Parallel algorithms and architectures based on pipelined optical buses.

ABSTRACT

Optical signals have some unique properties, such as unidirectional propagation and precisely predictable path delays in waveguides, which are not shared with their electronic counterparts. By taking advantage of these unique properties, we can use optical interconnections to achieve speed improvements in single-instruction stream, multiple-data streams (SIMD) computations. We first show how optical buses can be utilized advantageously in SIMD architectures to obtain fast solutions to several computational problems, including integer addition, counting and logical XOR, sorting, and fast Fourier transforms. We then present a new implementation of the optical buses to meet the unique requirements in highperformance optical-electronic computing systems. Such an implementation allows the transmission of messages at speeds ideal for optics and, in the meantime, the processing of data at speeds ideal for electronics, dealing successfully with the speed limitation by electronics in optical-electronic computers. The primary effects of this bimodal optical bus are twofold reduction of fiber lengths and reduction of system latency. Reduced latency is a unique advantage to an optical bimodal bus. Together, these observations make optical-bus-based architectures appear to be a promising approach to SIMD processing.