Diffractive optical elements for optical processing

Abstract

An optical processing system comprising a beam replication arrangement configured to receive a plurality of input beams encoding an input vector, and produce a plurality of spatially-separated replica beams of each input beam; a modulator comprising a plurality of modulator elements configured to encode values of a matrix, each modulator element configured to modulate one of the replica beams according to the respective encoded matrix value; and a light-summing optical arrangement configured to converge the modulated replica beams to encode an output vector representing a result of multiplying the matrix by the input vector. The light-summing optical arrangement comprises a second diffractive optical element.

Description

[0001] DIFFRACTIVE OPTICAL ELEMENTS FOR OPTICAL PROCESSING

[0002] The present invention relates to systems for optical computing, in particular for performing matrix-vector multiplications.

[0003] BACKGROUND

[0004] Optical processing represents an advancing technology increasingly employed to carry out complex computations, in lieu of traditional fully electronic processors, essential for artificial intelligence models, optimization challenges, and other extensive computational tasks. Optical processing is motivated by its ‘optical advantage’ over electronics [1], Optical processors have a fundamentally larger bandwidth than electronic equivalents which allows for large wavelength-division multiplexing (parallel processing of multiple optical signals with subtly different wavelengths, on the same system) and lower computing latency. High spatial parallelism can be realised with optical signals using free-space components such as liquid crystal-based modulators or digital light processors where the spatial density of simultaneous encoding is superior. Optical signals experience approximately zero amplitude loss when transmitted through free space, and very low loss in fibre and waveguide, as well as through optical components with high-reflectivity or anti -reflectivity coatings. Optical computation dynamics are often approximately lossless when they involve propagation through some linear or nonlinear configuration (e.g., a lens for linear Fourier transform).

[0005] Optical core processors (OCs) are potentially very advantageous as dedicated computation engines for large-scale matrix-vector operations which make up the majority of machine learning tasks. State-of-the-art artificial intelligence (Al) architectures such as large-scale machine learning (ML), deep learning (DL), and transformer-based large language models (LLMs) contain billions of trainable parameters, enabling their complex functionality. Using OCs for this role maximally utilises the optical advantage attributes listed previously and yields the largest performance improvement over incumbent technologies such as graphical and tensor processing units (GPUs and TPUs, respectively).

[0006] Analog optical computing hardware acceleration promises increased computing speed alongside improved energy and cost efficiency over the incumbent electronic processing units which support Al models. Optoelectronic hardware and optical computing architectures have the potential to decrease latency and energy consumption [2,3] and the development of optical processor layers of appropriate size and complexity is accelerating.

[0007] A critical sub-system within an optical processor is that which performs optical fan-out, i.e. the replication of optical input signals to perform parallelised computations. Optical fan-out can be realised in various ways such as using diffractive optical elements, microlens arrays, waveguide or fibre splitters, beam splitters etc. However, for an optical processor with very high channel counts (to increase the speed of the optical core), the size, weight, and complexity of the free-space optical system gets easily out of hand. The present invention seeks to address this and other problems with existing optical processors, as well as providing various other advantages as discussed in the following disclosure.

[0008] SUMMARY

[0009] According to a first aspect, there is provided an optical processing system comprising: a beam replication arrangement configured to receive a plurality of input beams encoding an input vector, and produce a plurality of spatially-separated replica beams of each input beam; a modulator comprising a plurality of modulator elements configured to encode values of a matrix, wherein each modulator element is configured to modulate one of the replica beams received from the beam replication arrangement according to the respective encoded matrix value; and a light-summing optical arrangement configured to converge the modulated replica beams output from the modulator elements to encode an output vector representing a result of multiplying the matrix by the input vector, wherein: the light-summing optical arrangement comprises a second diffractive optical element.

[0010] Using a diffractive optical element for the optical summation (or “fan-in”) is advantageous because it provides much greater flexibility in how modulated beams from the modulator are combined. This in turn allows for greater flexibility in and simplification of the optical arrangement in multiple parts of the optical processing system compared to alternative mechanisms for optical summation such as a micro-lens array that can only sum together beams that are physically adjacent to one another.

[0011] Optionally, the beam replication arrangement comprises a first diffractive optical element. Using a DOE for the fan-out as well as the fan-in is advantageous because it allows compact fan-out in multiple dimensions.

[0012] Optionally, an effect of the first diffractive optical element on light propagating through the first diffractive optical element in a forward direction relative to the first diffractive optical element is the same as an effect of the second diffractive optical element on light propagating through the second diffractive optical element in the forward direction relative to the second diffractive optical element. This simplifies the manufacturing and setup of the optical processing system, because the first and second DOEs can be physically similar or the same.Optionally, the beam replication arrangement is configured such that light propagates through the first diffractive optical element in the forward direction; and the light-summing optical arrangement is configured such that light propagates through the second diffractive optical element in a reverse direction opposite to the forward direction. This provides a very conceptually and physically simple arrangement in which the fan-in DOE reverses the effect of the fan-out DOE. This also provides significant flexibility in how the replica beams are arranged on the modulator.

[0013] Optionally, the plurality of modulator elements is further configured to control a phase of the modulated replica beams such that modulated replica beams corresponding to the same input beam are coherent. This provides a compact arrangement to ensure that the modulated replica beams can be combined accurately without loss of amplitude due to mismatched phase.

[0014] Optionally, the modulator comprises a second plurality of modulator elements configured to control a phase of the modulated replica beams such that modulated replica beams corresponding to the same input beam are coherent. Separating out the phase correction may simplify the construction and design of the two pluralities of modulator elements.

[0015] The invention may also be embodied in method of performing analogue computing, comprising: producing a plurality of spatially-separated replica beams of each of a plurality of input beams of light encoding an input vector; modulating the replica beams using a plurality of modulator elements encoding a matrix according to the respective encoded matrix value of each modulator element; and converging, using a second diffractive optical element, the modulated replica beams to encode an output vector representing a result of multiplying the matrix by the input vector.

[0016] LIST OF FIGURES

[0017] Embodiments of the invention will now be described, by way of example only, with reference to the accompanying drawings in which corresponding reference symbols represent corresponding parts, and in which

[0018] Fig. l is a schematic illustrating the logical components of an optical processing system; Fig. 2 depicts the layout of the system architecture;

[0019] Fig. 3 shows the beam replication process for a single input beam;

[0020] Fig. 4 schematically illustrates an exemplary modulator;

[0021] Fig. 5 shows a light-summing optical arrangement using a diffractive optical element;Fig. 6 depicts schematically how the optical processing system operates for a 3x3 array of input beams;

[0022] Fig. 7 depicts schematically the optical fan-out process; and

[0023] Fig. 8 depicts schematically the optical fan-in process.

[0024] DETAILED DESCRIPTION

[0025] The fundamental processes required for performing a matrix -vector multiplication (MVM) are: (1) the element-wise multiplication of two arrays, and (2) the row / column-wise summation of those resulting products to produce the elements of the output vector. Process (1) can be performed by modulation of a plurality of input beams, as will be described in more detail below. Process (2) requires superimposing appropriate subsets of the modulated beams to produce a smaller number of output beams. These processes are performed by the present system as described below.

[0026] Fig. l is a schematic that illustrates the logical components of the optical processing system 1 and the flow of light through the optical processing system 1. Examples of more specific components that can be used to implement each logical component are also given in Fig. 1.

[0027] The optical processing system 1 comprises an optical core processor (OC) 20, and may comprise an optical input arrangement 10 and an optical output arrangement 30. The optical input arrangement 10 produces input beams of light encoding the input vector. The optical core processor 20 receives the input beams and performs the MVM with a matrix encoded in the optical core processor 20. Finally, the optical output arrangement 30 receives the optical output of the optical core processor 20 and produces an output corresponding to the result of the MVM that can be transmitted to other systems, for example an electrical signal.

[0028] The components of the optical processing system 1 are coupled together through firee-space transmission of light between the different components. The ability to manipulate light in free space facilitates modularity and dynamic reconfigurability, making it ideal for adaptive computing paradigms. As such, the present optical core processor 20 is more scalable than photonic processors which are fabricated on 2D substrates and therefore a constrained 2D topology. For example, Mach-Zender interferometer meshes receive X input signals and require X-l layers of interferometer units to perform full MVM.

[0029] Fig. 2 depicts how the system architecture may be laid out, with the optical core 20 comprising most of the free-space optical components and being the main sub-system of concern. As shown in Fig. 2, the optical processing system 1 may further comprise anelectronic control board 5 to receive input data from a host machine and transmit output data back to the host system. The input data could comprise the input vectors and matrix that are to be multiplied. Based on the input data, the electronic control board 5 can control the optical input arrangement 10 and the optical core processor 20 to perform the MVM.

[0030] This makes the optical processing system 1 highly modular and suitable for integration in a wide range of existing electronic systems. Different host system interfaces could be implemented by changing the electronic control board 5 while the rest of the optical processing system 1 remains unchanged. The electronic control board 5 may be further configured to apply some post-processing steps to the resulting data, to decrease the computational burden on the host processor.

[0031] The optical processing system 1 may comprise an optical input arrangement 10 configured to provide a plurality of input beams of light encoding an input vector.

[0032] The optical input arrangement 10 may comprise an emitter array 11, as shown in Fig. 1. The emitter array 11 may be implemented as a plurality of modulated optical emitters configured to produce the plurality of input beams. Any suitable emitters could be used, such as laser diodes (e.g. a vertical -cavity surface-emitting laser, VCSEL). The emitters may be binary emitters.

[0033] Each emitter may provide one of the input beams, such that the number of emitters is equal to the number of input beams. Alternatively, the emitter array 11 could be implemented using one or more light sources and a plurality of optical modulators (such as pixels of an LCD screen) configured to modulate light from the one or more light sources to produce the plurality of input beams. In this case, the number of light sources may be smaller than the number of input beams, but the number of modulators may be equal to the number of input beams. Any suitable modulator could be used, such as an absorptive modulator or a refractive modulator. Each vector is prepared by simultaneously writing each vector element to one or more emitters and / or modulators. This is carried out by control electronics (not pictured).

[0034] The encoding of the input vector into the plurality of input beams can be performed in a variety of different ways depending on the priorities of the specific application. For example, the optical input arrangement 10 may be configured to encode the input vector using temporal modulation of the input beams. The emitters and / or modulators configured to produce the input beams are preferably able to change their output at very high frequencies, in order to maximise the throughput of the system 1. For example, the emitters and / or modulators may have a modulation frequency of at least 1 MHz, optionally at least 100MHZ, optionally atleast 1GHz, optionally at least 5GHz, optionally at least 10GHz, optionally at least 20GHz, optionally at least 40GHz.

[0035] The number of input beams used to encode each element of the input vector may also vary. For example, a single input beam may be used for each element of the input vector. Binary emitters and / or modulators could be used to deconstruct each n-bit analogue vector element as a bitstream of n binary signals in the corresponding input beam. In this case, each element would be transmitted by a single beam over n timesteps of the emitters and / or modulators. Alternatively, a multi-level emitter and / or modulator with more than two levels of output could be used to encode the elements.

[0036] Alternatively or additionally, the optical input arrangement 10 may be configured to encode each element of the input vector using plural input beams. For example, for an n-bit vector element, n input beams may be used to encode each bit of the input beam using a separate input beam. This maximises the speed of processing of each individual vector element, but reduces bandwidth.

[0037] The input beams are transmitted to the optical core processor (OC) 20. The OC 20 receives the input beams and combines the input beams in an appropriate way to perform the desired optical processing operation, such as an MVM.

[0038] The OC 20 comprises three main components: a beam replication arrangement 22, a modulator 24, and a light-summing optical arrangement 26. These will be discussed in turn below.

[0039] The beam replication arrangement 22 is configured to receive the plurality of input beams and produce a plurality of spatially-separated replica beams of each input beam. This process is also referred to as optical “fan-out”.

[0040] The beam replication arrangement 22 may be configured to produce the plurality of spatially-separated replica beams of each input beam as a two-dimensional array of replica beams. The replica beams may be evenly spaced in both dimensions of the 2D array. Each input beam may be replicated into at least 4 replica beams, optionally at least 16, optionally at least 32, optionally at least 64, optionally at least 128.

[0041] The beam replication arrangement 22 is coupled through free space to the optical input arrangement 10, such that the input beams are transmitted to the beam replication arrangement 22 through free space. The fan-out process also occurs in free space, such that the replica beams are transmitted from the beam replication arrangement 22 to the modulator 24 through free space.This is a key advantage of the present system compared to photonic systems. There is limited scope on photonic processor systems for optical fan-out, which is able to massively enhance the parallelism of the present system in a highly scalable way. This is also an inherent advantage of the present system over incumbent electronic systems whose processor units are again fabricated in 2D and can only be accessed across the 2D substrate.

[0042] The fan-out process is illustrated in more detail in Fig. 3. The input beams are provided as a 2D (JXK) array. Fig. 3 shows the process for a single input beam. The input beams are transmitted from the optical input arrangement 10 through free space to the beam replication arrangement 22. The beam replication arrangement 22 generates PXQ replica beams (in this case, 3x3) of each input beam, arranged in a 2D array. The replica beams are produced by passively replicating the incoming signal channels.

[0043] When this is repeated for all JXK input beams (in this case, a 3x3 2D input), this results in an MXN array of replica beams (where M = JP and N = KQ) which in total contain the same amount of power as the sum of the initial JXK array of input beams. Each replica beam will thus contain a factor PQ times less power than the input beam it originates from. The beam replication arrangement 22 may arrange the replica beams such that the pitch of the individual optical signals is reduced but the spot size is preserved.

[0044] As shown in the exemplary arrangement of the figures, the beam replication arrangement 22 may comprise a first diffractive optical element (DOE) 28. The first DOE 28 may implement a bespoke diffractive phase mask to produce the required pattern of replica beams. However, this is not essential, and in other embodiments, other optical components may be used to replicate the input beams. For example, the beam replication arrangement 22 may comprise a pair of micro-lens arrays, optionally arranged in series with a focusing lens. This is a similar arrangement as is sometimes used for an imaging homogenizer.

[0045] The beam replication arrangement 22 may comprise further optical components to properly condition and position the input beams for replication. As shown in Fig. 3, the beam replication arrangement 22 may use a configuration of two-dimensional lens arrays and a 4f imaging system. The first DOE 28 is then placed at the Fourier plane of the 4f imaging system. This design enables flexibility by allowing the OC 20 to be decoupled from the input arrangement, because the operation of the beam replication arrangement 22 can be made more independent of the exact arrangement of the input beams. This can enable functions such as wavelength multiplexing or interoperability with different types of emitter.

[0046] The OC 20 further comprises a modulator 24, also referred to as an optical encoder. Anexemplary modulator 24 is illustrated schematically in Fig. 4. The modulator 24 comprises a plurality of modulator elements configured to encode values of a matrix. The plurality of modulator elements is preferably arranged in rows and columns to form a 2D array of modulator elements. The replica beams from the beam replication arrangement 22 are transmitted through free space to the modulator, where the replica beams are imaged on to individual pixels of the modulator. Thereby, the modulator 24 modulates the replica beams according to the respective encoded matrix value of each modulator element. The modulator 24 may be arranged in a transmissive or reflective configuration, dictating the relative positioning in space of the optical fan-out and fan-in components.

[0047] The pitch of the replica beams in each dimension is chosen to match the pitch of the rows and columns of the modulator 24, such that each modulator element receives exactly one replica beam. Each modulator element is configured to modulate one of the replica beams received from the beam replication arrangement 22 according to the respective encoded matrix value.

[0048] As discussed above, the beam replication arrangement 22 may be configured to recover the individual spot sizes originating from the optical input arrangement 10, such that each replica beam has a spot size matching the spot size of the corresponding input beam. This may be achieved using a 4f imaging system. By controlling the magnification of the replica beams in the beam replication arrangement 22, the spot size of the replica beams can be controlled to ensure each modulator element of the modulator 24 is filled as efficiently as possible by the replica beam. This principle is scalable to any arbitrary dimensions of input beam and / or replica beam array and modulator 24.

[0049] As shown in Fig. 4, the modulator 24 applies weights to the replica beams according to the encoded matrix values. In particular, each modulator element applies a weight to the corresponding replica beam according to a particular encoded element of the matrix. The modulator 24 comprises M*N modulator elements such that each replica beam (representing an element of the input vector) is incident on a single modulator element (also referred to as a pixel) of the optical encoder. Thereby, each element of the input vector is multiplied with a respective element of the matrix, with all of the multiplications occurring simultaneously in parallel.

[0050] The input beams (and thereby the replica beams) may be modulated at very high rates (e.g. 1GHz or higher) such that the throughput of parallel computations is extremely high. Matrix values are fed to the modulator 24 via appropriate encoder driver circuitry. Themodulator elements may be modulated at a much lower rate than the input beams, for example a factor of 103 - 106 times lower. The modulator elements may be modulated at a frequency of at least 1kHz, optionally at least 10kHz, optionally at least IMhz.

[0051] Programming of the encoder driver enables desired sub-sections of the modulator 24 to be updated separately and more quickly than a full matrix update. Depending on the choice of modulator 24 and its programming, the precision of the matrix encoding can be varied to n-bit precision. Depending on the exact nature of the matrix operations that are to be computed, this can enable improved efficiency by appropriate design of how the matrix values are transmitted to the modulator 24.

[0052] Any suitable modulator 24 may be used. For example, the modulator 24 may comprise a spatial light modulator, SLM, (such as a liquid-crystal spatial light modulator, LC-SLM, or liquid-crystal on silicon spatial light modulator, LCoS-SLM), a digital micromirror device, DMD, or an optical micro-electromechanical, MEM, system. Which device is used depends on the priorities of a specific use case. For example, a DMD might be preferable for a 1 -bit MVM where speed is a priority, because DMDs can be modulated very quickly. An LCoS SLM might be preferable for an 8-bit MVM, where system precision is more important than processing speed. Combinations of different devices may also be used. For example, a DMD coupled with an LCoS SLM for intensity correction and n-bit precision can balance speed and precision. A single multi quantum well SLM may provide very high speed, n-bit precision.

[0053] The OC 20 further comprises a light-summing optical arrangement 26 configured to converge the modulated replica beams output from the modulator elements to encode an output vector representing a result of multiplying the matrix by the input vector. The lightsumming optical arrangement 26 may be configured to encode the output vector as a plurality of output beams encoding elements of the output vector. This process is also referred to as optical “fan-in”. Optical fan-in carries out local summation of the P*Q regions of the modulator 24. The resulting sums can then be coupled to detectors in the optical output arrangement 30.

[0054] The spatial size of the output beams is dictated by the design of the light-summing optical arrangement 26. The spatial size is preferably equal to or less than the aperture size of the corresponding receiving component (for example an optical fibre core or detector element) of the output arrangement 30 on which they are incident. This ensures the fidelity of the summation and thus the accuracy of the MVM. Additionally, the angle of incidence onto the receiving component of the output arrangement 30 of every output beam must be equal toor less than the angle of acceptance of the receiving component. Similar to the beam replication arrangement 22, the light-summing optical arrangement 26 may comprise further optical components to properly condition and position the modulated replica beams for summation.

[0055] As discussed above, performing an MVM requires: (1) the element-wise multiplication of two arrays, and (2) the row / column-wise summation of those resulting products to produce the elements of the output vector. Process (2) requires superimposing appropriate subsets of the modulated replica beams to produce a smaller number of output beams encoding the output vector, which is performed by the light-summing optical arrangement 26.

[0056] Previous solutions for process (2) might involve using a micro-lens array (MLA) to focus groups of neighbouring modulated beams together to produce a single output beam at a detector or output element such as an optical fibre. Modulated replica beams from each of a plurality of contiguous modulator regions, representing products of input vector elements and matrix elements, would be incident on a designated region of the MLA, where they experience local summation.

[0057] However, using an MLA has a number of drawbacks. Finding or designing an MLA with precisely-defined properties such as focal lengths, radii of curvature, and thickness to achieve the required superimposition of modulated beams is very difficult. This is because all of these properties are interdependent and difficult to fine tune individually to the required values.

[0058] In addition, an MLA can only sum together neighbouring beams, requiring the input beams to be arranged in specific patterns when they arrive at the MLA. Achieving this may be difficult or result in optical losses in the input arrangement, or require additional optical components to rearrange the modulated beams after the modulator, which increases complexity and introduces additional losses.

[0059] The present optical processing system 1 addresses these difficulties by using a diffractive optical element for optical fan-in. Consequently, the light-summing optical arrangement 26 comprises a second diffractive optical element (DOE) 29. This arrangement is shown in Fig. 5.

[0060] Typically, a diffractive optical element (DOE) is used to divide an incoming light beam into multiple distinct outgoing beams, for example as discussed for the beam replication arrangement 22 above. The DOE, often a precisely engineered microstructure with phase or amplitude variations, manipulates the incident light through diffraction. When a singleincoming light beam encounters the DOE, the beam’s wavefront is altered to produce controlled interference patterns, resulting in replicated beams at predetermined angles. For example, a DOE designed to transform a Gaussian beam into a 3x3 2D array will create nine equally spaced beams, each maintaining the Gaussian profile, effectively forming a grid-like distribution of light. This process is highly efficient, allowing parallel beam generation while maintaining uniform intensity across the array.

[0061] However, just as a DOE can replicate a beam in the forward pass, it can also combine beams in the reverse pass. For example, if a single beam is replicated into a 1 *M array of beams, a 1 *M array of beams incident on the DOE would combine them into a single beam. Fig. 5 demonstrates this combination for the 3x3 array generated by the beam replication arrangement 22 in Fig. 3.

[0062] A diffractive optical element (DOE) can function in reverse due to the principle of optical reciprocity, which states that the behaviour of light in a system is symmetrical when the direction of light propagation is reversed. If a DOE is designed to split a single Gaussian beam into an (M*N) array, it can also operate in reverse mode to recombine beams from that array into a single output. This occurs because the microstructure of the DOE, which manipulates the phase and amplitude of the incoming light to create multiple beams, can similarly guide multiple incident beams back into a coherent, focused single output when the beams strike it from the reverse direction. The same optical path and diffraction properties are used in both cases. Thus, when an (MxN) array of beams is incident on the DOE at the correct alignment and phase, the diffractive patterns merge the beams constructively, forming a unified output beam.

[0063] The present optical processing system 1 exploits this bidirectional functionality to provide an efficient and convenient way to converge and combine the modulated replica beams into the output beams.

[0064] Fig. 6 depicts schematically how the optical processing system 1 operates for a 3x3 array of input beams. In a typical implementation, the input beams may be carried by an input array of optical fibres spaced at 250pm with an effective dimension of 500x500pm.

[0065] Using a beam replication arrangement 22 with a first DOE 28 such as shown in Fig. 3, with a 4f imaging system in place, the input optical fibre array is set at a distance f from the first lens (having focal length f). The first lens collimates the input beams onto the first DOE 28 placed at the Fourier plane of the 4f imaging system. The first DOE 28 then, acting as a grating, replicates the input beams into the diffraction orders such that each input beam isreplicated, in this example, into a 3x3 array resulting in a total of 81 replica beams in a 9x9 array, each having a Gaussian profile. The replica beams associated with each input beam are located in a replicated zone 40, as shown in Fig. 6. Only three of the nine replicated zones are labelled in Fig. 6, Fig. 7, and Fig. 8 for ease of viewing. In this example, each replicated zone 40 contains 9 replica beams arranged in a 3x3 array. This is shown as step 1 in Fig. 6, and also illustrated in Fig. 7. These 81 replica beams will them be transmitted to the modulator 24 where they are modulated to perform the multiplication with the encoded matrix.

[0066] In order to perform the summation correctly at the final image plane at the output of the light summing arrangement 26 to correctly compute the MVM, each individual modulated replica beam from a given replicated zone 40 must be added to the corresponding modulated replica beam from each other replicated zone 40 (step 2). This is shown as step 2 in Fig. 6, and also illustrated in Fig. 8.

[0067] This summation cannot be accomplished directly using an MLA, as is commonly done in existing implementations. This is because, as mentioned above, an MLA can only converge adjacent beams such as the beams in the same replication zone 40. To perform the summation using an MLA therefore requires an extra step where the beams are rearranged and then summed up as shown in steps 2.1 and 2.2 of Fig. 6. The rearranging could be performed either on the replica beams before they reach the modulator 24, or on the modulated replica beams after they leave the modulator 24 but before they reach the light summing arrangement 26. In either case, the rearranging of the beams will require additional optical components and add size and complexity to the OC 20. Depending on the mechanism used to rearrange the replica beams, optical losses may also be increased.

[0068] However, the present system is based on the realisation that the reverse of step 2 in Fig.

[0069] 6 is essentially the same process as the original fan-out step 1 used to perform the beam replication. An appropriately-designed DOE with a particular phase pattern would be able to accomplish the fan-out (beam replication) step 1 as discussed above. Therefore, using a DOE in reverse can accomplish step 2 directly in a single operation. Consequently, the lightsumming optical arrangement comprises a second diffractive optical element (DOE) 29. The second DOE 29 has an appropriate design to converge the modulated replica beams output from the modulator elements to encode the output vector representing the result of multiplying the matrix by the input vector.

[0070] As discussed above, the beam replication arrangement may comprise a first DOE 28. In this case, an effect of the first diffractive optical element 28 on light propagating through thefirst diffractive optical element 28 in a forward direction relative to the first diffractive optical element 28 may be the same as an effect of the second diffractive optical element 29 on light propagating through the second diffractive optical element 29 in the forward direction relative to the second diffractive optical element 29. In other words, the first DOE 28 and the second DOE 29 may have the same design and / or microstructure such that their effect on propagating light is the same. This ensures that the second DOE 29 can reverse the effect of the first DOE 28 for fan-in. It also simplifies manufacturing and design of the system, because only one type of DOE needs to be designed and manufactured.

[0071] Where the first DOE 28 and the second DOE 29 have the same effect on propagating light, the beam replication arrangement 22 may be configured such that light propagates through the first diffractive optical element 28 in the forward direction, and the light-summing optical arrangement 26 may be configured such that light propagates through the second diffractive optical element 29 in a reverse direction opposite to the forward direction.

[0072] One caveat with the use of the second DOE 29 for optical fan-in is that the modulated replica beams arriving at the light-summing optical arrangement 26 should ideally be coherent. Incoherency will reduce the accuracy of the MVM result because the modulated replica beams will not be accurately combined by the second DOE 29 into the output beams.

[0073] The relative phases of the replica beams may be affected by factors such as minor variations caused by the different modulation applied to each replica beam. To compensate for this, the plurality of modulator elements may be further configured to control a phase of the modulated replica beams such that modulated replica beams corresponding to the same input beam are coherent.

[0074] In particular, the plurality of modulator elements may be configured to ensure that the modulated replica beams arriving at the light-summing optical arrangement 26 have the same relative phases as the corresponding replica beams produced by the beam replication arrangements, which may comprise the first DOE 28. However, it is not strictly necessary for the phase relationships among all the replica beams to be preserved after modulation by the modulator in order that improved accuracy is preserved. The modulated replica beams that are combined into a single output beam by the light-summing optical arrangement should be coherent such that the second DOE 29 is able to accurately and consistently sum them up into a single beam. The plurality of modulator elements may be configured to achieve this.

[0075] Modulated replica beams that are combined into different output beams by the second DOE 29 need not necessarily be coherent, and may be incoherent, with one another.The phase control may be provided by the same plurality of modulator elements as those that provide the modulation encoding the matrix values. This will produce the most compact optical system with the fewest individual components. However, accurately controlling both the amplitude and phase of each modulated replica beam simultaneously with a single modulator element may be difficult. Therefore, an alternative arrangement may be where the modulator 24 comprises a second plurality of modulator elements. The second plurality of modulator elements are then configured to control a phase of the modulated replica beams such that modulated replica beams corresponding to the same input beam are coherent. The number of modulator elements in the second plurality of modulator elements used to control the phase will generally be equal to the number of modulator elements in the plurality of modulator elements.

[0076] The use of the second DOE 29 for optical fan-in bypasses several disadvantages associated with existing fan-in solutions such as MLAs, as described above. A DOE can also provide a 100% fill factor of downstream optical components, such as optical fibres used to receive the output beams, thereby maximising usage of the optical components. This may be difficult or impossible to achieve with existing arrangements. A DOE can also correct for aberrations the system may face, for example introduced by lenses used to focus the input beams, replica beams, or output beams through a multi-step approach during fabrication of the DOE.

[0077] As illustrated in Fig. 5, a 4f imaging system may be used between the output of the modulator 24 and the second DOE 29. The second DOE 29 is placed a distance f from the imaging plane of the 4f system, where f in this case is the local length of the optical fan-in component, such that the beams are collimated after passing through the optical fan-in component. As for the beam replication arrangement, this design enables flexibility by allowing the OC 20 to be decoupled from the output arrangement 30.

[0078] Following the OC 20, the optical processing system 1 may further comprise an optical output arrangement 30 configured to receive the plurality of output beams and produce an output corresponding to the output vector based on the converged beams.

[0079] The optical output arrangement 30 may comprise a plurality of light-guiding elements configured to receive the plurality of output beams. The light-guiding elements may be configured as a three-dimensional light-guide. The light-guide of the output arrangement may be similar to the light-guide of the optical input arrangement 10. For example, the three-dimensional light guide may comprise a two-dimensional array of elongate light-guidingelements. The light-guiding elements may be optical fibres. The properties of the output fibre array including fibre core and cladding diameters and materials, are customisable to suit end-user needs. In one implementation, large mode area (LMA) optical fibres might be used to maximise the coupling efficiency of summed optical signals from free space to the fibre array. Multi-mode optical fibres (MMFs) may also be used.

[0080] The optical output arrangement 30 receives the summed signal groups (output beams) from the OC 20. For example, each output beam may be focussed into the core of one of the optical fibres of the light guide.

[0081] The optical output arrangement 30 may further comprise a detector arrangement configured to individually read out each of the plurality of output beams. The output beams may be directed to the detector arrangement by the light guide of the output arrangement, such that the light guide couples the detector arrangement to the output of the OC 20.

[0082] The detector arrangement may comprise an array of optical detectors, for example an array of photodiodes. Within the detector arrangement, the summed optical signals represented by the output beams from the OC 20 are converted to electronic signals which are then transferred to receiver electronics for processing. The result of the MVM calculation is read out of the receiver electronics at the same clock rate as the input signals are modulated.

[0083] As mentioned above, elements of the input vector may be encoded with plural input beams, and the intensities of the input beams may be related to one another by an exponential decay function. A similar effect can also be achieved when an n-bit analogue vector element is encoded as a bitstream of n binary signals in a single input beam. In this latter case, the detector arrangement may be tuned to have an appropriate RC time constant (matching the clock rate of the modulation of the input beam used to encode the bitstream) to exponentially weight the MVM output data to reconstruct the n-bit analogue result. For example, a transimpedance amplifier could be used with an appropriate time constant, such that in the sensing window of the amplifier the first bit of the bitstream corresponding to a single input vector element is amplified with a relative weight of 1, the second bit with a relative weight of 0.5, the third with a relative weight of 0.25, and so on. In this way, the overall value of the vector element represented by the entire bitstream can be accumulated in a single value quickly and straightforwardly without the need for further post-processing.

[0084] The optical processing system 1 can be implemented as a plugin card, for example in the style of a PCIe plugin card. This provides a small, compact, implementation suitable for desktop and workstation applications. Larger implementations of the optical processingsystem 1 with more input beams and modulator elements would provide increased throughput and scale more suitable for datacentres, for example implemented in a unit matching standard server rack dimension.

[0085] The OC 20 may also be configured to perform basic logic and mathematical functions, such as one or more of addition, subtraction, matrix transform, convolution, differentiation, and rotation. The availability of these operations removes the need for data that is being processed on the MVM system to be copied to the host system for an operation and then back to the MVM for continued processing. This removes memory copy overhead associated with moving data to and from the host system.

[0086] The use of three-dimensional light guides, such as optical fibre arrays, to connect the OC 20 to the optical input and output arrangements 10, 30 provides significant flexibility of system architecture. Moreover, the layout of the OC 20 can be fully customised and appropriately condensed by choice of beam-steering optics. For example, the OC 20 might be laid out in a linear configuration with a transmissive optical encoder, or in a raster configuration to consume least volume for compatibility with existing server racks in data centres.

[0087] The optical processing system 1 described above can be used to perform a method of performing analogue computing. The method comprises:

[0088] producing, using a first diffractive optical element, a plurality of spatially-separated replica beams of each of a plurality of input beams of light encoding an input vector;

[0089] modulating the replica beams using a plurality of modulator elements encoding a matrix according to the respective encoded matrix value of each modulator element; and converging, using a second diffractive optical element, the modulated replica beams to encode an output vector representing a result of multiplying the matrix by the input vector. REFERENCES E McMahon, P. L. The physics of optical computing. Preprint at https: / / doi.org / 10.48550 / arXiv.2308.00088 (2023).

[0090] 2. Guo, X., Barrett, T. D., Wang, Z. M. & Lvovsky, A. I. B ackpropagation through nonlinear units for the all-optical training of neural networks. Photon. Res., PRJ 9, B71-B80 (2021).

[0091] 3. Wang, T. et al. An optical neural network using less than 1 photon per multiplication. Nat Commun 13, 123 (2022).

Claims

CLAIMS1. An optical processing system comprising:a beam replication arrangement configured to receive a plurality of input beams encoding an input vector, and produce a plurality of spatially-separated replica beams of each input beam;a modulator comprising a plurality of modulator elements configured to encode values of a matrix, wherein each modulator element is configured to modulate one of the replica beams received from the beam replication arrangement according to the respective encoded matrix value; anda light-summing optical arrangement configured to converge the modulated replica beams output from the modulator elements to encode an output vector representing a result of multiplying the matrix by the input vector,wherein:the light-summing optical arrangement comprises a second diffractive optical element.

2. The optical processing system of claim 1, wherein the beam replication arrangement comprises a first diffractive optical element.

3. The optical processing system of claim 2, wherein an effect of the first diffractive optical element on light propagating through the first diffractive optical element in a forward direction relative to the first diffractive optical element is the same as an effect of the second diffractive optical element on light propagating through the second diffractive optical element in the forward direction relative to the second diffractive optical element.

4. The optical processing system of claim 3, wherein:the beam replication arrangement is configured such that light propagates through the first diffractive optical element in the forward direction; andthe light-summing optical arrangement is configured such that light propagates through the second diffractive optical element in a reverse direction opposite to the forward direction.

5. The topical processing system of any preceding claim, wherein the plurality of modulator elements is further configured to control a phase of the modulated replica beamssuch that modulated replica beams corresponding to the same input beam are coherent.

6. The optical processing system of any of claims 1 to 4, wherein the modulator comprises a second plurality of modulator elements configured to control a phase of the modulated replica beams such that modulated replica beams corresponding to the same input beam are coherent.

7. The optical processing system of claim 6, wherein a number of modulator elements in the second plurality of modulator elements is equal to a number of modulator elements in the plurality of modulator elements.

8. The optical processing system of any preceding claim, wherein the beam replication arrangement is configured to produce the plurality of spatially-separated replica beams of each input beam as a two-dimensional array of replica beams.

9. The optical processing system of any preceding claim, wherein the modulator comprises a spatial light modulator, for example a liquid-crystal spatial light modulator.

10. The optical processing system of any preceding claim, wherein the modulator comprises a digital micromirror device.

11. The optical processing system of any preceding claim, wherein the light-summing optical arrangement is configured to encode the output vector as a plurality of output beams encoding elements of the output vector.

12. A method of performing analogue computing, comprising:producing a plurality of spatially-separated replica beams of each of a plurality of input beams of light encoding an input vector;modulating the replica beams using a plurality of modulator elements encoding a matrix according to the respective encoded matrix value of each modulator element; and converging, using a second diffractive optical element, the modulated replica beams to encode an output vector representing a result of multiplying the matrix by the input vector.

13. The method of claim 12, wherein producing the plurality of spatially-separated replica beams is performed using a first diffractive optical element.

Images