Massively Parallel Processing with Optical Interconnections: What Can Be, Should Be and Must Not Be Done By Optics

Abstract

What is wrong about Optical Computing is the implied search for “general purpose computing”. We think that such an attempt has little chance to result in a practical system for, at least, the next ten years. The main reason is the economical justification. What such an “optical computing” system may offer has to be compared with the value of the application and the alternatives (electronics). On the other hand, communication in general is an area where optics has proved to be a real blessing. Long distance communication is most economically done today using optical fibers. We think that another realistic search for good optical applications should now be done for shorter distances. A possible good direction may be the communication needs of Massively Parallel Processing (MPP) systems. In such a system, large number (10’s of thousands) of Processing Elements (PEs) are to be interconnected. A PE can be seen as made of a high-end single chip CPU available today, with memory and communication circuits. We do not view the other possible meaning of MPP, namely processing and interconnections at the single gate or device level, as practical to consider. This paper describes the views of the author from the computer architecture’s standpoint, with the hope to serve as a pointer to the “Optical Computing” community. Although much has been done in the area of optical communication technology for the past 10-20 years, and many optical network experimental systems have been proposed, it seems that optics has not yet found its expected place as the interconnection technology of choice for MPP systems. In this paper we try to suggest some possible reasons preventing the common use of optical interconnections in MPP systems, in a hope to focus attention on what really needs to be done to advance the field. We would suggest focusing on searching for a processing-less solution rather than trying to mimic the existing thinking of electronic networks. We outline several key principles essential to follow to reach realistic and economical solutions of optical interconnections for MPP systems. An example of using such principles for an MPP, free-space network is presented in [1].