Complexity of combinational logic networks realized by fiber optic logic elements is analyzed. The fiber-optic elements consist of optical fibers, photo detectors, and light emitting diodes, and they act as NOR (or NAND) gates. We assume that both true and complemented variables are available as inputs, and that the networks are at most three levels. The measures of complexity include number of gates, fan-in, fan-out, connections, and array size when realized by programmable logic arrays. The complexity for arbitrary functions, symmetric functions, parity functions and adders are derived by using combinatorial mathematics, and for other arithmetic functions, randomly generated functions, control networeks for microprocessors are derived by computer simulation. These measures are useful for estimating the amount of hardware and power dissipation to realize various functions by fiber-optic logic networks.

The evolution of silicon submicron technologies will yield very powerful single chip U LSI processors (possibly processor arrays) and high performance advanced packaging technologies, providing significant opportunities to realize very compact, distributed computing systems. However, exploiting that opportunity will require development of very high performance communication networks, scaled to the much smaller size and more monolithic realization of such future distributed systems. Optical communication is presently being applied to larger scale versions of such networks which, if scalable to the smaller, more monolithic world of future system structures, may help overcome several physical limits of scaled electrical networks. We review general system-level limits of scaled optical networks, assuming cointegration of Si CMOS logic, GaAs-based optoelectronics and waveguides within a common monolithic technology. The system limits suggest that a number of performance limits remain. Resolving such limits will be critical in exploiting the considerable advantages of scaled, optical interconnections for such future, highly integrated systems.

This paper discusses basic component technologies for the development of digital optical computing systems. The establishment of integration technologies, including optoelectronic and photonic integrations, is important for the application of optoelectronic devices in computing systems, especially for use in optical interconnections. The development of optical bistable devices is a prerequisite for improving computing performance. Both of these technologies are a significant basis for optoelectronic computing systems, and can be built up by using III-V semiconductor materials. Recent advances in these fields will be presented.

The properties of a GaAs-based integrated multiquantum well device capable of optoelectronic amplification are described. The device consists of a heterostructure bipolar transistor acting as a controller and a multiquantum well Stark effect p-i-n modulator grown in single step epitaxy. The integrated device also has cascadable properties and fan-out and fan-in values of 8 and 6, respectively, are measured.

The use of new interconnection technologies to address the interconnect problem encountered in massively parallel computing are discussed. Emphasis is given to a new approach in which the computational units are stacked in the third dimension rather than spread out horizontally on a board. The average interconnect length decreases significantly, resulting in large reduction in system power consumption due to much lower parasitic impedances. Another advantage of shorter interconnects lies in reduced interconnect delays, alleviating problems associated with memory-access/logic-cycle-time discrepancies. This three-dimensional architecture uses wafer-scale integration for each of the planes, eliminating costly and space-consuming packaging at both the chip and board levels.

Optoelectronic component arrays for use in the optical interconnection of VLSI circuits are described. The design of Stark-effect modulators and of detectors for 'flip-chip' hybridization onto integrated circuits are discussed, and preliminary results are presented.

The Modulo 61 M/A has been designed using a custom multilayer board, semicustom IC drivers, a light plate, and alignment assembly. A light plate and an accompanying alignment assembly and procedure have been designed to form in excess of 500 optical interconnections at once to interconnect the M/A. A discrete 3 x 3 LUT will perform at 100 MHz; the Modulo 61 M/A will run at between 100 and 200 MHz.

The design, fabrication, and performance of an advanced distributed optoelectronic computing array architecture is presented. The system is used to compute a series of morphological image processing operations constituting a skeleton algorithm. The entire computation is performed in pure combinatorial logic and is effected in a single pass through the system.

An optical computing model is described that makes use of arrays of optical logic gates interconnected in free space with regular connection patterns. An advantage of this approach is that interconnection complexity is moved to free space which is virtually free of faults. A cost of this approach is that the regular interconnection patterns restrict connectivity to the extent that a number of logic gates are inaccessible. Methods are described for moving the operational, inaccessible logic gates into the positions of faulty logic gates. The result is that the effective yields of optical logic arrays are increased and device processing tolerances are relaxed.

Corner cube retroreflectors with dihedral angles slightly different from right angles are used for optical chip-to-chip interconnections. Their spot patterns are used as interconnection patterns. Alignment is a critical issue for free space interconnection systems. This approach reduces the requirements for alignment accuracy by about hundred times. The corner cubes can also be used for space invariant interconnects that have potential for large fan-out numbers and high spatial densities. Experimental results are shown.

Modelocked semiconductor lasers are a source of subpicosecond pulses at high repetition rates with very low phase noise. This makes them ideal sources for clock/strobe signals for OEIC's and optical computing, for high data rate input devices and for use in optical digital-to-analog converters. These applications will be described in this paper, together with an outline on the state of the art in modelocked semiconductor lasers.

This paper reviews the optical hardware required for photonic switching and optical computing systems that are based on free-space digital optics. It includes a comparison of the devices, interconnects, and systems that have been proposed for both types of systems.

The arrival of optical technology changes the interconnection structure of highly-parallel multiprocessor systems. We classify and investigate the topologies of three-dimensional multistage interconnection networks. We also discuss the routing algorithms in the three-dimensional environment. Certain network topologies and controls are better utilizing the optical features. These networks are given special attention in our study. Key Words: Optical Switch, I)ilation, Spatial Invariance, 3-D Interconnections, Multistage Interconnection Network, Multiprocessors, Network Topology.

The application of photorefractive crystals (PRCs) to optical interconnection and switching networks is discussed. Examining the advantages and disadvantages of PRCs compared to existing optical interconnection approaches, it is shown that their major applications are reconfigurable interconnection networks. Systems using volume hologram superposition to provide fast reconfiguration (compared to the photorefractive response time) are of particular interest. Three such systems currently being investigated are discussed.

In this paper the characteristics and methods of forming substrate-mode holographic optical elements are described. These components use total internal reflection to translate an optical beam, and holographic elements to couple light in and out of the guiding substrate. These gratings also have several useful polarization properties which are then discussed and demonstrated. The paper concludes with a description of the application of polarization-sensitive substrate-mode holographic elements to an optical bus system.

Advances in high performance computers and signal processing systems have led to parallel system architectures. The main limitation in achieving the performance expected of these parallel systems has been the realization of an efficient means to interconnect many processors into a effective parallel system. Electronic interconnections have proved cumbersome, costly and ineffective. The Optical Fiber Crossbar Switch (OFCS) is a compact low power, multi-gigahertz bandwidth multi-channel switch which can be used in large scale computer and telecommunication applications. The switch operates in the optical domain using GaAs semiconductor lasers to transmit wideband multiple channel optical data over fiber optic cables. Recently, a 32 X 32 crossbar switching system was completed and demonstrated. Error free performance was obtained at a data bandwidth of 410 MBPS, using a silicon switch IC. The switch can be completely reconfigured in less than 50 nanoseconds under computer control. The fully populated OFCS has the capability to handle 12.8 gigabits per second (GBPS) of data while switching this data over 32 channels without the loss of a single bit during switching. GaAs IC technology has now progressed to the point that 16 X 16 GaAs based crossbar switch Ics are available which have increased the data bandwidth capability to 2.4 GBPS. The present optical interfaces are integrated GaAs transmitter drivers, GaAs lasers, and integrated GaAs optical receivers with data bandwidths exceeding 2.4 GBPS. A system using all Ill-V switching and optoelectronic components is presently under development for both NASA and DoD programs. The overall system is designed to operate at 1.3 GBPS. It is expected that these systems will find wide application in high capacity computing systems based on parallel microprocessor architecture which require high data bandwidth communication between processors. The OFCS will also have application in commercial optical telecommunication systems where high bandwidth communication has already exceeded 2 GBPS. There are also critical requirements for a secure fiber optic switching system in military Command Control Communication (C3) situations.

A family of fixed and reconfigurable optical interconnection networks offering the potential of Terabit throughput are described. The designs are based on a bus architecture that combines space, wavelength and other multiplexing methods and have the following attractive features:- graceful growth; wide-sense non-blocking operation; low space switch crosspoint count; and simple control algorithm.

The National Science Foundation has granted funding to the Centerfor Optoelectronic Computing Systems located in Boulder, Colorado. A new initiative has been undertaken there in collaboration with the Center for Telecommunications Research at Columbia University. This project is to design and build a multi-GHz optoelectronic data transport network using self-routing packets in a multi-hop network. The single electronic word packet payloads are highly compressed using optical techniques, and remain optical from source to target while traversing the switching nodes. Optical packet switching is performed with custom LiNbO3 directional coupler switches. The routing is done with a lean, self-routing hot potato protocol in order to avoid the need for data storage at the switching nodes and to provide a fixed node latency equivalent to a few meters of fiber. Sustainable throughput both in to and out of Ihe electronic host at each node should exceed 1 Gb/s, with bursts close to the 10-100 Gb/s peak link bandwidth. Some technical details of the optical compression and decompression schemes, the hot potato switching protocol, and the wrap-around shuffleexchange interconnection network will be given. The project timetable anticipates a lower speed, proof-of-principle four node network in three years, and a higher speed, larger, engineering demonstration in five years. The project received NSF funding in September 1989.

The optical computer and its development are crucial to performance enhancement of CI networks. The development of components, techniques, and devices necessary to the realization of an optical computer has made significant strides. What are the remaining areas of concern in the development of a practical optical computer for this application? This paper will take a close look at the steps to be taken in this quest for such a practical optical computer.

High speed photodetectors are required for use in high speed optical interconnects. Recent results on high speed photodetectors are reviewed, and devices with important implications for future high speed photodetector operation are considered.

The routing ability of waveguide structures patterned by impurity induced layer disordering of AlGaAs superlattice and graded barrier heterostructures is discussed. Also considered are issues associated with the integration of multiple optical devices into the waveguides.

Electroabsorption modulation (EAM) in strained In(x)Ga(1-x)As/GaAs multiple-quantum-well structures grown by MBE is investigated for spatial-light-modulator applications. Since EAM near the exciton absorption peak always has significant insertion loss, the enhancement of modulation depth is discussed primarily in terms of the dynamic range, optimizing the change in transmission (Delta-T), instead of the contrast (on/off) ratio. The material properties of the InGaAs/GaAs strained-layer heterojunction make possible the growth of a large number of quantum wells, above the pseudomorphic thickness limit. The use of reflection instead of transmission modulators does not produce an increase in the maximum Delta-T but allows a reduction in the driving voltage.

The use of quantum well material resulted in speed and efficiency advantages in devices such as injection lasers and opto-electronic absorption modulators. Opto-electronic circuits need also waveguides, switches and detectors. The challenge lies further in finding suitable materials and process technologies for their integration. Because of the electric field dependence of quantum well transitions one finds an extremely strong electric field dependence of the refractive index in the vicinity of the wavelength which corresponds to the quantum well transition. This phenomenon can be exploited to make electro-refractive wave guide switches and modulators of higher speed and efficiency, operating at voltages which are comparable to those in electronic ICs.

Glassy nonlinear optical polymers can be processed into channel waveguides. When poled, the channels become electrooptic and can switch and modulate light. Using lithographic and machining techniques familiar to the chip industry, it should be possible to integrate large numbers of electrooptic switches into a board-level package or module, and thus provide the additional benefits of active switching and reconfiguration to passive hybrid optical interconnect modules. Some of the properties of the materials, some process methods, and potential applications in optical interconnection are described.

An all-optical processing loop circuit, pumped entirely by semiconductor diode lasers, has been constructed and operated. Functional features include optically programmable logic, thresholding, and synchronization; these are achieved using three bistable interference filter devices. The circuit is presently single-channel, however 15 x 15 capability of the devices has been demonstrated using Dammann holograms and array-to-array coupling of a pair of bistable plates; potential parallelism is in excess of lO. Circuit simulations and tolerancing are also described.

A new scheme of digital optical switching, optical analog-to-digital conversion, and wavelength bistability is presented which can be applied to digital optical, memory operations, utilizing the mode-hopping phenomenon of Fabry-Perot cavity-type injection laser diode. The measured speed of switching was within the nsec region, which was limited by the driving capabilities and the light detection response, not by the phenomenon itself.

Recent results in the modeling and optimization of semiconductor etalon-based devices for optical switching and signal-processing applications are reviewed. Critical device performance criteria including contrast, cascadability, switching speed and energy are evaluated for both one- and two-wavelength device concepts.

A novel design for a dynamically programmable fiber optic based Johnson counter with built-in code conversion is presented. The design essentially exploits low attenuation, precise signal propagation delay, and high fan-in/fan-out factors in optical fibers. The approach is fully compatible with existing high speed digital integrated circuits. A 116 MHz version has been built that demonstrates the potential and the simplicity of fiber-optics in sequential logic processing.

The modified signed-digit (MSD) number system, because of its inherent weak interdigit dependance, has been suggested as a useful means for a fast and parallel digital arithmetic. To maintain a fast processing speed, a single-stage holographic optical content-addressable memory (CAM) based MSD algorithm was suggested. In this paper, a novel non-holographic opto-electronic CAM based fast MSD addition processing architecture is proposed. The proposed concept has been verified with our first-order proof-of-principle experiments. A figure of merit comparison of this and other existing approaches is also presented. Based on this key opto-electronic CAM element, implementation of more sophisticated I'VISD arithmetic, such as optical MSD subtraction and multiplication operations, are proposed.

Pred.kin gates matrix using LCLV has been realized on the base of bichromatic logic. With this Predkin biohromatio matrix the following results have been obtained: logical and arithmetic operations, programmable oommutat ion ne t and binary image processing. The work is a step to a polychromatic computer.

An architecture for an optically implemented adaptive array processor which operates in quadratic residue number systems (QRNS) is presented. The unit provides four adaptive degrees of freedom but can be expanded, through additional modular processing elements, to 12 or more degrees. Separate subprocessors construct a sample covariance matrix and solve for the complex weights via Gauss elimination and back substitution. A third subprocessor performs scaling and integer reconstruction operations through mixed radix conversion (MRC). The system uses optical clocking to operate at an internal rate of 200 MHz, and it will produce complete weight solutions approximately once per microsecond with a total latency of about 2.3 psec.

This paper describes the system considerations in implementing a bit serial optical computer using lithium niobate directional couplers and optical fiber. I discuss the three main problems in constructing such a computer: computation of fiber lengths, estimation of and compensation for power losses, and details of the physical construction of the machine.

Because of the well-known advantages of fiber transmission, a technologically important and increasing amount of data is transmitted optically, often in time-slot format. In-line optical time-slot interchangers would obviously be of great use in such systems. In a time-multiplexed opticai computer environment, a time-slot interchanger corresponds to a multiprocessor interconnection network in a spatially parallel multiprocessor system. For fundamental physics reasons, 2x2 optical exchange elements are among the easiest optical logic elements to construct. Until now, for a frame of N time-slots, designs have used a planar array of switches like their space analogs and required a few times N of these switches to implement. By taking a new approach, a powerful structure called a serial array architecture has been developed which has demonstrated very high potential for substantial hardware reduction. A time-domain version of the spatial shuffle-exchange network has been implemented in this architecture which has resulted in the time analog of the NxN space switch to be collapsed to O(log2N) switches. The perfect shuffle permutation is the more interesting part of the stage and consumes most of the switches for the interchanger. By recursively shrinking the size of the shuffle train, like in a Benes network, further reduction in switch count has been achieved. The resulting architecture yields such savings in optical switch and related hardware that in-line time-slot interchangers, even in LiNbO, may now be practical.

Optically switchable directional couplers can function as logically complete building blocks for constructing all-optical computational engines. Logically, such devices operate as all-optical five-terminal gates, where two output signals are logical functions of two guided input signals and an input control signal. Such an optically controlled exchange element is a promising functional unit for constructing general-purpose digital optical logic circuits. We have implemented such all-optical five-terminal gates using Ti:LiNbO eleciro-optic directional couplers, with singlemode optical fiber for all gate interconnections. Effectively DC-coupled control circuitry converts short, low-power (<5ns, -25 dBm) 1 .3 mm optical pulses into electrical pulses capable of switching the transfer state of low-voltage lumped electrode directional couplers. These optical logic gates, while containing significant electronics in the control arm, are logically identical to more fundamental optically controlled exchange elements, and utilize mature technology developed for the optical fiber telecommunications industry, simplifying construction of robust optical logic hardware. In addition, the use of significant electronics in the control terminal input offers considerable advantages over more fundamental optical switching effects, i.e. greater noise immunity, wider dynamic range, adjustable detection threshold, simple phase compensation, and conirol over the quiescent switching state. We have demonstrated the use of such Ti:LiNbO3 five-terminal optical logical gates in the construction of various simple circuits, such as oscillators and divide-by-N circuits. Such circuits demonstrate many of the issues arising in the construction of all-optical digital computing systems, and are fundamental subsections of an all-optical bit serial computer design, which we are currently consiructing.

The paper reveals a new operating principle for a planar photonic transistor in which no electronic components or nonlinear optical modules are used for digital switching. The author investigates the design constraints associated with the switching function, the contribution of the "noise" factors as well as the integration of the switch in a logic processor.

Characteristics of digital synchronous delay-line memories that use pulse stretching to compensate for phase variations are presented. When optical fiber is used as the delay medium, the choice of carrier wavelength determines which mechanism (thermal variation or dispersion) limits the maximum memory size. If implemented at the dispersion-minimum wavelength of 1310 nm using a laser with lmewidth 1.55 nm and 50% duty cycle, a memory could store up to 22 million bits before dispersion dominates, but such a system would require thermal stabilization to within 0.002C. A digital fiberoptic delay line memory will be built for a bit-serial optical computer, where each switching element is a lithium niobate directional coupler having an electro-optic control terminal. Non-idealities in this type of switch, such as attenuation, crosstalk, and polarization losses, will have negligible effects on the memory. Intermittent regeneration errors at the electrooptic boundary will also be minor. For a bit modulation frequency of 100 MHz, a single-line 2000-bit memory can be reliably implemented without thermal compensation for a lab variation of A memory system incorporating L delay lines would reduce both the average access time and the thermal sensitivity of the system by a factor of L. The number of additional switches required by such a memory is roughly 5L in a system accounting for regeneration needs.

The theory of classical artificial neural networks has been used to solve pattern recognition problems in image processing that is different from traditional pattern recognition approaches. In standard neural network theory, the first step in performing a neural network calculation involves the linear operation of multiplying neural values by their synaptic strengths and adding the results. Thresholding usually follows the linear operation in order to provide for non-linearity of the network. This paper presents the fundamental theory for a morphological neural network which, instead of multiplication and summation, uses the non-linear operation of addition and maximum. Several basic applications which are distinctly different from pattern recognition techniques are given, including a net which performs a sieving algorithm.

We demonstrate both nematic and ferroelectric liquid crystal SLM's in two separate optical connectionist machines on which two layer neural network algorithms are executed. An analysis of their system performance with respect to non-ideal operation of the component parts is presented; the analysis being carried out with the aid of a computer simulation. A direct comparison between the two machines is then made demonstrating the improved performance of the compact ferroelectric liquid crystal based machine.

Neural network researchers have recently addressed the problem of speech recognition. Most networks, however, offer solutions to the problem through complicated structures or complicated learning schemes. It is our goal to present extremely simple and fast learning algorithms that provide large discrimination between words. This large discrimination will allow this system to be designed into a simple optical computer architecture with minimum learning and yet maximum computation capacity.

The relationship of a neural network to an optical correlator architecture is discussed and a design concept is formed. This concept design promotes a modular system which can be cascaded to form successive layers of a neural network. The design approach uses the concept of solid optics to mitigate optical alignment problems in field enviroments. Current results on performance is discussed.

A hybrid optical/digital neural net is described. Initial tests on the optical components are provided with the first simulated neural net results addressing various optical system error sources included. Attention is given to the accuracy required for each optical component, the dominant error source and the cumulative effect of multiple optical system error sources.

Gabor-type motion sensors can be used to extract the spatio-teniporal frequencies of moving image patches. We propose a energy minimization scheme for the computation of optical flow from spatio-temporal frequencies. This computational scheme can be implemented by an optical neural system to perform rea1time visual motion analysis.

We demonstrate that novel electron trapping (ET) materials are capable of performing optical formation of an interconnection matrix based on the Hopfield prescibed learning rule. An association has been performed using the optically formed matrix.

The convergence of a neural network model based on optical resonator designs is examined for Boolean logic operations. Computer simulations are performed to investigate convergence performance and to assess possible optical implementations. The model is a simple and general mathematical formulation obtained using standard methods in which plane wave amplitudes and phases are specified at discrete times separated by the resonator period. The model is trained and tested as an associative memory neural network using an input state vector and a hologram matrix that evolves in time according to a set of coupled nonlinear difference equations. In general, these equations represent a high-order threshold logic, and the hologram matrix is a function of the outer product matrix of the evolving complex-element state vector. Model parameters are explored to provide insight on convergence mechanisms, robustness to input perturbations, and optimization of convergence times for both training and testing. The model is of interest for optical resonator designs that incorporate (1) dynamic holograms for massively parallel interconnection and storage functions and (2) nonlinear components such as phase conjugate mirrors (with thresholding and gain) for decision operations.2 These components are often incorporated into resonator loops to provide feedback and adaptation interactions. The neural

Symbolic substitution operations can be realized optically on a correlator. This is a very attractive and efficient architecture for symbolic substitution. It allows parallel multichannel realization with a fixed set of filters (on film or easily realized on low space bandwidth product spatial light modulators) using space and frequency-multiplexing or sequential filters. All basic logic, numeric and morphological image processing functions can be achieved by symbolic substitution. Moreover, all operations are possible on one multifunctional optical processor. Morphological operations are felt to be essential for ATR and pattern recognition preprocessing in clutter. They greatly improve the role for optics by allowing the same optical architecture to be used for low, medium and high level vision.

Several new higher-order spatial symbol recognition methods for optical symbolic substitution-based calculations are presented. In case of logic processing, higher-order symbolic substitution (SS) rules can be applied to implement multi-variable logic functions. In binary arithmetical calculations requiring carry propagation, the simultaneous processing of a number of bits increases computational speed. Finally, using higher-order SS rules, image processing can be perform using larger windows. Both, multiplicative and additive techniques for a spatial symbol recognition are discussed. Four different optical architectures, a multi-reflecting technique using an optical cavity, a correlation, a phase conjugation and a content-addressable memory (CAM) techniques, are suggested. Optical either dual-rail (DR) or triple-rail (TR) spatial encoding is employed. Some preliminary experimental results are also included.

By combining two powerful optical computing techniques, namely, optical symbolic substitution (OSS) and polarization-encoded optical shadow-casting (POSC), morphological or shape transformation operations are demonstrated. Accordingly, erosion, dilation, opening, and closing operations are realized using both OSS and POSC schemes. These morphological operations are used for noise removal in binary images.

The architectural relationship between two powerful optical computing techniques namely, symbolic substitution and optical shadow-casting is discussed. A common basis for both the techniques are devel oped and their roots are traced back to fundumental principles of logic design. It has been shown that both shadow-casting and symbolic substitution based optical computing operations can be expressed as a logical sum of product expression.

Over half of backplane layers in computers are often required for clock distribution because of isolation and termination problems associated with high frequency signals on electrical wires. Glass and plexiglass sheets have been demonstrated for optical distribution of these signals. We present advantages and disadvantages of laser glasses allowing gain to compensate for output losses.

The technique of laser direct-writing was used to fabricate low-loss, rib-like waveguides on GaAs/A1GaAs material. Since this is a inaskless fabrication process, laser-etched structures such as bends, tapers, and Y-branches are readily demonstrated. The ability to smoothly control the etch depth, and thus the effective index of refraction across the wafer, was used to make a compact directional coupler. Finally, the reduction of losses in waveguide tapers and bends by grading the effective index will be discussed.

Optical shadow-casting based edge detection is hardware implemented using the spatial light rebroadcasting (SLR) materials. The digitized input image is encoded in accordance to the polarization-encoded optical shadow-casting design algorithm. An edge-detection operation of the coded image is performed using the SLR material both asspatial light modulator and logical device.

We describe an optoelectronic neural network implementation with electronic neurons and synapses and optical storage of weights and present experimental results.

A compact size optical neural network using high resolution liquid crystal televisions (LCTVs) has been constructed. System design consideration and an experimental demonstration of the LCTV neural network are provided.

We ask and give only very preliminary answers to two questions which must arise when we consider quantum mechanical computers with significant quantunt indeterminacy. First, how does this impact our belief in Church's thesis? Second, how does this impact our belief in freedom of thought?