine is realized with an integrated silicon/electrooptic modulator technology (e.g., Si/PLZT) to implement the PE arrays and 3-D holographic storage medium such as photorefractive crystals to support the programmable interconnects. The architecture of POEMs can support any variation of the parameters commonly used to classify parallel architectures: granularity (fine, coarse, or large grain), synchrony (SIMD or MIMD), and topology (local or global). Here we describe a fine-grain POEMs machine containing a very large number (100,000 or more) of simple 1-bit silicon PEs. An optoelectronic controller is used to broadcast optically an instruction stream to the PEs for SIMD processing. Unlike conventional parallel systems, there is no fixed Interconnection topology among the processors. Instead the programmable optical interconnects are determined by the optoelectronic controller. The programmer implements a topology that best matches his algorithm. Experimentally, we have designed a scalable prototype of such POEM machines with eight PEs. Each PE has the capability of performing logic, data movement, conditional execution, memory, and I/O operations. Among a wide variety of algorithms to which it was designed to apply parallelism, the fine-grain POEMs machine is especially effective for the rapid execution of symbolic information processing tasks and graph algorithms because of the programmability of optical interconnects and the large number of simple PEs. In particular, it is expected to offer extremely high performance in the rapid execution of semantic networks, production systems, management of large knowledge bases, parallel databases, transportation and communication optimization problems, computer-aided design, VLSI circuit simulation, and game playing. Some sample algorithms are presented and their performance compared with that of the conventional systems.

© 1988 Optical Society of America