Introduction

Optical processors hold tremendous potential for processing many information channels in parallel, each at very high bandwidth, and in small volumes with low power consumption (VanderLugt, 1992). Major classes of operations that optics can perform include integral transformations such as Fourier transforms and correlations, and matrix operations, e.g. vector–matrix multiplication (Lee, 1987). Many, but not all, of these are made possible by the remarkable property that a ‘Fourier transform lens’ generates, at the back focal plane, the two-dimensional (2-D) Fourier transform of phase and amplitude information input at the front focal plane.

Initial feasibility demonstrations of optical signal processors are usually performed in the laboratory using existing components, but these often do not come close to fully exploiting the potential of the optical processor, and often disregard practical implementation issues such as: how much phase distortion is permissible in a critical lens, or what will be the effect of vibration or temperature changes in the field, and is it feasible to have a given dynamic range and information capacity in the same processor? Therefore the aim of most subsequent optical hardware development is to answer these questions and overcome any deficiencies.

Much effort has gone into the development of the critical active optical devices, such as infrared laser diodes, spatial light modulators (SLMs) and photodetector arrays, and new device concepts and improvements on existing devices continue to occur (Lee & VanderLugt, 1989).

Introduction

This chapter is dedicated to the evaluation of optical interconnects between electronic processors in multiprocessor systems. Each processor is fully electronic except for the incorporation of a number of photodetectors and optical signal transmitters (e.g., laser diodes, light emitting diodes (LEDs) or optical modulators).

The motivation for this analysis stems from fundamental advantages of optics as well as those of electronics. While electrons are charged fermions, subject to strong mutual interactions and strong reactions to other charged particles, photons are neutral bosons, virtually unaffected by mutual interactions and Coulomb forces. Thus, unlike electrons, multiple beams of photons can cross paths without significant interference. This property allows holographic interconnects to achieve a 3-D (three-dimensional) connection density with only 2-D optical elements. Similarly, photons can propagate through transparent materials without appreciable attenuation or power dissipation. Thus, neglecting speed of light delays, the speed of an optical link is limited only by the switching speed and capacitance of the transmitters and detectors. (For a 10 cm connection length, and a 50° hologram deflection angle, the speed-of-light delay is 0.5 ns.) Hence, the speed and power requirements of an optical interconnect are independent of the connection length. Since electrical very large scale integration (VLSI) connections have a switching energy directly proportional to the line length and an RC delay that grows quadratically with line length, for long enough communication links, optical connections will dissipate less power and provide faster data rate communication.

Introduction

When a scientist or engineer is provided with a given processing problem and asked whether optical processing techniques can solve it, how does he answer the question and approach the problem? To address this, he first considers the basic operations possible. They are often not a direct match to a new problem. Hence new algorithms arise to allow new operations to be performed (often on a new optical architecture). The final system is thus the result of an interaction between optical components, operations, architectures, and algorithms. This chapter describes these issues for the case of image processing. By limiting attention to this general application area, specific examples can be provided and the role for optical processing in most levels of computer vision can be presented. (Other chapters address other specific optical processing applications.)

Section 1.2 presents some general and personal philosophical remarks. These provide guidelines to be used when faced with a given data processing problem and to determine if optical processing has a role in all or part of a viable solution. Section 1.3 describes optical feature extractors. These are the simplest optical systems. This provides the reader with an introduction and summary of some of the many different operations possible on optical systems and how they are of use in image processing. A major application for these simple systems, product inspection, is described to allow a comparison of these optical systems and electronic systems to be made.

Section 1.4 addresses the optical correlator architecture, since it is one of the most powerful and most researched architectures. The key issues in such a system used for pattern recognition are noted.

The design of optical hardware for signal and image processors and in computing systems can be expected to become an increasingly important activity as optics moves from the laboratory to actual usage. Over the past decades a variety of concepts have been developed to exploit the capabilities of optics for ultrahigh-bandwidth and massively-parallel processing such as for analog front-end operations, image-processing and neural-net applications, high-bandwidth interconnection of electronic systems such as computers and microwave subsystems, execution of logic and other nonlinear operations for routing of optical interconnects, as well as for general reduction of hardware size, weight and power consumption. Original concept investigations usually identify theoretical advantages of performing various tasks with optics, but one then needs to address whether the advantages can be maintained as one addresses various new issues involving usage in a real-world environment; e.g., can demonstrations of optical processing speed and parallelism, optical routing, and communication bandwidth be retained in full-scale application? Addressing these questions very often requires an optical approach that employs different algorithmic/architectural approaches than conventional alternatives; hence, new architectures are devised to utilize maximally the strengths of optics. Finally, the capabilities of optical technology may dictate specific optical hardware designs.

It is therefore difficult to predict whether a specific optical hardware configuration will still possess the originally identified advantages. Although full-scale demonstrations can provide a direct answer, careful design and analyses must be performed to ensure an effective demonstration. The design and analysis stage can involve fundamental issues that are quite removed from the original optical-processing concept.

Introduction

This chapter discusses the issues involved in the design of a free-space digital optical logic system. The first section is an introduction to this field with an overview of the justification of pursuing this exciting and challenging course. The second section is a discussion of the basic characteristics of the devices that must be used in a system of this type with specific examples of the behavior of three of the most popular devices. Section 6.3 is a discussion of the optical and mechanical constraints involved in the design process. The fourth section is a design example using logic gates made from symmetric self electro-optic effect devices (S-SEEDs). The final section summarizes the likely future directions.

Background of digital optics

Two methods exist for communicating high bandwidth information: light and electricity. It is clear that the optimum choice for long distances (>1 km) is light; hence the adoption of fiber for telephony. It is equally clear that over short distances (<1 mm), the optimum choice is electricity hence its use for gate-to-gate interconnection in a chip. It has been shown that as the distance between the transmitter and receiver increases, the energy required by an electrical connection increases more rapidly than if optics is used (Miller, 1989). It is difficult to justify the use of optics over short distances and conversely, electrical connection is not ideal except over the shortest distances. The ultimate optimum choice for chip-to-chip and board-to-board interconnection is unclear.

Introduction

Optical linear algebra processors (OLAPs) perform numerical calculations such as matrix–vector multiplication by equating a particular property of light, normally its intensity level, to a number. Through various effects to be discussed in this section, uncertainty is introduced either through random fluctuations or the addition of a bias. The fluctuating intensity can now be considered a random variable for which a mean and standard deviation can be determined. After corrections for bias, the mean value can be taken as the correct level that relates to the expected numerical value while the nonzero standard deviation represents processor noise.

In the design of analog OLAPs such as matrix–vector and matrix–matrix processors, the effect of noise from both device and system sources has a major impact on the choice of system architecture and components. The ability of a particular system design to meet system performance specifications such as signal-to-noise ratio (SNR), dynamic range, and accuracy is directly connected to the noise properties of the system and its components. As an example, the choice of the spatial light modulator (SLM) has a significant impact (Casasent & Ghosh, 1985; Perle & Casasent, 1990; Taylor & Casasent, 1986; Batsell, Jong, Walkup & Krile, 1989a, 1990; Jong, Walkup & Krile, 1986). This chapter examines the effects of noise on these system specifications and their implications for OLAP design.

We begin with a theoretical analysis of a simple matrix–vector multiplier and then generalize the results for N cascaded matrix–vector multipliers.