The benefits of optical interconnections over distances of hundreds of metres or more are well established. However, the increasing complexity of computers and other electronic equipment places severn demands on interconnection capabilities over much shorter distances where optics can also offer potential advantages. A comparison of optical and electrical interconnects at the board and backplane levels of modem mainframe computers has been performed using the interconnect criteria appropriate at these levels. For example, within the CPU (central processor unit) the most important criterion is latency. In electrical interconnections this arises from a combination of the propagation time together with signal distortion caused by the frequency dependent attenuation characteristic of the transmission line and reflections from terminations, vias and other discontinuities. This gives latencies which increase supralinearly with increasing line length. The optical signal, however, propagates faster and has no frequency dependent delay, but a length independent delay is incurred in the electrical/optical interfaces, setting a lower limit on the distance over which optics offers an advantage. This trade-off is examined. Power consumption and interconnect density are also compared. Optical backplane geometries allowing unrestricted bi-directional connections (ie. any point to any other point) are less well defined, but a comparison has been made of one particular geometry based on a wideband bus consisting of up to eight nodes, each operating at 32 Gb/s.

Two important issues will greatly influence the success of mapping optical interconnections into future waferlevel distributed computing systems: (1), the scalability of active optical devices with cointegration along side ULSI components, and (2), the scalability of optical networks and components to the wafer level. If these criteria can be met, planar integrated and free-space optics can potentially provide a very high performance communication network within the multi-wafer environment. With the predominantly planar geometry and processing of waferlevel circuits, process compatible integrated planar optical interconnections are especially attractive for providing network passive connectivity. As with their electrical counterparts, spatial, as well as time division multiplexing of optical interconnections is desirable, given that layout and area constraints are not too severe. Therefore here, emphasis is shifted away from the individual behavior of traditional long distance lightwave single mode waveguides towards the collective system behaviour (i.e. density, coupling, layout, etc.) of large dense arrays of multimode optical waveguides. In this paper, initial experimental optical coupling results are presented for arrays of multimode polysilyne polymer waveguides, both for straight configurations and for arrays with radial right angle bend layouts.

The design and the modelling results of an 8 channel parallel optical chip to chip interconnection consisting of a laser diode (LD) array, a single-mode waveguide (WG) array, and a photodiode (PD) array with 8 channels each are presented. The separation of the channels is 125 im, so the overall width of the 8 channel line is only I mm. The electronic and the optoelectronic components will be mounted on a silicon substrate wafer and the waveguides on a second silicon wafer which will be fixed upside down on the substrate. The LDs are envisaged to be AlGaAs singlequantum well types though the first implementation will be realized with conventional A1GaAs MCRW semiconductor lasers with a wavelength of 0.85 rim. The PDs are fabricated in standard silicon technology, the silica WGs with the flame hydrolysis technique and reactive ion etching. The trade off between large fabrication tolerances and the desired high coupling efficiencies is discussed. Mounting techniques for the LD- and PD-arrays are presented. A comparison between this optical interconnection and an equivalent electrical one is given.

The Deift Parallel Process, based on a multi-computer architecture and constructed with a Multi-Broadcast Interconnection topology, is a 16 processing no&s Multiple Instruction stream, Multiple Data stream computer developed at the Deift University of Technology in 1977 to evaluate software algorithms, architectures and interconnection topologies experimentally. The Processing Elements in the nodes, containing Weitek Floating Point Multiplier/ALU combinations, are going to be interconnected in a full parallel Electro-Optic Communication System. The Routing Elements, containing Transputers, are linked by an optical reconfigurable network, which enables a simultaneous communication between all the REs.

Providing the required interconnections between the processors of a parallel computer is a difficult problem: latency, switching control, cost, and crosstalk effects have to be taken into account. It is widely believed that the design might be simplified if optical technology is used. However, even optical interconnections cannot cater for an unlimited number of processors. For systems with up to about 1O - 106 processors, it seems that the arrangement of the processors in space so that each can "see" the others is the limiting factor; A large part of the volume involved must stay empty for the information-bearing light beams. above that, diffraction effects and power requirements place a limit on the number of processors.

We present an optical, free-space, crossbar-like N x N interconnection network based on N routing switches of the form 1 to N. The network is used to connect Processing Elements (PEs) in a large-scale, parallel computer, utilizing a 3-D space layout. Each PE controls and sets its own switch, using a distributed algorithm. The switch deflects a monochromatic, coherent light beam towards one of N possible spots. This beam, carrying information from the sending PE, is thereafter passively deflected by a system of multifacet holograms, until it reaches the target PE. Unlike other schemes, this layout uniformly fills the 3-D space with PEs. Thus, inter-PE communication can be minimized when algorithms exploit locality of reference.

We present here the state of development of our MILORD machine. MILORD is a multiprocessor architecture organized around a 32 x 32 optical crossbar network. This network interconnects 16 INMOS T4 14 microprocessors with four bidirectional links each. Once the system integration was made, we have noted the existence of errors on the optical links caused by the presence of modal noise at the output of the optical fibers. A theroretical analysis of modal noise effect is presented and solutions to avoid it in optical networks are proposed. This study is still in progress at the Optics Department. We examine the possibilities to improve the three main parameters of such networks : the number of links, the reconfiguration time, the data rate. We give a theoretical and quite general analysis about the bit error rate performance as a function of the mean power available at each output of such matrixvector inner product optical networks.

In this paper lange dimension photonic switching networks and their performances are analyzed. The characteristics of completeness, planarity, reconfiguration times, non-blocking, and the fault-tolerance property are taken into account. The shuffle type interconnection network results the better choice. Moreover, the control algorithm is considered, and the results of the simulations performed by a digital computer are given.

Micromachining in silicon is a well established technique for manufacturing sensors and actuators. Another field is fiber alignment components such as V-grooves and reflecting mirrors. This work is aimed towards a batch fabricated opto coupler with the use of both V-grooves and perpendicular inflecting mirrors on the same silicon wafer, with a maximum utilization of the single crystalline lattice structure. Since both the (1 1 1) and (100)-etchstop planes are given by the crystal lattice structure, it is possible to obtain a well defmed 45 degree angle between the [1 10] and [100]-directions. The main function of the (2x2) coupler is that a perpendicular wall has a height that covers half the fiber core. The result is that half of the incoming light beam passes over the wall and the other half is reflected. There will obviously be some geometrical limitations for the distance between the fibers, which introduces optical loss that must be considered. The measured excess loss is 4dB above the theoretical limit.

GaAs layers have been grown on silicon subsirates by Molecular Beam Epitaxy (MBE), from which photoconductive circuit elements (PCE) have been fabricated. A fabrication procedure will be described which is fully compatible with standard siicdn IC processing technology. Results will be presented demonstrating the reliance of GaAs PCE performance on epitaxial growth conditions and subsequent processing steps. PCE response speeds ranging from <10 to 60 P5 have been observed.

We report about the design of an optical intrachip clock distribution using an IR-laser diode 1 ocated at one edge of an electronic chip and 4 on-chip photoreceivers, all components connected via a reflective HOE. The high divergence angle of the laser beam and its oblique incidence onto the HOE together wfth the 4 fan-out directions require HOE-feature sizes which are in the order of the operation wavelength X0.86 tm. This study considers both multiplex and multifacet computer-generated HOEs . The pattern complexity, the respective pattern resolutions, the beam rouung abilities and the spot quality of the focused output beams are discussed. A calculation method for the estimation of the alignment tolerances is explained and the results will show that our arrangement is feasible using mounting techniques based on micromechanics in preferential etched silicon.

This paper reports the design and fabrication of transmission holographic optical elements (HOEs) for clock distribution. First, we have studied and fabricated a multi-focus doublet HOE. The aberrations due to the wavelength shift between recording ( = 488 nm) and reconstruction ( = 780 nm) have been minimized by an appropriate recording and readout geometry. The diffraction efficiency has been optimized by a copying technique. Second, we have investigated the near-field internal reflection (TIR) holographic recording technique to solve the problems of miniaturization. With this method, we have recorded a lOOxlOO lenlet array with focal lengths of f =400 μm.

In this paper, optical neural networks under development in our laboratory are reported, together with several novel devices for building advanced systems. This paper is divided into three parts. At first, neural network models which permit us to use commercially-available binary-operating devices are described, including both the feedback model and the feedforward model. Secondly, a GaAs/A1GaAs optical associative neuro-chip based on the feedback model, in which 32 neurons are implemented in an optical integrated structure, is introduced. Next a three-layered dynamic optical network which can recognize 26 characters of alphabet is described, using a spatial light modulator as the synaptic interconnection device. The emphasis is placed on the unique learning rule using the feed forward model. In the third part, fututre trend of optical neurochips is discussed.

We present a hybrid opto-electronic implementation of a neural network. In our system the linear part of the algorithm (calculation intensive) is performed optically. The non4inear feedback is controlled by a microcomputer which takes the system output detected using a CD camera, thresholds this output and then uses the thresholded signal to commtnd the spatial light modulator which provides the input to the optical part of the system. One of the major advantages of optics in this type of architecture is its ability to work in parallel on two•dimensional data structures. However, a four-dimensional synaptic matrix is required to stoie the connections between two images. We recall and explain the Frequency Multiplexed Raster (FMR) coding scheme we have developed to enable us to store this four..dimensional array on a planar component (CGH). In addition we present some modifications to the basic synaptic matrix algorithm which improve its performance with correlated images. Finally, after presenting some first experimental results we extend the FMR coding scheme for use in generalized optical interconnections.

Two dimensional jmuem recognition related tasks require the introduction of invariant recognition ftors such as shift, rotation, scale and deformation. In a preliminary work' we showed the inherent capability of our FMR coding to perform shift-invariant retrieval. This capability is essential for other invariances which can be transformed into translated representations. In this paper we propose using this translation invariant coding in a general scheme, whereby a we-processor handles the invariance issues and aligns the input probe to a fixed baseline utilized by the recognition mechanism. This higher order layer includes S "validation switches", each switch corresponding to a position index of translated objects. A validation switch will be eon" f the stimulus matches one of the objects, translated to the position handled by this switch. These switches condition the search, and where a search is needed, they supply the required transformation for stimulus alignment. Several proximity measures, needed for the realization of these switches, are presented and compared, in the context of different weight prescription algorithms.

The non-interacting nature and free spsce propagation of light allows optical interconnectioni with higher degree of connectivity then is pouible with conventional electronic integrated circuits where the interacting electroni must travel along paths which do not meet and where the geometry is confined to two dimensions. Optical interconnection can be realised in free space with arrays of computer generated holograms (CGHs) and are suitable for the interconnection of neural networks. Fan-out presents a problem of low diffraction efficiency and quantisation noise due to the dynamic range of the associated Fourier transform used to encode the hologram. Manipulation of the phase of individual output spots diminLshes the dynamic range and helps to alieviate this problem. Systems assembled using commercially available SLMs, detector arrays and computer workstations together with the CGH arrays are described and the realisation of more complex systems discussed.

We report, for the first time, a simple method using an array of logic XNOR gates to execute the optical process of vector-matrix multiplication or inner-product correlation, where the two levels of light intensity -on and off- can be used to represent bipolar binary vectors. Optical implementation of neural networks based on the Hopfield model is described as an example to show the application of this novel method.

A technique for the determination of the holographic function required to be encoded as a CGH (Computer Generated Hologram) using the complex zeros of the far field profile and a simulated annealing algorithm is to be outlined. This algorithm uses the error-reduction algorithm for an initial estimate of the function required and then optimises this function.

Polymer lightguides ofPMMA, doped with azo 4yes are investigated with respect to their opto-optical or all optical properties. The TRANS-CIS photoisomerisation of the azo dye molecules leads to a reasonable polarization dependent birefringence. This azo dye system is taken as an example for a bifunctional dye system: Adding a UV photorefractive material a combination of UV-structurability and photoinduced reversible birefringence can be obtained.

We consider a synthetic Fourier-domain holographic interconnect, which consists of a vertical array of horizontal stripes that have the structure of a II) Dammann grating [H. Dammann and K. Görtler, Opt. Commun. 3, 312 (1971)]. Two different reconstruction geometries are given, and the realization of the hologram using both conventional (lithographic) techniques and an acousto-optic (reconfigurable) scheme is discussed.

Hybridised optoelectronic modulator arrays with applications for the optical interconnection of VLSI circuits are described. The design criteria and predicted performance of long-wavelength MQW devices are described and progress towards the realisation of optically interfaced high performance silicon circuits is reported.

Self-pulsating, irregular fluctuations and deterministic chaos in the output intensity produced by a BaTiO3 self-pumped phase conjugate mirror has been observed. The behaviour is seen to vary with the incident laser intensity and especially with the specific region of interaction within the probe, which can be influenced by the orientation of the crystal with respect to the input beam. Q uaiitative examinations of the nature of these fluctuations - like the Fourier transform and the autocorrelation function - show signs for deterministic chaotic behaviour. As a qualitative analysis method of nonlinear dynamics, the Grassberger-Procaccia analysis is carried out to compute a lower bound of the Hausdorff dimension of the temporal field. That dimension varies with the critical input parameter, the region of interaction, between 2.25 and 8.5 and allow thus conclusions on the theoretical model needed to describe the system.

The phase profile of the wave scattered by a GRIN lens has been determined using the eikonal approximation for the scattering amplitude. This wave acts as the object wave in a holographic register to construct a GRIN holocoupler. Different geometries in the phase profile distribution are presented as functions of the taper shape.

Tunable filters, wavelength division multiplexers and focusing elements may be needed in future fibre optic networks. Holographically generated elements can meet some of these requirements. Three holographically generated elements will be described in this paper. a tunable reflection filter a polarisation insensitive transmission grating; an array of microlenses. All three elements have been made in volume phase materials, either dichromated gelatin or rehalogenated silver halide. The phase material properties have to be different for each application: thickness and refrztive index modulation need to be optimised in each case. This will be discussed.

A new software module for automated design and optimization of alignment components for optical systems has been developed. The module is a part of the program package for optical design OPTAL. Up to one hundred parameters of the analyzed system, including both optical and mechanical components can be optimized. The optimization is carried out with respect to a pre-determined value which is characteristic of the optical system. The algorithm is a selective directed search of a surrounding n-dimensional grid of points to find a local direction in which the defined value of the optical system decreases and keeps this direction as far as the aberration or the deviation from the ideally aligned system decrease. Then a new direction of variable corrections is determined. The current position of the optical components is computed as a sum of the initial position and the current tolerance of the corresponding elements of the mechanical alignment system. When the final value of the minimization function is reached the current tolerance set is stored for further use. In this way the user can get the distributed production tolerances of the optical and mechanical components in dependence of the desired final result and localize the most important dimensions in the whole optical system, which require the most precise control. These dimensions will be accompanied by the smallest tolerances after the optimization.

Electrooptical modulators are key elements in today's optical telecommunication systems [1] and provide a wide range of applications in parallel optical signal processing [2] and microwave optoelectronics. Because of their low capacitance planar metal-semiconductor-metal (MSM) structures are well suited for applications at very high frequencies that hardly can be realized using conventional pn-devices. MSMphotodetectors have been reported exceeding frequency limits of 100 GHz [3]. Similar results are expected for planar reflection modulators.

Neural networks are a primary candidate architecture for optical computing. One of the major problems in using neural networks for optical computers is that the information holders: the interconnection strengths (or weights) are normally real valued (continuous), whereas optics (light) is only capable of representing a few distinguishable intensity levels (discrete). In this paper a weight discretization paradigm is presented for back(ward error) propagation neural networks which can work with a very limited number of discretization levels. The number of interconnections in a (fully connected) neural network grows quadratically with the number of neurons of the network. Optics can handle a large number of interconnections because of the fact that light beams do not interfere with each other. A vast amount of light beams can therefore be used per unit of area. However the number of different values one can represent in a light beam is very limited. A flexible, portable (machine independent) neural network software package which is capable of weight discretization, is presented. The development of the software and some experiments have been done on personal computers. The major part of the testing, which requires a lot of computation, has been done using a CRAY X-MP/24 super computer.