Optical device for performing matrix-vector multiplication

Abstract

Basic block, for use in photonic processor, is presented, comprising an optical device having N (N≥1) waveguide structures between device's input and output ports. Each i-th waveguide structure defines M light guiding units (M≥2), each comprising active medium, excitable by a respective control light signal to affect intensity of j -th light signal, passing through j -th light guiding unit, with a weighting factor defined by power of the control light signal. This provides NxM weighted light signals at NxM output ports of the TV waveguide structures. The device's input port has TV splitters producing AV light signals, each of power Xy indicative of i-th component of vector X=[X1,...,XN] to be multiplied by matrix or vector of size NxM. The output port has a number of combiners providing output light signal of power Y j indicative of j -th component of vector Y=[Y1,...Y M ], presenting a matrix-vector or vector-vector multiplication of the input light signal.

Description

[0001] OPTICAL DEVICE FOR PERFORMING MATRIX- VECTOR MULTIPLICATION TECHNOLOGICAL FIELD AND BACKGROUND

[0002] The present disclosure is in the general field of optical computational devices and more particularly, in the field of photonic matrix-vector multiplication.

[0003] Matrix computation is one of the most widely used and indispensable tools of information processing in science and engineering. Matrix computation contributes most of the computational overheads in modern signal processing and artificial intelligence algorithms. Most signal processing, such as the discrete Fourier transform and convolution operation, can be attributed to matrix computations.

[0004] On the other hand, since the concept of artificial intelligence (Al) was put forward in 1956 for the first time, artificial neural networks (ANNs) have been rapidly developed and widely used in various fields. Due to the continuous substantial increase in information capacity, general electronic processors seem to be incapable of executing high-complexity Al tasks in the foreseeable future. To solve this challenge, chips oriented to Al applications have emerged, such as neural network processing units (NPUs). At present, Al chips have been widely used in almost every type of big data processing in areas such as search, news, e-commerce, cloud computing, and inverse design of functional devices. Typically, neural network algorithms represented by deep learning, such as forward neural networks (FNNs), convolutional neural networks (CNNs) and spiking neural networks (SNNs), are characterized by many training parameters, especially in heavy matrix computations.

[0005] Traditionally, matrix computation is completed by an electrical digital signal processor, and its speed and power consumption are greatly limited by the nature of the electronic devices themselves. Therefore, traditional electrical methods are hard to simultaneously achieve high-capacity and low-latency matrix information processing limited by the Moore’s law. However, for some applications, such as ultrafast neural networks, large bandwidth and low latency are simultaneously required.Photonic accelerators are designed to accelerate specific categories of computing in the optical domain, especially matrix multiplication, to address the growing demand for computing resources and capacity. Photonic matrix multiplication has much potential to expand the domain of telecommunication, and artificial intelligence benefiting from its superior performance.

[0006] Various photonic processors are described in US10,429,580 and US10,838,139, as well as in the following patent publications being assigned to the assignee of the present application: US2022 / 291569; US12,001,947; WO2023 / 248218.

[0007] GENERAL DESCRIPTION

[0008] There is a need in the art for a novel approach in photonic matrix-vector multiplication (MVM).

[0009] Recent research in photonic matrix multiplication has flourished and may provide opportunities to develop applications that are unachievable at present by conventional electronic processors. Photonic circuits are well suited to high-performance implementations of neural networks for two predominant reasons: interconnectivity and linear operations. Connections between pairs of artificial neurons are described by a scalar synaptic weight (a primary memory element), so the layout of interconnections can be represented as a matrix-vector operation, where the input to each neuron is the dot product of the output from connected neurons attenuated by a weight vector. Optical signals can be multiplied by transmission through tunable waveguide elements, and they can be added through wavelength-division multiplexing (WDM) by accumulation of carriers in semiconductors, electronic currents, or changes in the crystal structure of a material induced by photons. Neural networks require relatively long-range connections to perform non-trivial distributed information processing. When comparing metal wire connections with photonic waveguides, optical signals experience lower attenuation and generate less heat (the latter provided the light source of off-chip) as a function of distance.

[0010] Connections between a pair of neurons are weighted by their intervening synapse. These synaptic weights are scalar multipliers. Before being received by downstream neurons, the weighted signals from upstream neurons are summed. The weighted interconnects can therefore be represented by a matrix whose entries are the weightvalues, with each entry multiplying a particular synapse’s input signal. One purpose of the photonic system is to perform that matrix multiplication.

[0011] The present disclosure provides a system and a method for photonic matrix-vector multiplication (MVM). The photonic MVM technique of the present disclosure is implemented using an optical device, which can be used as a basic block in a photonic processor, for performing a vector-matrix multiplication operation on a plurality of input optical signals representing an input vector.

[0012] In an exemplary MVM configuration of the present disclosure, the input vector X is passed through a transmission matrix W towards an output, to generate an output vector Y, representing the product of the matrix W and the vector X, i.e., Y = XW. Thus, the optical device / basic block has a respective number of inputs and outputs and is configured and operable to implement said transmission matrix W.

[0013] Additionally, the present disclosure provides a photonic processor including a number K (K>1) of the basic blocks of the disclosure enabling cascaded MVM.

[0014] The basic block of the present disclosure provides considerable advantages compared to existing optical computation systems utilizing free space optics. For example, the basic block and / or the photonic processor of the present disclosure may be easily packaged in a small and handy package, or as a single component on a printed circuit board or on a chip (e.g., glass chip). The photonic processor further provides an enhanced modularity, as several modules (basic blocks) may be packaged together to build up a more complex system.

[0015] Thus, according to a broad aspect of the present disclosure, it provides a basic block for use in a photonic processor for performing light signals multiplication, the basic block comprising an optical device, wherein:

[0016] the optical device comprises a number N (N>1) of waveguide structures between input port and output port of the optical device, each i-th waveguide structure (z = 1,..., N) being configured to define M light guiding units (M>2) extending between M input ports and M output ports of the waveguide structure, respectively, to enable directional propagation of AT light signals through the AT light guiding units, each j -th light guiding unit (j=l,...M) comprising active medium, which, while being pumped by a predetermined control light signal propagating through the j -th light guiding unit, affectsintensity of a j -th light signal, interacting with said medium while passing through said j-th light guiding unit, with a corresponding weighting factor defined at least by a power of the control light signal, said optical device thereby providing NxM weighted light signals a NxM output ports of the N waveguide structures;

[0017] the input port of the optical device comprises the number TV of splitting assemblies associated and aligned with the inputs of the N waveguide structures, respectively, each i-th splitting assembly being configured and operable for splitting input light of a first wavelength Xu of power into / W input light powers forming the A / light signals, each of power Xij (j=l,...M), directed into the respective M light guiding units of the i-th waveguide structure, said M light signals of the power Xy being indicative of an i-th component of a vector X=[XI,...,XN] to be multiplied by a matrix or vector of size NxM,' the output port of the optical device comprises a predetermined number of combiner assemblies associated with the NxM output ports and configured to provide an output light signal of power Yj being indicative of a / -th component of a vector Y=[YI,...,YM], said vector Y=[YI,...,YM] presenting a matrix-vector or vector-vector multiplication of the input light signal.

[0018] The operation of the basic block utilizes the control light signals for affecting the M light signals passing through the AT light guiding units. Thus, in some embodiments, the input port of the optical device further comprises the number N of control ports associated with the A / light guiding units of each of the N waveguide structures and being configured and operable to input NxM control light signals into the NxM light guiding units to thereby amplify the light signals passing through said light guiding units.

[0019] In some embodiments, the basic block comprises a single waveguide structure (N=l). In this case the output port comprises a single combiner, said basic block thus performing vector-vector multiplication V of the input light signal X producing the output signal Y=XV.

[0020] In some embodiments, the basic block comprises an array (stack) of said N (N>2) waveguide structures, and the output port of the optical device comprises the number AT of combiner assemblies associated with the NxM output ports. Each combiner assembly is configured for combining N light signals being output from the N j-th light guiding units of the N waveguide structures to thereby provide an output light signal of power Yjbeing indicative of the / -th component of a vector Y=[Y

[0021]

[0022] said vector Y=[Y I,...,YM] presenting the matrix-vector multiplication =XW.

[0023] In some embodiments, the basic block is configured such that spectral properties of the control signals are different from spectral properties of the respective light signals.

[0024] In some embodiments, the active medium comprises ions excitable by the spectral properties of the control signal and capable of responding to such excitation by spontaneous emission of light of the spectral properties of the light signal propagating through the respective light guiding unit.

[0025] In some embodiments, the light guiding units of the N waveguide structures comprise the same active medium, whereas in other embodiments, the light guiding units of the N waveguide structures comprise the light guiding units containing different active media.

[0026] In some embodiments, light signals of the i-th waveguide structure have the wavelength Xu of about 1550 nm, and said control signals being input into said light guiding units of the i-th waveguide structure are of a second wavelength X21 of 980 nm.

[0027] The basic block of the present disclosure may be configured such that the control signals being input into the A / light guiding units of the waveguide structure comprise the control signals of different powers, thereby providing amplification or attenuation of the light signals with different weighting factors.

[0028] In the embodiments, where the basic block comprises an array (stack) of said N (N>2) waveguide structures, the control signals being input into the N waveguide structures may comprise the control signals of different powers, to thereby amplify the light signals propagating through the N waveguide structures with different weighting factors.

[0029] The basic block of the present disclosure may be configured such that the light signals being input into the M light guiding units of the i-th waveguide structure have substantially equal power. Alternatively, in some embodiments, the light signals being input into the M light guiding units of the i-th waveguide structure comprise the light signals of different powers.In the embodiments, where the basic block comprises an array (stack) of said N (N>2) waveguide structures, the light signals being input into the N waveguide structures are of substantially the same wavelength An. In these embodiments, the light signals being input into the N waveguide structures may have coherent optical condition, or alternatively, may comprise the light signals of incoherent optical condition.

[0030] Additionally, in the embodiments, where the basic block comprises an array (stack) of said N (N>2) waveguide structures and M light guiding units (M>2) extending between M input ports and M output ports of the waveguide structure, the numbers N and M satisfy the condition: N=M, whereas in other embodiments, the numbers N and A / may satisfy the condition: N^M.

[0031] The basic block of the present disclosure may be configured such that the waveguide structure of the optical device is formed by a number L (L>1) of glass chips, said number of glass chips defining said AT light guiding units.

[0032] In some embodiments, the optical device comprises at least one flexible fiberbased structure.

[0033] According to another broad aspect of the present disclosure, it provides a photonic processor comprising a number K (K>1) of the basic bocks configured according to any one of the embodiments described above, providing cascaded matrix-vector or vectorvector multiplication.

[0034] The photonic processor may comprise a detection system at the output port of the k-th basic block. The detection system may comprise at least one optical detector collecting the light signals of vector YR, and a processor configured and operable to translate said light signals of the vector YR into a corresponding mathematical value.

[0035] The processor of the detection system may be further configured and operable to correct said corresponding mathematical value to compensate effects of coherent interactions between the light signals.

[0036] The photonic processor may further include at least one pair of optically coupled basic blocks. The pair of optically coupled basic blocks comprises: a first preceding basic block, which is formed by A waveguide structures, each containing AT light guiding units, and which provides the vector-matrix or vector-vector multiplication; and a secondsuccessive basic block which comprises one or more waveguide structures, each being formed by the light guiding units and providing vector-matrix or vector-vector multiplication.

[0037] In some other embodiments, the photonic processor comprises at least one pair of optically coupled basic blocks, the pair comprising: a first preceding basic block, which is formed by N waveguide structures, each containing M light guiding units, and which provides the vector-matrix or vector-vector multiplication Y i(lxM)= X(lxN)xWi(NxM); and a second successive basic block, which is formed by M waveguide structures, each containing P light guiding units, and which provides vector-matrix multiplication Y2(1XP)= YI(1XM)XW2(MXP). In these embodiments, said N waveguide structures of the first basic block may be arranged in a spaced-apart substantially parallel relationship along a first axis, and said M waveguide structures of the second basic block may be arranged in a spaced-apart substantially parallel relationship along a second axis forming a substantially 90 degrees tilt with respect to the first axis.

[0038] In some other embodiments, the photonic processor comprises a number K / 2 of said pairs of the optically coupled basic blocks, wherein odd basic blocks are configured as said first preceding basic block and even basic blocks are configured as said second successive basic block.

[0039] According to yet another broad aspect of the present disclosure, it provides a basic block for use in a photonic processor for performing matrix-vector multiplication, the basic block comprising an optical device, wherein:

[0040] the optical device comprises an array (stack) of N (N>2) waveguide structures between an input port and an output port of the optical device, wherein each i-th waveguide structure (i = 1,..., N) is configured to define M light guiding units (M>2) extending between M input ports and M output ports of the waveguide structure, respectively, to enable directional propagation of M light signals through the M light guiding units, and wherein each j-th light guiding unit (j=l,...M) of the N waveguide structures comprises active medium, which, while being pumped by predetermined control light signal propagating through the j-th light guiding unit, affects intensity of a light signal, interacting with said medium while passing through said light guiding unit, with a corresponding weighting factor defined at least by power of the control light signal,said optical device thereby providing NxM weighted light signals at NxM output ports of the N waveguide structures;

[0041] the input port of the optical device comprises the number TV of splitting assemblies associated and aligned with the input ports of the N waveguide structures, respectively, each i-th splitting assembly being configured and operable for splitting an input light of a first wavelength Xu of power Xi into M input light powers forming the M light signals, each of power Xij (j=l, ...M), directed into the respective M light guiding units of the i-th waveguide structure, said M light signals of the power Xij being indicative of an i-th component of a vector X=[X1,...,XN] to be multiplied by a matrix W of size NxM,' the output port of the optical device comprises the number M of combiner assemblies associated with the NxM output ports, each combiner assembly being configured for combining TV light signals being output from the N j-th light guiding units of the N waveguide structures to provide an output light signal of power Yj being indicative of a j-th component of a vector Y=[Y1,...,YM], said vector Y=[Y1,...,YM] presenting the matrix-vector multiplication Y=XW.

[0042] According to yet another broad aspect of the present disclosure, it provides a basic block for use in a photonic processor for performing matrix-vector or vector-vector multiplication, the basic block comprising an optical device, comprising:

[0043] the optical device comprises a glass chip comprising a number N (N>1) of waveguide structures between an input port and an output port of the optical device, each i-th waveguide structure (i = 1,..., N) being configured to define M light guiding units (M>2) extending between AV input ports and M output ports of the waveguide structure, respectively, to enable directional propagation of M light signals through the M light guiding units, each j-th light guiding unit (j=l,...M) comprising active medium, which, while being pumped by a predetermined control light signal propagating through the j-th light guiding unit, affects intensity of a j-th light signal, interacting with said medium while passing through said light guiding unit, with a corresponding weighting factor defined at least by power of the control light signal, said optical device thereby providing NxM weighted light signals at NxM output ports of the N waveguide structures;

[0044] the input port of the optical device comprises the number N of splitting assemblies associated and aligned with the input ports of the N waveguide structures, respectively, each i-th splitting assembly being configured and operable for splitting an input light ofa first wavelength n of power Xi into M input light powers forming the M light signals, each of power Xij (j=l,...M), directed into the respective M light guiding units of the i-th waveguide structure, said M light signals of the power Xij being indicative of an i-th component of a vector X=[X1,...,XN] to be multiplied by a matrix or vector of size NxM;

[0045] the output port of the optical device comprises a predetermined number of combiner assemblies associated with the NxM output ports, to provide an output light signal of power Yj being indicative of a j-th component of a vector Y=[Y1,...,YM], said vector Y=[Y1,...,YM] presenting the matrix-vector or vector-vector multiplication of the input light signal.

[0046] The present disclosure also provides a neural network comprising one or more neuron units, wherein each neuron unit comprises the above-described photonic processor.

[0047] BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to better understand the subject matter that is disclosed herein and to exemplify how it may be carried out in practice, embodiments will now be described, by way of non-limiting examples only, with reference to the accompanying drawings, in which:

[0049] Figs. 1A and IB show schematically two examples of a basic block configured as an optical device, for use in a photonic processor for performing light signals multiplication according to the principles of the present disclosure;

[0050] Fig. 1C shows a photonic processor including a number K (K>1) of the basic bocks, each configured as the optical device of Fig. IB;

[0051] Figs. 2A to 2C show an example of an input port and a single (z-th out of N) waveguide structure including four light guiding units;

[0052] Fig. 3A shows an example of an optical device configured for vector-vector multiplication (lx4)x(4xl);

[0053] Fig. 3B exemplifies an optical device of size 4x4;

[0054] Figs. 4A and 4B show a photonic processor including a pair of optically coupled basic blocks; andFig. 5 exemplifies an optical device of size 8x16 where four glass chips, each including a waveguide structure with four light guiding units, define the total of 16 light guiding units.

[0055] DETAILED DESCRIPTION OF EMBODIMENTS

[0056] Reference is made to Figs. 1A and IB showing schematically two examples of a basic block 10 for use in a photonic processor 100 for performing matrix-vector multiplication according to the principles of the present disclosure. The basic block 10 includes an optical device 12 which includes a number N of waveguide structure(s) 14 between input port 16A and output port 18A of the optical device.

[0057] To facilitate understanding, the same reference numbers are used to identify functionally identical elements.

[0058] Generally, the optical device 12 includes one or more waveguide structures, i.e., N>1. In the example of Fig. 1A, the optical device 12 includes a single waveguide structure 14 between input port 16A and output port 18A of the optical device and, as will be described further below the device 12 is configured for vector multiplication.

[0059] In the example of Fig. IB, the optical device 12 includes an array (stack) of N (N>2) waveguide structures 14 between input port 16A and output port 18A of the optical device, and, as described below, this optical device operates as vector-matrix multiplication.

[0060] In both configurations, each i-th waveguide structure 14 (z = 1 , .. N) is configured to define M light guiding units 20 (M>2) extending between M input ports 16B and M output ports 18B of the waveguide structure 14, respectively, to enable directional propagation ofM light signals, ILy (e.g., ILn,...,ILiMor ILNI,...,ILNM), through the M light guiding units 20.

[0061] The waveguide structure (e.g., fiber-based structure or photonic crystal based structure) can be implemented as a planar structure, e.g., glass structure. It should also be noted that the light guiding unit may in some embodiments be configured as a single-core or multi-core structure.

[0062] Each j-th light guiding unit 20 (j=l,...M) of the N waveguide structures 14 includes active medium, which, while being pumped by predetermined control lightsignal, CSij, propagating through the j-th light guiding unit 20, affects (amplifies or attenuates) the respectively light signal ILij passing through said light guiding unit 20 with a corresponding weighting factor defined by the power of the control light signal CSij.

[0063] It should be noted that in the description below the effect of pumping light on the respective light signal is described as "amplification". However, this term should be interpreted broadly covering also an effect of attenuation (involving weighting factors between 0 and 1).

[0064] Thus, the light propagation through each z-th waveguide structure results in M amplified / weighted outputs Wij (j=l .. M). The optical device 12 thereby provides NxM weighted light signals, WLij, at NxM output ports 18B of the N waveguide structures 14.

[0065] The input port 16A of the optical device 12 includes the number N of splitting assemblies 22.

[0066] In the example of Fig. 1 A, there is a single splitting assembly SA which splits an input light IL of certain first wavelength Xu of power X into M input light components / powers, ILj (j = 1,..., M), forming the AT light signals, each of power A), directed into the respective AT light guiding units 20 of the waveguide structure 14. The AT light signals of the power Xj(j= 1 , ... , M) are indicative of M components of a vector X=[Xi, ... ,XM] to be multiplied by a vector V of size Mxl.

[0067] Also, the input port 16A includes or is appropriately connected to a control port (as shown and described below with reference to Figs. 2A-2C) associated with the A / light guiding units 20 of the waveguide structure WS. The control port (26 in Fig, 2B) is configured and operable to provide AT control light signals, CSj, which are input into the M light guiding units 20. As noted above, these control signals CSj are used to affect (amplify / attenuate) the light signals ILj passing through the AT light guiding units 20, via interaction with the active medium.

[0068] The output port 18A of the optical device 12 of Fig. 1A includes a combiner assembly 24 associated with the AT output ports. The combiner assembly CA is configured for combining M light signals WLj being output from the AT light guiding units 20 of the waveguide structure 14. This provides an output light signal OL of power T being indicative of a vector T presenting the vector multiplication =XV.In the example of Fig. IB, the input port includes multiple splitting assemblies, i.e., N splitting assemblies SAi (z = 1,..., N), associated with and aligned with the input ports 16B of the N waveguide structures 14, respectively. Each i-th splitting assembly 22 is configured and operable for splitting a respective input light, ILi, of a first wavelength n of power Xi into M input light components / powers, ILij (j = 1,..., M), forming the M light signals, each of power Xy, directed into the respective M light guiding units 20 of the i-th waveguide structure 14. The M light signals of the power Xy ( / = 1,..., M) are indicative of an i-th component (z = 1,..., N) of a vector X=[XI,...,XN] to be multiplied by a matrix of size NxM.

[0069] Similarly, the input port 16A of the optical device 12 further includes or is appropriately connected to the number N of control ports (as described below with reference to Figs. 2A-2C) associated with the AT light guiding units 20 of each of the N waveguide structures. The control ports are configured and operable to provide NxM control light signals, CSij, which are input into the NxM light guiding units 20. As noted above, these control signals CSij are used to amplify the light signals ILij passing through the light guiding units 20, via interaction with the active medium.

[0070] The output port 18A of the optical device 12 of Fig. IB includes the number M of combiner assemblies 24 (CAj (j = 1,..., M)) associated with the NxM output ports. Each combiner assembly CAj is configured for combining N light signals WLij (z = 1,..., N) being output from the number N of j-th light guiding units 20 of the N waveguide structures 14. This provides an output light signal OLj (j = 1,..., M) of power Yj being indicative of a / -th component of a vector Y=[Y YMJ. The vector Y=[Y I,...,YM] presents the matrix-vector multiplication =XW.

[0071] As shown in Figs. 1A and IB, the photonic processor 100 is associated with (i.e., includes or is connected to) a detection system 30 at the output port 18A of basic block 10. The detection system 30 may include at least one optical detector 32 collecting the output light corresponding to the vector Y. The detection system includes or is connectable to a processor 34 configured and operable to translate the light signal(s) of the vector Y into a corresponding mathematical value.

[0072] Reference is made to Fig. 1C which shows schematically a photonic processor 100 including a plurality of K (K>2) basic bocks 10. To facilitate understanding, in allthe examples, the same reference numerals are used to denote functionally similar elements.

[0073] In the multi-block photonic processor shown in Fig.lC, each block includes an optical device (ODk) 12 which is generally similar to the device of Fig. IB. It should, however, be understood that the photonic processor may include multiple basic blocks configured as the optical device of Fig. 1A.

[0074] Such arrangement provides a cascaded matrix-vector multiplication (or cascaded vector multiplication), as will be described in detail further below. In such embodiments, the at least one optical detector 32 may be configured to collect the light signals OLj (j = 1,..., M) of vector Yk, and the processor 34 may be configured and operable to translate said light signals of vector Yk into a corresponding mathematical value.

[0075] Reference is made to Figs. 2A to 2C more specifically showing a non-limiting example of an z-th input port 16A associated with a waveguide structure 14 including M light guiding units 20 - four such light guiding units in this example. The input port 16A includes a splitting assembly 22 configured to split the input light ILi( n) into four light signals ILy (j = 1,..., 4) propagating through the four guiding units 20. Each of the four guiding units 20 includes active medium (AM) which, while being pumped by predetermined control light signal propagating through the / -th light guiding unit or being coupled to said light guiding unit from an adjacent pump waveguide, affects (causes amplification or attenuation of) the corresponding light signal ILy passing through said light guiding unit with a corresponding weighting factor defined by the power of the control light signal.

[0076] The active medium may be implemented, for example, by a gain medium doped by Erbium ions. The weighting factor can be controlled, at least, by defining parameters of the control light signal, gain medium, material concentrations, uniformity, and / or effective length of the amplifying channel (i.e., light guiding unit).

[0077] Further, the input port 16A includes or is coupled to a control port 26, shown in Fig. 2B (the components of the waveguide structure 14 are not shown, for clarity), configured and operable to input respective number M (four in this example) control light signals CS (2t2) into the four light guiding units 20 to thereby amplify, via interaction with the active medium, the light signals passing through the light guiding units. Fig.2C showsmore specifically the input port 16A including the splitting assembly 22 and the control port 26.

[0078] It should be noted, and is also exemplified in the figures, that the pump light (control signals) can be guided and then coupled to the erbium doped light guiding units of the waveguide structure, or, alternatively, the pump light can be directly coupled to the input port 16A. The use of a separate control port 26 might be better for manufacturing control.

[0079] In some embodiments, spectral properties of the control signals CSij are different from spectral properties of respective light signals ILij. Specifically, the light guiding units 20 may be configured to enable controlled amplification of light signals having a first wavelength Xu propagating therethrough, by transmitting into the light guiding unit 20 control light signals having a second wavelength i (e.g. a pump light signal). The i light may be transmitted into the light guiding unit 20 through the control port 26.

[0080] According to some embodiments, the active medium within (extending along) the light guiding unit 20 may include (i.e., or doped by) ions excitable by light of the spectral properties of the control signal CSij (e.g., ib light) and capable of spontaneous emission of light of the spectral properties of the light signal ILij (e.g., having a first wavelength n) propagating through the respective light guiding unit 20.

[0081] For example, during the controlled amplification in the doped guiding unit, the dopant ions are excited following absorption of ib photons, and relaxation of the excited ions is facilitated by n photons through a process known as stimulated emission. In the relaxation process, the excited ions emit additional n photons with the same phase as the facilitating Xu photons, thereby increasing the power of the already present Xu light in the light guiding unit 20. By controlling the power of the ib light (control signal) directed into the light guiding unit 20, the level of amplification of the Xu light signal may be controlled.

[0082] It should be noted that, for example, in some cases (like Erbium and probably other doping), without sufficient pump light there will be absorption of the light signal while propagating through the light guiding unit, thus resulting in attenuation. As pump light is stronger, the absorption diminishes until a specific pump level which results in noabsorption at all, but also no gain (0 gain). From there, additional pump light will generate net amplification.

[0083] It should be noted that the active medium may or may not be the same in all the light guiding units 20 of the N waveguide structures 14 of the optical device 12.

[0084] In some examples, the light signals directed into the z-th waveguide structure 14 have the wavelength Xu of about 1550 nm and the control signals being input into the light guiding units 20 of the z-th waveguide structure 14 have a second wavelength X21 of 980 nm.

[0085] It was noted above that the active medium affects (amplifies or attenuates) the light signal passing through the light guiding unit 20 with a corresponding weighting factor defined by the power of the control light signal. Thus, the power of the control light signal determines NxM weighted light signals, WLy, at NxM output ports 18B of the N waveguide structures 14.

[0086] In some embodiments, the control signals being input into the M light guiding units 20 of the same waveguide structure 14 include the control signals of different powers. This provides affecting the power of the light guiding signals (e.g., providing amplification of the light signals) with different weighting factors.

[0087] Similarly, in some embodiments, the control signals being input into the N waveguide structures 14 include the control signals of different powers. In this case, the light signals propagating through the N waveguide structures 14 are affected (amplified) with different weighting factors.

[0088] As described above, the splitting assembly 22 is configured to split the input light, ILi, of a first wavelength Xu of power Xi into M input light components / powers, ILij (j = 1,..., M). Such splitting can be designed to equally split the input light, i.e., the light signals ILij being input into the M light guiding units 20 have substantially equal power. Alternatively, or additionally, the splitting provides at least some split light components of different powers.

[0089] In some embodiments, the basic block 10 is configured such that the light signals ILij being input into the N waveguide structures 14 are of substantially the same wavelength Xu. As a result, the light signals ILij may have coherent optical condition. Inthis case, the photonic processor 34 may be further configured and operable to correct the corresponding mathematical values of vector Y=[YI,...,YM] presenting the matrix-vector multiplication =XW, to compensate for effects of coherent interactions between the light signals ILij.

[0090] Alternatively, or additionally, the light signals ILij being input into the N waveguide structures 14 include the light signals ILij of incoherent optical condition.

[0091] Reference is made to Fig. 3A showing a specific non-limiting example of an optical device 12 configured for vector-vector multiplication (lx4)x(4xl), i.e., including a single waveguide structure 14 (N = 1) which is configured to define four light guiding units 20 (M = 4). The light guiding units extend between four input ports 16B and four output ports (only output ports 18B are shown in the figure and the waveguide structures are shown transparent for clarity). In this specific not-limiting example, the Erbium -doped light units guide light signals of 1550nm, and the pump light is of 980nm wavelength.

[0092] It should be noted, although not specifically shown, that the optical devices of Fig.3A can be arranged in a cascade fashion, such that the output of the preceding device having M light guiding units is input (together with the corresponding control signals) into the successive optical device. The latter can be configured generally similar to the preceding device namely including the single waveguide structure with P light guiding units (P being equal or different than AT), or may include a stack of multiple waveguide structure, as will be described more specifically further below.

[0093] With respect to control signal, irrespective of the number N of waveguide structures being used in the basic block, it should be noted for different Erbium concentrations being used, the pump light wavelength is properly selected. Similarly, for use of other materials / material compositions (such as Bismuth or Ytterbium), a different wavelength of pump light and different other parameters (e.g., intensity) are properly selected. The optimal values of pump light parameters can be determined by using training procedures for the function / task being performed by the system.

[0094] Reference is made to Fig. 3B showing a specific non-limiting example of an optical device 12 having a size of 4x4, i.e., including an array of four waveguide structures 14 (N = 4) in which each z-th waveguide structure 14 (z = 1,...,4) is configured to definefour light guiding units 20 (M = 4). The light guiding units extend between four input ports 16B and four output ports 18B (only output ports 18B are shown in the figure and the waveguide structures are shown transparent for clarity).

[0095] The optical device 12 of Fig. 3B is configured to receive input light signals ILi (z = 1,...,4) of respective first wavelengths {Xu} of respective powers {Xi}. Here, Xi is indicative of an z-th component (z = 1,..., 4) of a vector X=[Xi,..., X4] to be multiplied by a matrix W of size 4x4. As mentioned above, the respective first wavelengths {Xu} may be of substantially the same wavelength Xu or different wavelengths.

[0096] It should be noted, although not specifically shown here, that each z-th splitting assembly 22 is configured and operable for splitting an input light, ILi, of a first wavelength Xu of power X, into four input light components / powers, ILij (j = 1,..., 4), forming the four light signals, each of power Ay, directed into the respective four light guiding units 20 of the z-th waveguide structure 14. The four light signals of the power Xjj (j = 1,..., 4) are indicative of an z-th component (z = 1,..., 4) of the vector X=[Xi,..., X4] to be multiplied by a matrix W of size 4x4.

[0097] It should be noted that splitting signal of power X into 4 signals of power Xi each yield 4 substantially identical signals with power X / 4. Hence, the gain (in the erbium portion) is configured to compensate for the splitting loss (as well as for other losses such as input coupling loss).

[0098] Each / -th light guiding unit 20 (j=l,...4) of the four waveguide structures 14 includes active medium (not shown in the figure), which, while being pumped by predetermined control light signal, CSij, affects the power / intensity (e.g., amplifies) of the light signal ILij passing through the light guiding unit 20 with a corresponding weighting factor defined by the power of the control light signal CSij. The optical device 12 thereby provides 4x4 weighted light signals, WLij, at 4x4 output ports 18B of the four waveguide structures 14.

[0099] The four output ports 18A include, respectively, four combiner assemblies 24 being configured for combining the four light signals WLij (z = 1,..., 4) being output from the four / -th light guiding units 20 of the four waveguide structures 14 to provide an output light signal OLj (j = 1,..., 4) of power Yj being indicative of a / -th component of a vector Y=[Y I,...,Y4]. The vector Y=[Y I,...,Y4] presents the matrix-vector multiplication =XW.It is noted that in the non-limiting example of Fig. 3B, the matrix IK is a square matrix, i.e., N = M. However, the present disclosure is not limited to square matrices and may be implemented to perform matrix-vector multiplication using non-square matrices W, i.e., satisfying condition N^M.

[0100] Considering planar and rigid configuration of the waveguide structures (i.e., stacked configuration), the N waveguide structures (and accordingly the N associated input ports are arranged in a spaced-apart relationship along a first axis (vertical axis in the figure). In this case, if the configuration is such that the output ports 18A of the optical devices 12 are implemented as / embedded in M planar rigid structures, e.g., formed by M glass chips, the output ports are arranged in a spaced-apart relationship along a second axis substantially perpendicular to the first axis.

[0101] The above configuration is taken into account in arranging the waveguides of the adjacent basic blocks in a multi -block photonic processor 100 configured to provide cascaded matrix-vector multiplication.

[0102] Specifically, with reference to Figs. 4A and 4B, the photonic processor 100 includes at least one pair of adjacent optically coupled basic blocks 10, e.g., ODi and OD2. The first preceding basic block, ODi, is formed by N waveguide structures, each containing M light guiding units, and provides the vector-matrix multiplication Yi(lxM)= X(lxN)xWi(NxM). The second successive basic block, OD2, is formed by M waveguide structures, each containing P light guiding units, provides vector-matrix multiplication Y2(lxP)= YI(1XM)XW2(MXP). In the non-limiting example of Figs. 4A and 4B, each one of the basic blocks ODi and OD2, represents a square matrix of size 4x4.

[0103] As shown, the N (N = 4) waveguide structures of the first basic block ODi are arranged in a spaced apart substantially parallel relationship along a first axis, and the M waveguide structures (M = 4) of the second basic block OD2 are arranged in a spaced apart substantially parallel relationship along a second axis forming 90 degrees tilt with respect to the first axis. This tilt is required if a compact architecture of the photonic processor 100 is preferred. In some embodiments, the 90 degrees tilt may be avoided if the optical devices or at least their input and output ports are configured as flexible fiberbased structures.Thus, in some embodiments, where the photonic processor 100 configured as described above includes multiple pairs of the adjacent optically coupled basic blocks, the odd basic blocks may be configured as the first preceding basic block (e.g., ODi of Fig. 4B) and even basic blocks may be configured as the second successive basic block (e g., OD2 of Fig. 4B).

[0104] The optical device of the present disclosure can be implemented by a number L (L> 1) of glass chips, where the number of glass chips defines the M light guiding units. This is exemplified in Fig. 5 showing an optical device 12 where four glass chips, GSi, GS2, GS3, and GS4 define 16 light guiding units, defining thereby an 8x16 matrix.

[0105] The principles of the present disclosure can advantageously be used to implement a processor panel for use in a photonic computing system. Such processor panel can form a module enabling its housing in a network rack.

Claims

CLAIMS:

1. A basic block for use in a photonic processor for performing light signals multiplication, the basic block comprises an optical device, wherein:the optical device comprises a number N (N>1) of waveguide structures between input port and output port of the optical device, each i-th waveguide structure (z = 1,..., N) being configured to define M light guiding units (M>2) extending between M input ports and M output ports of the waveguide structure, respectively, to enable directional propagation of M light signals through the M light guiding units, each j -th light guiding unit (j=l,...M) comprising active medium, which, while being pumped by a predetermined control light signal propagating through the j -th light guiding unit, affects intensity of a j-th light signal, interacting with said medium while passing through said j-th light guiding unit, with a corresponding weighting factor defined at least by power of the control light signal, said optical device thereby providing NxM weighted light signals at NxM output ports of the N waveguide structures;the input port of the optical device comprises the number N of splitting assemblies associated and aligned with the inputs of the N waveguide structures, respectively, each i-th splitting assembly being configured and operable for splitting input light of a first wavelength Xu of power Xi into M input light powers forming the M light signals, each of power Xy (j=l,...M), directed into the respective M light guiding units of the i-th waveguide structure, said M light signals of the power Xy being indicative of an i-th component of a vector X=[Xi,..., XN] to be multiplied by a matrix or vector of size NxM,' the output port of the optical device comprises a predetermined number of combiner assemblies associated with the NxM output ports and configured to provide an output light signal of power Yj being indicative of a / -th component of a vector Y=[YI,...,YM], said vector Y=[YI,...,YM] presenting a matrix-vector or vector-vector multiplication of the input light signal.

2. The basic block of claim 1, wherein the input port of the optical device further comprises the number N of control ports associated with the A / light guiding units of each of the N waveguide structures, and being configured and operable to input NxM control light signals into said NxM light guiding units to thereby amplify the light signals passing through said light guiding units.

3. The basic block of claim 1, comprising a single waveguide structure (N=l), the output port comprising a single combiner, said basic block performing vector-vector multiplication V of the input light signal X producing the output signal Y=XV.

4. The basic block of claim 1 , comprising an array (stack) of said N (N>2) waveguide structures, the output port of the optical device comprising the number M of combiner assemblies associated with the NxM output ports, each combiner assembly being configured for combining TV light signals being output from the TVj -th light guiding units of the N waveguide structures to provide an output light signal of power I) being indicative of a / -th component of a vector Y=[YI,...,YM], said vector Y=[YI,...,YM] presenting the matrix-vector multiplication =XW.

5. The basic block of claim 1, wherein spectral properties of the control light signals are different from spectral properties of respective light signals.

6. The basic block of claim 5, wherein the active medium comprises ions which are excitable by light of the spectral properties of the control light signal and are capable of spontaneously emitting light of the spectral properties of the light signal propagating through the respective light guiding unit.

7. The basic block of claim 4, wherein:the active medium comprises ions which are excitable by light of the spectral properties of the control light signal and are capable of spontaneously emitting light of the spectral properties of the light signal propagating through the respective light guiding unit; andthe light guiding units of the N waveguide structures comprise the same active medium.

8. The basic block of claim 4, wherein:the active medium comprises ions which are excitable by light of the spectral properties of the control light signal and are capable of spontaneously emitting light of the spectral properties of the light signal propagating through the respective light guiding unit; andthe light guiding units of the N waveguide structures comprise the light guiding units containing different active media.

9. The basic block of claim 5, wherein said light signals of the i-th waveguide structure have the wavelength Xu of about 1550 nm, and said control light signals being input into said light guiding units of the i-th waveguide structure are of a second wavelength X21 of 980 nm.

10. The basic block of claim 1, wherein the control light signals being input into the M light guiding units of the waveguide structure comprise the control light signals of different powers, thereby providing amplification or attenuation of the light signals with different weighting factors.

11. The basic block of claim 4, wherein the control light signals being input into the N waveguide structures comprise the control light signals of different powers, to thereby amplify the light signals propagating through the N waveguide structures with different weighting factors.

12. The basic block of claim 1, wherein the light signals being input into the AT light guiding units of the i-th waveguide structure have substantially equal power.

13. The basic block of claim 1, wherein the light signals being input into the AT light guiding units of the i-th waveguide structure comprise the light signals of different powers.

14. The basic block of claim 4, wherein the light signals being input into the N waveguide structures are of substantially the same wavelength An.

15. The basic block of claim 14, wherein the light signals being input into the N waveguide structures have coherent optical condition.

16. The basic block of claim 14, wherein the light signals being input into the N waveguide structures comprise the light signals of incoherent optical condition.

17. The basic block of claim 4, wherein N=M.

18. The basic block of claim 4, wherein N^M.

19. The basic block of claim 1, wherein the waveguide structure of the optical device is formed by a number L (L>1) of glass chips, said number of glass chips defining said AT light guiding units.

20. The basic block of claim 1, wherein said optical device comprises at least one flexible fiber-based structure.

21. A basic block for use in a photonic processor for performing matrix-vector multiplication, the basic block being configured as an optical device, wherein:the optical device comprises a stack formed by an array of N (N>2) waveguide structures between an input port and an output port of the optical device, wherein each i-th waveguide structure (z = 1,..., N) is configured to define AT light guiding units (M>2) extending between M input ports and M output ports of the waveguide structure, respectively, to enable directional propagation of M light signals through the M light guiding units, and wherein each j-th light guiding unit (j=l,...M) of the N waveguide structures comprises active medium, which, while being pumped by predetermined control light signal propagating through the j-th light guiding unit, affects intensity of a light signal, interacting with said medium while passing through said light guiding unit, with a corresponding weighting factor defined at least by power of the control light signal, said optical device thereby providing NxM weighted light signals at NxM output ports of the N waveguide structures;the input port of the optical device comprises the number N of splitting assemblies associated and aligned with the input ports of the N waveguide structures, respectively, each i-th splitting assembly being configured and operable for splitting an input light of a first wavelength n of power Xi into AT input light powers forming the M light signals, each of power j(j=l,...M), directed into the respective AT light guiding units of the i-th waveguide structure, said M light signals of the power A; / being indicative of an i-th component of a vector X=[XI,...,XN] to be multiplied by a matrix W of size NxM,' the output port of the optical device comprises the number M of combiner assemblies associated with the NxM output ports, each combiner assembly being configured for combining N light signals being output from the N j-th light guiding units of the N waveguide structures to provide an output light signal of power Yj being indicative of a / -th component of a vector Y=[YI,...,YM], said vector Y=[YI,...,YM] presenting the matrix-vector multiplication =XW.

22. A basic block for use in a photonic processor for performing matrix -vector or vector-vector multiplication, the basic block being configured as an optical device, comprising:a glass chip comprising a number N (N>1) of waveguide structures between an input port and an output port of the optical device, each i-th waveguide structure (i = 1,..., N) being configured to define M light guiding units (M>2) extending between M input ports and M output ports of the waveguide structure, respectively, to enable directional propagation of M light signals through the M light guiding units, each j -th light guiding unit (j=l,...M) comprising active medium, which, while being pumped by a predetermined control light signal propagating through the j -th light guiding unit, affects intensity of a j-th light signal, interacting with said medium while passing through said light guiding unit, with a corresponding weighting factor defined at least by power of the control light signal, said optical device thereby providing NxM weighted light signals at NxM output ports of the N waveguide structures;the input port of the optical device comprises the number N of splitting assemblies associated and aligned with the input ports of the N waveguide structures, respectively, each i-th splitting assembly being configured and operable for splitting an input light of a first wavelength Xu of power Xi into M input light powers forming the M light signals, each of power Xij (j=l, ...M), directed into the respective M light guiding units of the i-th waveguide structure, said M light signals of the power Xij being indicative of an i-th component of a vector X=[X1,...,XN] to be multiplied by a matrix or vector of size NxM;the output port of the optical device comprises a predetermined number of combiner assemblies associated with the NxM output ports, to provide an output light signal of power Yj being indicative of a j-th component of a vector Y=[Y1,...,YM], said vector Y=[Y1,...,YM] presenting the matrix-vector or vector-vector multiplication of the input light signal.

23. A photonic processor comprising a number K (K>1) of basic bocks, each basic block being configured according to claim 1, providing cascaded matrix -vector or vectorvector multiplication.

24. The photonic processor of claim 23, comprising a detection system at the output port of the k-th basic block.

25. The photonic processor of claim 24, wherein the detection system comprises at least one optical detector collecting the light signals of vector YR, and a processor configured and operable to translate said light signals of vector YR into a corresponding mathematical value.

26. The photonic processor of claim 25, wherein said processor of the detection system is further configured and operable to correct said corresponding mathematical value to compensate effects of coherent interactions between the light signals.

27. The photonic processor of claim 23, comprising at least one pair of optically coupled basic blocks, the pair of the optically coupled basic blocks comprising:a first preceding basic block, which is formed by N waveguide structures, each containing M light guiding units, and which provides the vector-matrix or vector-vector multiplication; anda second successive basic block which comprises one or more waveguide structures, each of said one or more waveguide structures being formed by the light guiding units and providing vector-matrix or vector-vector multiplication.

28. The photonic processor of claim 23, comprising at least one pair of optically coupled basic blocks, the pair of the optically coupled basic blocks comprising:a first preceding basic block, which is formed by N waveguide structures, each containing M light guiding units, and which provides the vector-matrix or vector-vector multiplication Yi(lxM)= X(lxN)xWi(NxM); anda second successive basic block, which is formed by M waveguide structures, each containing P light guiding units, and which provides vector-matrix multiplication Y2(1XP)= YI(1XM)XW2(MXP).

29. The photonic processor of claim 28, wherein said N waveguide structures of the first basic block are arranged in a spaced-apart substantially parallel relationship along a first axis, and said AT waveguide structures of the second basic block are arranged in a spaced-apart substantially parallel relationship along a second axis forming a substantially 90 degrees tilt with respect to the first axis.

30. The photonic processor of claim 28, comprising a number K / 2 of said pairs of the optically coupled basic blocks, wherein odd basic blocks are configured as said first preceding basic block and even basic blocks are configured as said second successive basic block.

31. A neural network comprising one or more neuron units, wherein each neuron unit comprises the photonic processor of claim 23.

Images