Photon tensor calculation method, photon tensor core, computer device and storage medium

Abstract

The application relates to a photonic tensor calculation method, a photonic tensor core, a computer device and a storage medium. The method comprises the following steps: generating a multi-wavelength optical signal, modulating optical signals of each wavelength in the multi-wavelength optical signal through a first loopback arrayed waveguide grating, and multiplexing to obtain a first optical signal with multiple wavelengths; splitting each first optical signal to obtain a plurality of first optical signal copies, and modulating optical signals of each wavelength in a single first optical signal copy through a plurality of second loopback arrayed waveguide gratings, and multiplexing to obtain a second optical signal; and photoelectrically converting each second optical signal to obtain a tensor calculation result. The method can improve the photonic tensor calculation accuracy and scalability.

Description

Technical Field

[0001] This application relates to the field of photon tensor computation technology, and in particular to a photon tensor computation method, a photon tensor kernel, a computer device, and a storage medium. Background Technology

[0002] With the development of artificial intelligence technology, photon tensor computing has become an important solution for accelerating AI computation.

[0003] Traditional techniques typically employ Mach-Zehnder interferometer (MZI) arrays based on coherent light and microring resonator arrays based on incoherent light to perform photon tensor calculations. However, due to the large thermo-optic coefficient of silicon, changes in ambient temperature can easily cause phase fluctuations in MZI devices, leading to phase errors in photon tensor calculations and resulting in problems such as wavelength drift and deterioration of isolation. Furthermore, the microring structure is limited by the radius of the microring, restricting the number of usable channels and significantly impacting scalability and accuracy.

[0004] It is evident that the existing architecture suffers from insufficient computational accuracy and limited scalability in practical applications, making it difficult to meet the computational efficiency and accuracy requirements of large-scale neural networks. Summary of the Invention

[0005] Therefore, it is necessary to provide a photon tensor calculation method, photon tensor kernel, computer equipment, and storage medium that can improve the accuracy and scalability of photon tensor calculation, in response to the above-mentioned technical problems.

[0006] In a first aspect, this application provides a method for calculating photon tensors, the method comprising:

[0007] A multi-wavelength optical signal is generated, and each wavelength of the optical signal is modulated separately through a first loop array waveguide grating, and then multiplexed to obtain a multi-channel multi-wavelength first optical signal; wherein each channel of the first optical signal is isolated from each other in the spectrum; a first modulation module is coupled on the loop path of the first loop array waveguide grating, which is used to modulate the optical signal of each wavelength according to the input data to be processed;

[0008] Each of the first optical signals is split to obtain multiple copies of the first optical signal. The optical signals of each wavelength in each copy of the first optical signal are modulated by multiple second loop array waveguide gratings and multiplexed to obtain a single second optical signal. A second modulation module is coupled on the loop path of the second loop array waveguide grating to modulate the optical signals of each wavelength according to the input coefficient data.

[0009] The second optical signal is converted into photoelectric signal to obtain tensor calculation results.

[0010] In one embodiment, the multiplexing to obtain a first optical signal with multiple wavelengths includes:

[0011] The modulated optical signals of each wavelength are grouped according to a preset number of channels, with each group corresponding to one of the first optical signals; the frequency domain spacing between adjacent wavelengths in each of the first optical signals is the product of the frequency domain spacing between adjacent wavelengths in the multi-wavelength optical signals and the preset number of channels.

[0012] In one embodiment, the method further includes, prior to modulating the optical signals of each wavelength in a single copy of the first optical signal, the following steps:

[0013] Spectral shaping is performed on a single copy of the first optical signal so that each wavelength component in the copy of the first optical signal falls within a preset flat-top passband.

[0014] In one embodiment, the data to be processed includes neural network activation values; the coefficient data includes convolution kernel weights; and the tensor calculation result includes the output vector of matrix-vector multiplication.

[0015] Secondly, this application provides a photonic tensor kernel, the photonic tensor kernel comprising:

[0016] The light source unit is used to generate multi-wavelength light signals;

[0017] The parallel data loading optical engine includes a first loop-back array waveguide grating, which is used to modulate the optical signals of each wavelength in the multi-wavelength optical signals and then multiplex them to obtain multiple multi-wavelength first optical signals; wherein each first optical signal is isolated from each other in the spectrum; a first modulation module is coupled on the loop-back path of the first loop-back array waveguide grating, which is used to modulate the optical signals of each wavelength according to the input data to be processed;

[0018] The beam splitting unit is used to split each of the first optical signals to obtain multiple copies of the first optical signal;

[0019] The photonic dot product engine includes a second loop array waveguide grating, which is used to modulate the optical signals of each wavelength in a single copy of the first optical signal and multiplex them to obtain a single second optical signal; wherein, a second modulation module is coupled on the loop path of the second loop array waveguide grating, which is used to modulate the optical signals of each wavelength according to the input coefficient data;

[0020] The photoelectric detection module is used to perform photoelectric conversion on each of the second optical signals to obtain tensor calculation results.

[0021] In one embodiment, the first loopback array waveguide grating includes a first star coupler, a second star coupler, an array waveguide connecting the first star coupler and the second star coupler, and a first loopback path connecting the output side of the second star coupler and the input side of the first star coupler.

[0022] The first modulation module is coupled to the first loopback path and is used to modulate the optical signal of each wavelength according to the input data to be processed, and send the modulated optical signal of each wavelength back to the first star coupler.

[0023] In one embodiment, the second loopback array waveguide grating includes a third star coupler, a fourth star coupler, an array waveguide connecting the third star coupler and the fourth star coupler, and a second loopback path connecting the output side of the fourth star coupler and the input side of the third star coupler;

[0024] The second modulation module is coupled to the second loop path and is used to modulate the optical signal of each wavelength according to the input coefficient data, and send the modulated optical signal of each wavelength back to the third star coupler.

[0025] In one embodiment, the second loop array waveguide grating further includes a flat-top spectral response input unit;

[0026] The third star coupler acquires a copy of the first optical signal after spectral shaping through the flat-top spectral response input unit; each wavelength component in the first optical signal copy falls within a preset flat-top passband.

[0027] In one embodiment, the flat-top spectral response input unit employs a Mach-Zehnder interferometer or a multimode interferometer.

[0028] In one embodiment, the multiplexing to obtain a first optical signal with multiple wavelengths includes:

[0029] The modulated optical signals of each wavelength are grouped according to a preset number of channels, with each group corresponding to one of the first optical signals; the frequency domain spacing between adjacent wavelengths in each of the first optical signals is the product of the frequency domain spacing between adjacent wavelengths in the multi-wavelength optical signals and the preset number of channels.

[0030] In one embodiment, the beam splitting unit includes a plurality of beam splitters; the number of beam splitters corresponds to the number of preset channels; each beam splitter is used to split the corresponding first optical signal to obtain a plurality of first optical signal copies; the number of first optical signal copies is equal to the number of photon multiplication engines.

[0031] In one embodiment, both the first modulation module and the second modulation module adopt one of the following: Mach-Zehnder interference structure, resonant micro-ring structure, and electro-absorption modulation structure.

[0032] In one embodiment, the light source unit includes at least one of a superradiative diode light source, a multi-wavelength laser light source, and an optical frequency comb light source.

[0033] Thirdly, this application provides a computer device including a memory and a processor, wherein the memory stores a computer program and the processor includes a photon tensor kernel as described above.

[0034] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.

[0035] The aforementioned photonic tensor calculation method, photonic tensor kernel, computer equipment, and storage medium, by generating multi-wavelength optical signals and using a first loop-back arrayed waveguide grating for spectral isolation and modulation, can achieve high-density signal processing based on wavelength division multiplexing. Employing a two-stage modulation architecture—a first modulation module loading the data to be processed and a second modulation module loading the coefficient data—decouples the data from the weights. Combined with a parallel processing mechanism of beam splitting and multiple second loop-back arrayed waveguide gratings, it effectively avoids the data transfer bottleneck in traditional electronic computing. By summing the calculation results through photoelectric conversion, it transforms data domain operations into optical domain operations. This allows for the multiplication and summation of input data and corresponding weight data while significantly increasing the number of parallel computing channels and the signal-to-noise ratio. Furthermore, it maintains overall wavelength consistency despite changes in ambient temperature and manufacturing process deviations, thus avoiding the impact of these factors on the accuracy of photonic tensor calculations. Moreover, it allows for flexible configuration of arrayed waveguide gratings of different wavelengths based on the data to be processed and the coefficient data, achieving high scalability and ultimately improving both the accuracy and scalability of photonic tensor calculations. Attached Figure Description

[0036] Figure 1 This is a diagram illustrating the application environment of the photon tensor calculation method in one embodiment;

[0037] Figure 2 This is a flowchart illustrating a photon tensor calculation method in one embodiment;

[0038] Figure 3 Here is a block diagram of the photon tensor kernel in one embodiment;

[0039] Figure 4 This is a schematic diagram of a novel high-precision scalable photonic tensor kernel for accelerating AI convolutional computation in one embodiment.

[0040] Figure 5 This is a schematic diagram of the structure of a parallel data loading optical engine in one embodiment;

[0041] Figure 6 This is a schematic diagram of the photon dot product engine in one embodiment;

[0042] Figure 7 This is a schematic diagram of the waveguide position distribution of a 1×K loop AWG in one embodiment;

[0043] Figure 8 This is a schematic diagram of the structural transformation of a loop AWG in one embodiment;

[0044] Figure 9 This is a schematic diagram of a typical spectral response of a 1×K loop AWG in one embodiment;

[0045] Figure 10 This is a schematic diagram of the waveguide position distribution of a 1×K loop AWG in one embodiment;

[0046] Figure 11 This is a schematic diagram of the structural transformation of a 1×K loop AWG after adding a modulator in one embodiment;

[0047] Figure 12 This is a schematic diagram of the typical spectral response of a 1×K loop AWG after adding a modulator in one embodiment;

[0048] Figure 13 This is a schematic diagram of the principle of a 1×1 loop AWG in one embodiment;

[0049] Figure 14 This is a schematic diagram of the structural transformation of a 1×1 loop AWG in one embodiment;

[0050] Figure 15 This is a schematic diagram of the typical spectral response of a 1×1 loop AWG in one embodiment;

[0051] Figure 16 This is a schematic diagram of the waveguide position distribution of a 1×1 loop AWG after adding a modulator in one embodiment;

[0052] Figure 17 This is a schematic diagram of the structural transformation of a 1×1 loopback AWG after adding a modulator in one embodiment;

[0053] Figure 18 This is a schematic diagram of the typical spectral response of a 1×1 loop AWG after adding a modulator in one embodiment;

[0054] Figure 19 This is a schematic diagram of a multi-wavelength laser source based on the combination of multiple lasers in one embodiment;

[0055] Figure 20 This is a schematic diagram of a frequency-sparse optical source based on a pump laser and a silicon nitride microring in one embodiment;

[0056] Figure 21 This is a schematic diagram of a broadband light source based on a superluminescent diode and its typical light source spectrum in one embodiment;

[0057] Figure 22 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0059] The photon tensor calculation method provided in this application embodiment can be applied to, for example, Figure 1 The application environment shown. The photonic tensor kernel comprises: a light source unit 1 generating multi-wavelength optical signals; a parallel data loading optical engine 2 modulating each wavelength of the multi-wavelength optical signals through a first loop array waveguide grating 21 and multiplexing them to obtain multiple multi-wavelength first optical signals; wherein each first optical signal is spectrally isolated from the others; a first modulation module coupled to the loop path of the first loop array waveguide grating 21 modulates each wavelength of the optical signals according to the input data to be processed; a beam splitting unit 3 splits each first optical signal to obtain multiple copies of the first optical signals; a photonic dot product engine 4 modulates each wavelength of the optical signals in a single copy of the first optical signal through multiple second loop array waveguide gratings 41 and multiplexing them to obtain a single second optical signal; wherein a second modulation module coupled to the loop path of the second loop array waveguide grating 41 modulates each wavelength of the optical signals according to the input coefficient data; and a photoelectric detection module 5 performs photoelectric conversion on each second optical signal to obtain tensor calculation results.

[0060] In one embodiment, such as Figure 2 As shown, a photon tensor calculation method is provided. Taking the application of this method in a photon tensor calculation scenario as an example, the photon tensor calculation method includes:

[0061] Step S100: Generate multi-wavelength optical signals, and modulate each wavelength of the multi-wavelength optical signals separately through the first loop array waveguide grating, and then multiplex them to obtain a multi-channel multi-wavelength first optical signal.

[0062] In this context, the multi-wavelength optical signal can be a set of optical carriers containing multiple wavelength components, thus serving as a physical carrier for parallel data transmission. Each first optical signal is spectrally isolated from the others. Correspondingly, generating the multi-wavelength optical signal can be achieved by using a light source device to generate continuous or pulsed light waves containing multiple specific wavelengths, thus generating the multi-wavelength optical signal as a computational carrier. For example, the light source device can be implemented using one or more methods, such as a multi-wavelength laser array, a single laser combined with an optical frequency comb generator, etc.

[0063] The first loop-back arrayed waveguide grating can be an arrayed waveguide grating structure with a loop-back reflection path, used for wavelength demultiplexing and multiplexing. In this embodiment, the first loop-back arrayed waveguide grating can utilize the optical path difference of the waveguide array to generate interference, guiding different wavelengths to a specific port and returning via the loop-back path, thereby enabling at least two signal transmissions and / or processing within the same arrayed waveguide grating. For example, the first loop-back arrayed waveguide grating can be one or more of the following, depending on the substrate material: silicon-based arrayed waveguide grating, indium phosphide-based arrayed waveguide grating, silicon dioxide-based arrayed waveguide grating, etc.

[0064] The optical signals of each wavelength in the multi-wavelength optical signal are modulated separately by a first loop-loop array waveguide grating. For example, the grating can separate different wavelengths into independent paths and modulate them independently for each wavelength, thereby completing the data loading. Furthermore, the loop-loop path can be realized by a loop-loop waveguide or by optical signal transmission materials such as a mirror; this embodiment is not limited to this.

[0065] In this embodiment, a first modulation module is coupled along the loop path of the first loop array waveguide grating, which is used to modulate the optical signal of each wavelength according to the input data to be processed.

[0066] The first modulation module can be an electro-optic modulation unit located on the first loop path. In this embodiment, the first modulation module can modulate the optical signal by receiving data to be processed in the form of an electrical signal and changing the physical properties of the passing optical signal. For example, depending on the modulation mechanism, the first modulation module can include one or more of the following: an electroabsorption modulator, a Mach-Zehnder modulator, and a micro-ring modulator.

[0067] The data to be processed can be a set of input values ​​that need to participate in tensor operations. In an exemplary embodiment, the data to be processed can be the activation values ​​of neurons in the forward propagation of a neural network. Furthermore, depending on the data precision, the data to be processed can include floating-point data, fixed-point data, binary data, etc.

[0068] Furthermore, the first modulation module modulates the optical signal of each wavelength according to the input data to be processed. This can be done by converting the data to be processed into control voltage or current, changing the amplitude or phase of the optical signal, thereby completing the mapping of the input data to the optical domain.

[0069] In this embodiment, the first optical signal can be a multi-channel optical signal that has been modulated at the first level and whose spectrum is isolated from each other. Each group of signals carries N data and is transmitted through the corresponding wavelength, so that it can be used to carry the data information to be processed, and is further divided into multiple signal copies with equal power by the beam splitting module.

[0070] In step S200, each first optical signal is split into multiple copies of the first optical signal, and each wavelength of the optical signal in a single copy of the first optical signal is modulated by multiple second loop array waveguide gratings and multiplexed to obtain a single second optical signal.

[0071] The process of splitting each first optical signal into multiple copies can be achieved by using a beam splitting module to split each first optical signal into multiple copies, thus obtaining multiple copies of the first optical signal corresponding to each of the multiple first optical signals. For example, this can be achieved by using a Y-type branched waveguide for two-way beam splitting, or by using a multimode interference coupler for multi-beam splitting; this embodiment does not limit the specific implementation.

[0072] The first optical signal copy can be a replicated signal generated by power distribution of the first optical signal. By distributing the energy of one signal to multiple channels, it can support parallel multiplication operations of the same input data with multiple different coefficients. Each copy enters an independent second loop array waveguide grating.

[0073] In this embodiment, the second loop-loop waveguide grating may have the same structure as the first loop-loop waveguide grating. Modulating each wavelength of the optical signal in a single copy of the first optical signal using multiple second loop-loop waveguide gratings means that each second loop-loop waveguide grating processes one copy of the first optical signal and modulates each wavelength of the optical signal in that copy separately. Furthermore, the modulation method may be the same as or similar to the modulation method of the first loop-loop waveguide grating; this embodiment does not limit this.

[0074] In this embodiment, a second modulation module is coupled on the loop path of the second loop array waveguide grating, which is used to modulate the optical signal of each wavelength according to the input coefficient data.

[0075] The second modulation module is an electro-optic modulation unit located on the second loop path. It receives coefficient data in the form of electrical signals and performs secondary modulation on the passing copy of the first optical signal. For example, the second modulation module may use the same structure as the first modulation module. The coefficient data may be a set of weight parameters used in tensor calculations. For example, the coefficient data may be the connection strength and transformation matrix between neural network layers.

[0076] The second modulation module modulates the optical signal of each wavelength according to the input coefficient data. For example, it can convert the coefficient data into a control signal and perform secondary attribute changes on the copy of the first optical signal, thereby completing the mapping of weight data to the optical domain and realizing multiplication operations in the optical domain.

[0077] The second optical signal can be an optical product signal that simultaneously carries the data to be processed and the coefficient data. It realizes the multiplication of the data to be processed and the weight data carried by the wavelength, representing the multiplication result of the input data and the weights. It can be accumulated and output to the detector to complete the summation. For example, the second optical signal can be one or more of the following, depending on the signal state: coherent superimposed optical signal, incoherent superimposed optical signal, polarization multiplexed optical signal, etc.

[0078] Step S300: Perform photoelectric conversion on each second optical signal to obtain tensor calculation results.

[0079] The tensor calculation result can be the sum of electrical signals obtained after photoelectric conversion, which serves as the final operation value of the photonic computing unit. Furthermore, the photoelectric conversion of each second optical signal to obtain the tensor calculation result can be achieved using a photodetector to receive the second optical signals and perform photoelectric conversion on each of them. In some exemplary embodiments, single-ended photodiode detection, differential detection using balanced photodetectors, or other methods can be used to complete the conversion from optical domain multiplication to electrical domain addition, outputting the final calculated value.

[0080] This embodiment provides a photonic tensor calculation method that generates multi-wavelength optical signals and uses a first loop-back arrayed waveguide grating for spectral isolation and modulation. This enables high-density signal processing based on wavelength division multiplexing (WDM). A two-stage modulation architecture is employed, with a first modulation module loading the data to be processed and a second modulation module loading the coefficient data. This decouples the data from the weights. Combined with beam splitting and a parallel processing mechanism using multiple second loop-back arrayed waveguide gratings, this effectively avoids the data transfer bottleneck in traditional electronic computing. The calculation results are aggregated through photoelectric conversion, transforming data domain operations into optical domain operations. This allows for the multiplication and summation of input data and corresponding weight data while significantly increasing the number of parallel computing channels and the signal-to-noise ratio. Furthermore, it maintains overall wavelength consistency despite changes in ambient temperature and manufacturing deviations, thus avoiding the impact of these factors on the accuracy of photonic tensor calculations. Moreover, different wavelength arrayed waveguide gratings can be flexibly configured according to the data to be processed and the coefficient data, achieving high scalability and improving both the accuracy and scalability of photonic tensor calculations.

[0081] The various optional embodiments are further described in detail below.

[0082] In one embodiment, multiplexing to obtain multiple multi-wavelength first optical signals includes:

[0083] The modulated optical signals of each wavelength are grouped according to the preset number of channels, and each group corresponds to one first optical signal; the interval between adjacent wavelengths in the frequency domain of each first optical signal is the product of the interval between adjacent wavelengths in the multi-wavelength optical signal and the preset number of channels.

[0084] The preset number of channels can be a configuration parameter used to specify the signal grouping size. The preset number of channels can be used as a grouping basis to determine the number of wavelengths contained in each first optical signal. For example, the value of the preset number of channels can be equal to the total number of the first loopback output channels.

[0085] Each group corresponds to one first optical signal, and a group can be a logical set composed of several modulated wavelength signals. In this embodiment, each group of optical signals can correspond to one physical transmission channel. Furthermore, the number of wavelengths in each group and the total number of channels together determine the total wavelength scale of the first optical signal.

[0086] Understandably, the total number of wavelengths may not be exactly equal to the product of the preset number of channels and the number of wavelengths in the first optical signal. Grouping the modulated optical signals of each wavelength according to the preset number of channels can be achieved by cutting a continuous sequence of modulated wavelength signals into multiple subsets containing the same or nearly the same number of wavelengths based on the preset number of channels parameter.

[0087] Furthermore, the modulated optical signals of each wavelength can be grouped according to a preset number of channels. This can be achieved by spatial grouping using the inherent dispersion characteristics of the arrayed waveguide grating, thereby establishing a structured mapping relationship between the signal channels.

[0088] In this embodiment, the frequency domain spacing of adjacent wavelengths can refer to the distance between two adjacent wavelength components within the same first optical signal in the frequency dimension. It is understood that the frequency domain spacing of adjacent wavelengths can prevent overlap between different wavelength components within the same signal, avoiding signal interference between different transmission channels. In a specific embodiment, the frequency domain spacing of adjacent wavelengths can adopt equal spacing, non-equal spacing, or gradually varying spacing distributions, etc.

[0089] The spacing between adjacent wavelengths in a multi-wavelength optical signal can be the inherent frequency distance between the fundamental adjacent wavelength components in the generated original multi-wavelength signal. In this embodiment, the spacing between adjacent wavelengths in the multi-wavelength optical signal can reflect the spectral characteristics of the light source itself and affect the isolation of the multiplexed signal. In an exemplary embodiment, the spacing between adjacent wavelengths in the multi-wavelength optical signal can be the inherent spacing of the laser array, the spacing of the optical frequency comb, or the spacing of the filter channels. This spacing can be determined according to the type of light source device, and this embodiment does not limit it.

[0090] This embodiment provides a photonic tensor calculation method that establishes a structured mapping relationship between signal channels by grouping modulated optical signals according to a preset number of channels. This ensures clear logical isolation between multiple signals. By setting the frequency domain interval between adjacent wavelengths in each signal to be the product of the original interval and the number of channels, the spectral guard band between channels is artificially expanded, effectively avoiding crosstalk caused by spectral overlap between different channels and solving the problem of isolation degradation. Furthermore, this photonic tensor calculation method, through a frequency interval expansion-based multiplexing approach, allows for linear expansion of calculation channels by increasing the preset number of channels without increasing the complexity of physical devices. This overcomes the channel number bottleneck limited by the radius of the micro-ring structure, thereby improving the system's parallel processing capability while ensuring signal purity. Therefore, it achieves the effect of improving the accuracy and scalability of photonic tensor calculation.

[0091] In one embodiment, the method further includes modulating the optical signals of each wavelength in a single first optical signal copy separately, prior to:

[0092] The spectrum of a single copy of the first optical signal is shaped so that each wavelength component in the copy of the first optical signal falls within a preset flat-top passband.

[0093] Among them, spectrum shaping can be used to shape each wavelength component in the first optical signal copy so that the optical signal can exhibit a flat-top characteristic in the spectrum display.

[0094] Furthermore, spectral shaping can be achieved by processing a single copy of the first optical signal using a loop-arrayed waveguide grating input unit. The loop-arrayed waveguide grating input unit can be positioned within the optical signal transmission path, serving as a loop-arrayed waveguide grating structure with a flat-top spectral response, to adjust the spectral power distribution characteristics of the optical signal. Furthermore, the loop-arrayed waveguide grating input unit can be used to eliminate spectral unevenness of the light source, ensuring the consistency of optical power across wavelength channels before entering the modulation stage. For example, the loop-arrayed waveguide grating input unit can be an input unit based on a Mach-Zehnder interferometer structure, or a single-mode / multimode interference coupler structure can be used to attenuate or enhance the amplitude of each wavelength component of the input optical signal.

[0095] The preset flat-top passband can be a pre-defined spectral window region with flat transmission response characteristics formed by the input units of the loop-loop arrayed waveguide grating. This provides a uniform transmission environment for wavelength components and avoids amplitude errors caused by misalignment of the channel spectra between the first and second loop-loop arrayed waveguide gratings due to manufacturing process variations. Furthermore, the passband edge frequencies and top flatness can be determined through the structural design of the input units in the arrayed waveguide grating, forming a specific spectral transmission curve.

[0096] It is understandable that by using the input unit of the loop array waveguide grating with a flat-top spectral response to receive the copy of the first optical signal after beam splitting, and adjusting the relative power of each wavelength component, the amplitude error caused by the misalignment of the channel spectrum between the first loop array waveguide grating and the second loop array waveguide grating due to the deviation in the manufacturing process can be corrected, thereby optimizing the signal-to-noise ratio of the signal and reducing the nonlinear distortion in the subsequent photoelectric conversion process.

[0097] This embodiment provides a photonic tensor calculation method that performs spectral shaping on a single first optical signal replica to ensure that each wavelength component in the first optical signal replica falls within a preset flat-top passband. By preprocessing the first optical signal replica using the input unit of a loop-loop waveguide grating with a flat-top spectral response, the method forces each wavelength component to fall within the preset flat-top passband. This eliminates amplitude errors caused by misalignment of the channel spectra between the first and second loop-loop waveguide gratings due to manufacturing process deviations. Simultaneously, by employing a two-stage modulation architecture that loads the data to be processed and the coefficient data using a modulator in the loop-loop waveguide, the data and weights are decoupled. Combined with the parallel processing mechanism of beam splitting and multiple second loop-loop waveguide gratings, this method effectively avoids the data transport bottleneck in traditional electronic computing and the phase error of traditional MZI arrays. While ensuring signal isolation, phase stability, and spectral consistency, it significantly increases the number of parallel computing channels and the signal-to-noise ratio, thereby achieving the technical effect of improving the accuracy and scalability of photonic tensor calculation.

[0098] In one embodiment, the data to be processed includes neural network activation values; coefficient data includes convolution kernel weights; and tensor computation results include the output vector of matrix-vector multiplication.

[0099] In this scheme, the neural network activation values ​​can be feature data passed between neural network layers, serving as input operands for photonic computation and mapped to modulation data of multi-wavelength optical signals. Furthermore, the neural network activation values ​​can be generated by the previous layer of the neural network, and after digital-to-analog conversion, drive the first loop-loop waveguide grating and modulator array. Depending on the network layer, the neural network activation values ​​can also include one or more of the following: input layer feature activation values, hidden layer intermediate activation values, and output layer predicted activation values. Correspondingly, when the data to be processed includes neural network activation values, the activation value data format in the neural network can be identified, converted into the electrical signals required to drive the loop-loop waveguide grating and modulator module, and the data to be processed can be obtained. This achieves semantic integration between the algorithm model and photonic hardware, eliminating the accuracy loss caused by data format conversion.

[0100] The convolutional kernel weights can be fixed parameter matrices used for feature extraction in a convolutional neural network. In this scheme, they serve as coefficient operands for photon computation and are loaded onto the corresponding optical wavelength. In this embodiment, the convolutional kernel weights can be used to define the transformation relationship between input and output features, realizing multiplication operations in the optical domain. For example, the convolutional kernel weights can be pre-trained and stored, then read and converted into control signals to be loaded into the modulation module. For example, when the coefficient data includes the convolutional kernel weights, the stored weight parameters can be read and converted into control signals required to drive the second modulation module. Furthermore, the loading of weight data into the physical properties of the optical domain can be completed through methods such as statically configuring the modulator bias and dynamically refreshing and updating the modulator state.

[0101] The output vector of matrix-vector multiplication can be the result sequence obtained by multiplying the input vector with the weight matrix and accumulating the results. In this scheme, it corresponds to the electrical signal converted by the photodetector PD array and is used as the calculation result of the current layer neural network.

[0102] This embodiment provides a photon tensor computation method that maps neural network activation values ​​to photon computation data to be processed, maps convolution kernel weights to photon computation coefficient data, performs photon tensor computation, and outputs a matrix-vector multiplication output vector. This method can meet the computational efficiency and accuracy requirements of large-scale neural networks, avoid the power consumption bottleneck of data transfer in traditional electronic computing, and ensure the accuracy of computation results while maintaining high parallelism. Ultimately, it achieves the technical effect of improving the accuracy and scalability of photon tensor computation, and meeting the computational efficiency and accuracy requirements of large-scale neural networks.

[0103] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0104] Based on the same inventive concept, this application also provides a photonic tensor kernel for implementing the photonic tensor calculation method described above. The solution provided by this photonic tensor kernel is similar to the implementation described in the above method. Therefore, the specific limitations in the one or more photonic tensor kernel embodiments provided below can be found in the limitations of the photonic tensor calculation method above, and will not be repeated here. Furthermore, the specific limitations in the one or more photonic tensor kernel embodiments provided below can also serve as supplementary content to the above photonic tensor calculation method.

[0105] In one embodiment, such as Figure 3 As shown, a photonic tensor kernel is provided, which includes:

[0106] Light source unit 1 is used to generate multi-wavelength light signals;

[0107] The parallel data loading optical engine 2 includes a first loop-back array waveguide grating 21, which is used to modulate the optical signals of each wavelength in the multi-wavelength optical signal separately and then multiplex them to obtain multiple multi-wavelength first optical signals; wherein, each first optical signal is isolated from each other in the spectrum; a first modulation module 22 is coupled on the loop-back path of the first loop-back array waveguide grating 21, which is used to modulate the optical signals of each wavelength according to the input data to be processed;

[0108] Beam splitting unit 3 is used to split each first optical signal to obtain multiple copies of the first optical signal;

[0109] The photonic dot product engine includes a second loop array waveguide grating 41, which is used to modulate the optical signals of each wavelength in a single copy of the first optical signal and multiplex them to obtain a single second optical signal; wherein, a second modulation module 42 is coupled on the loop path of the second loop array waveguide grating 41, which is used to modulate the optical signals of each wavelength according to the input coefficient data.

[0110] The photoelectric detection module 5 is used to perform photoelectric conversion on each second optical signal to obtain tensor calculation results.

[0111] The output of the light source unit 1 can be connected to the input of the parallel data loading optical engine 2 via a waveguide, generating an optical carrier that carries computational information and provides a carrier for subsequent modulation. For example, the light source unit 1 may include one or more of the following: a distributed feedback laser array with a wavelength division multiplexer, an external cavity laser with an optical frequency comb generator, and a superluminescent diode.

[0112] The input end of the parallel data loading optical engine 2 is connected to the light source unit 1, and the output end is connected to the beam splitter unit 3, thereby loading the data to be processed onto the optical carrier and realizing parallel optical domain encoding of the input data. It is understandable that traditional MZI arrays are greatly affected by temperature, leading to the accumulation of phase errors. By using the first loop-back arrayed waveguide grating 21 for wavelength routing and cooperating with the first modulation module 22 in the loop to complete signal modulation, at least two multiplexing of the arrayed waveguide gratings can be achieved, avoiding phase shifts caused by different arrayed waveguide gratings being affected by different temperatures.

[0113] The first loop-loop array waveguide grating 21 can utilize the optical path difference of the waveguide array to generate interference, guiding different wavelengths to a specific port to achieve demultiplexing and multiplexing of optical signals. In this embodiment, the first loop-loop array waveguide grating 21 includes a loop path, and the first modulation module 22 is coupled to the loop path of the first loop-loop array waveguide grating 21.

[0114] Furthermore, the first modulation module 22 may include an electrical signal terminal and a mapping terminal, wherein the electrical signal terminal is connected to the data source and is used to change the optical signal properties according to the data to be processed, thereby completing the optical domain mapping of the activation value. In an exemplary embodiment, the first modulation module 22 may receive the data to be processed in the form of an electrical signal and change the physical properties of the transmitted optical signal. Exemplarily, the first modulation module 22 may be implemented using one or more of an electroabsorption modulator, a Mach-Zehnder modulator, etc.

[0115] The first optical signal, which is multi-channel and multi-wavelength, is output by the parallel data loading optical engine 2 and input to the beam splitting unit 3. It carries the data to be processed and is a spectrum-isolated optical signal.

[0116] The beam splitting unit 3 is used to divide the input optical signal equally into multiple output ports with equal power. In this embodiment, each first optical signal copy can be divided into multiple first optical signal copies by the beam splitting unit 3, and then enter different dot multiplication engines for parallel operation.

[0117] In this embodiment, the number of photonic dot multiplication engines can be the product of the number of output ports of beam splitting unit 3 and the number of beam splitting units 3 (i.e., the number of output ports of parallel data loading optical engine 2). The input end of the photonic dot multiplication engine is connected to the beam splitting unit 3, and the output end is connected to the photodetector module 5, which is used to realize the optical domain multiplication operation of input data and weights.

[0118] The second loop-loop arrayed waveguide grating 41 receives a copy of the first optical signal at its input side, is coupled to the second modulation module 42 via a loop-loop path, and is connected to the photodetector module 5 at its output side. It can be used to provide independent wavelength channels for coefficient modulation, ensuring optical isolation between channels. In one specific embodiment, the second loop-loop arrayed waveguide grating 41 can be a 1×1 loop-loop arrayed waveguide grating, performing wavelength separation and loop routing again to prevent signal interference between different weighted arrays and ensure computational accuracy.

[0119] The second modulation module 42 is coupled to the loopback path of the second loopback array waveguide grating 41. It may include an electrical signal terminal, a storage unit for connecting coefficient data, and encodes the coefficient data onto an optical carrier to realize multiplication operations in the optical domain. In this embodiment, the architecture of the second loopback array waveguide grating 41 can be the same as that of the first loopback array waveguide grating 21. Similarly, the second modulation module 42 can also adopt the same structure as the first modulation module 22.

[0120] Understandably, in traditional schemes, data and weights are tightly coupled, making flexible configuration difficult. In this embodiment, the photonic dot multiplication engine can use the second loop array waveguide grating 41 and the second modulation module 42 coupled to the loop path to perform secondary modulation on the copy of the first optical signal.

[0121] The photoelectric detection module 5, for example, may include multiple photodetectors, wherein the input end of each photodetector can be connected to a corresponding photonic multiplication engine, and the output end can be connected to a subsequent digital processing circuit, thereby converting the optical domain multiplication result into an electrical signal, accumulating it, and outputting the final calculated value. For example, the photoelectric detection module 5 may employ one or more of the following, including but not limited to single-ended photodiode detectors and balanced photodetectors.

[0122] This embodiment provides a photonic tensor kernel that generates multi-wavelength optical signals through a light source unit 1 and combines the wavelength division multiplexing characteristics of the first loop array waveguide grating 21 and the second loop array waveguide grating 41. Utilizing the high channel density advantage of the array waveguide grating, it can effectively overcome the channel number bottleneck caused by the radius limitation of the micro-ring structure, significantly improving the system's scalability. Simultaneously, by adopting a parallel data loading optical engine 2 and a photonic dot product engine-separated architecture, the data to be processed and the coefficient data are independently loaded through the first modulation module 22 and the second modulation module 42, respectively. This also avoids the direct interference of phase fluctuations caused by the large thermo-optic coefficient of silicon material in traditional Mach-Zehnder interferometer arrays with the calculation process. Combined with the beam splitting unit 3 and the photoelectric detection module 5 to complete signal distribution and photoelectric conversion, it can reduce the impact of phase errors on computational accuracy while ensuring spectral isolation, thereby achieving the technical effect of improving the accuracy and scalability of photonic tensor calculation.

[0123] In one embodiment, the first loopback array waveguide grating includes a first star coupler, a second star coupler, an array waveguide connecting the first star coupler and the second star coupler, and a first loopback path connecting the output side of the second star coupler and the input side of the first star coupler.

[0124] The first modulation module is coupled to the first loopback path and is used to modulate the optical signals of each wavelength according to the input data to be processed, and send the modulated optical signals of each wavelength back to the first star coupler.

[0125] The input side of the first star coupler is connected to an external optical signal source and the optical signal returned from the first loopback path, while the output side is connected to the input end of the arrayed waveguide. It can be used as the input / output interface of the arrayed waveguide grating, enabling coupling and distribution of the optical signal between the free-space propagation region and the waveguide array. In this embodiment, the first star coupler can receive the input optical signal and, by utilizing the diffraction effect of the planar waveguide region, disperse the input optical signal to each channel of the arrayed waveguide. For example, the first star coupler can be a star coupler based on a Rowland circle structure, or other star coupler structures that meet the phase-matching condition.

[0126] The input side of the second star coupler is connected to the output end of each channel of the array waveguide, and the output side is connected to the input end of the first loopback path. It can be used as the other end interface of the array waveguide grating to complete the convergence and conversion of optical signals from each channel of the array waveguide to the loopback path. In this embodiment, the second star coupler can receive array waveguide signals, receive array waveguide signals with different optical path differences, and perform interference convergence to obtain a converged, i.e., multiplexed optical signal.

[0127] The arrayed waveguide connects the first star coupler and the second star coupler, and can be used to provide a precise fixed optical path difference, so as to realize the spatial separation and phase relationship control of optical signals of different wavelengths.

[0128] The first loopback path connects the output side of the second star coupler to the input side of the first star coupler. It can be used to construct an internal loop path for the optical signal, guiding the optical signal from the output side of the second star coupler back to the input side of the first star coupler for modulation. This avoids the phase instability caused by long-path interference, which is overly sensitive to environmental temperature changes or manufacturing deviations, common in traditional Mach-Zehnder interferometer arrays. In this embodiment, the first loopback path can be a closed loop formed by receiving the output optical signal and connecting it via a physical waveguide. This allows the optical signal to complete wavelength-based spatial routing before entering the modulation module, resulting in a cyclic optical signal.

[0129] This embodiment provides a photonic tensor kernel, which uses a first star coupler and a second star coupler in conjunction with an arrayed waveguide to form an arrayed waveguide grating structure. Compared to the micro-ring resonator arrays in traditional technologies that are limited by the radius of the micro-rings, this effectively overcomes the physical limitation on the number of channels and significantly improves the scalability of the system. The output side of the second star coupler is connected to the input side of the first star coupler through a first loop path, forming an internal loop path for the optical signal. This allows the first modulation module to couple onto this path to modulate the optical signal, thereby effectively avoiding the phase fluctuation problem caused by long-path interference in traditional Mach-Zehnder interferometer arrays, which is too sensitive to changes in ambient temperature or manufacturing deviations. Furthermore, by using the spatial beam splitting characteristics of the star coupler and the optical path difference design of the arrayed waveguide, the impact of phase error on the computational accuracy can be reduced while ensuring multi-wavelength signal processing, thus achieving the technical effect of improving the accuracy and scalability of photonic tensor calculation.

[0130] In one embodiment, the second loopback array waveguide grating includes a third star coupler, a fourth star coupler, an array waveguide connecting the third star coupler and the fourth star coupler, and a second loopback path connecting the output side of the fourth star coupler and the input side of the third star coupler.

[0131] The second modulation module is coupled to the second loop path and is used to modulate the optical signals of each wavelength according to the input coefficient data, and send the modulated optical signals of each wavelength back to the third star coupler.

[0132] In this embodiment, the input side of the third star-shaped coupler is connected to the second loop path, and the output side is connected to the input end of the arrayed waveguide. It can serve as the input / output interface of the arrayed waveguide grating, enabling the divergence and convergence of optical signals between the free propagation region and the arrayed waveguide. In this embodiment, the third star-shaped coupler can receive the input optical signal and, by utilizing the diffraction effect of the planar waveguide region, uniformly distribute the input optical signal to multiple arrayed waveguide channels to obtain the output optical signal.

[0133] The input side of the fourth star coupler is connected to the output end of the arrayed waveguide, and the output side is connected to the second loop path. It can serve as the other end interface of the arrayed waveguide grating, multiplexing and focusing the optical signal transmitted through the arrayed waveguide. In this embodiment, the fourth star coupler can receive the optical signal transmitted through the arrayed waveguide and, through a phase difference generated based on the optical path difference, focus optical signals of different wavelengths to different positions or the same position on the output side to obtain the multiplexed optical signal.

[0134] The modulated optical signals of each wavelength are output from the second modulation module and input to the input side of the third star coupler, thereby enabling the formation of a copy of the first optical signal carrying part of the data to be processed with the coefficient data.

[0135] This embodiment provides a photonic tensor kernel. A waveguide structure connecting the two via a third and fourth star coupler enables the construction of a stable optical signal transmission and multiplexing path. Compared to traditional Mach-Zehnder interferometer architectures, which are susceptible to phase fluctuations due to ambient temperature, this optical path structure, utilizing star couplers and arrayed waveguides, exhibits better temperature stability, reducing phase errors from a physical perspective. Furthermore, this structure supports parallel processing of multi-wavelength signals within a compact space, overcoming the channel number bottleneck caused by the radius limitation of micro-ring structures. Combined with modulation mechanisms on the loopback path, it enables the loading of coefficient data and signal multiplexing. While ensuring spectral isolation, it also reduces the impact of phase errors on computational accuracy, thereby improving the accuracy and scalability of photonic tensor calculations.

[0136] In one embodiment, the second loop array waveguide grating further includes a flat-top spectral response input unit;

[0137] The third star coupler obtains a copy of the first optical signal after spectral shaping through the flat-top spectral response input unit; each wavelength component in the first optical signal copy falls within the preset flat-top passband.

[0138] In this embodiment, the flat-top spectral response input unit serves as the input unit for the arrayed waveguide grating, receiving the optical signal from the preceding waveguide structure. It is understood that traditional arrayed waveguide gratings have narrow passbands, and variations in ambient temperature and manufacturing process deviations can easily cause misalignment of the channel spectra between the preceding and following arrayed waveguide gratings, leading to computational errors. This embodiment uses a flat-top spectral response input unit to ensure that the optical wavelength signal from the preceding arrayed waveguide grating always falls within the corresponding range of the flat-top spectrum of the following arrayed waveguide grating, thereby mitigating or even avoiding the aforementioned problems. For example, the flat-top spectral response input unit can be a Gaussian spectral response modified into a flat-top shape, ensuring that the narrow channel spectrum of the preceding arrayed waveguide grating always falls within the broadband flat-top spectrum of the following arrayed waveguide grating, avoiding computational errors caused by misalignment of the spectral center wavelengths due to temperature changes and manufacturing process errors.

[0139] This embodiment provides a photonic tensor kernel that, by introducing an input unit with a flat-top spectral response, performs spectral shaping on the optical signal entering the third star coupler, ensuring that each wavelength component falls within a preset flat-top passband. The flat-top passband provides a wider tolerance range, effectively tolerating wavelength drift caused by environmental temperature changes or process deviations, reducing signal crosstalk and isolation degradation caused by wavelength drift. Simultaneously, spectral shaping ensures the power stability of the optical signal during transmission, reducing the interference of phase errors on the calculation results, and guaranteeing the accuracy of signal processing at the physical level. Therefore, it achieves the technical effect of improving the accuracy and scalability of photonic tensor calculations.

[0140] In one embodiment, the flat-top spectral response input unit employs a Mach-Zehnder interferometer or a multimode interferometer.

[0141] The Mach-Zehnder interferometer is an optical phenomenon based on the principle of two-beam interference. Its core structure consists of two beam splitters. By separating the incident light into two independently transmitted beams, which are then recombined after different paths or phase modulation, an interference effect is generated. The light intensity output of this phenomenon follows a sinusoidal law, and its mathematical model can be expressed as a function of the output light intensity and the phase difference between the two arms. In this embodiment, the Mach-Zehnder interferometer can be used as a specific implementation of a flat-top spectral response input unit and utilizes the interference effect to shape the optical signal spectrum to match the input requirements of the arrayed waveguide grating. For example, when the phase difference between the two arms remains constant or changes slowly within a specific wavelength range, the interference output can form an approximately flat response curve within that band, thereby achieving flat-top spectral characteristics.

[0142] A multi-mode interferometer (MMI) can also be used as one of the specific implementations of a flat-top spectral response input unit. Specifically, a multi-mode interferometer can utilize the self-image effect in a multi-mode waveguide to redistribute the input optical field at the output port at a specific length. By optimizing the waveguide width, length, and input position, a relatively flat transmission response can be obtained within the target wavelength range.

[0143] This embodiment provides a photonic tensor kernel that uses a Mach-Zehnder interferometer or a multimode interferometer as a key component for spectral shaping. By utilizing its interference characteristics, it constructs a spectral response curve with a flat top. Compared to traditional microring resonators, it can have a wider passband tolerance and can effectively tolerate the misalignment of the center wavelengths of the first and second loop array waveguide grating channels caused by environmental temperature changes due to the large thermo-optic coefficient of silicon material and processing errors. This effectively reduces the risk of signal crosstalk and isolation degradation, thereby achieving the technical effect of improving the accuracy and scalability of photonic tensor calculation.

[0144] In one embodiment, multiplexing to obtain multiple multi-wavelength first optical signals includes:

[0145] The modulated optical signals of each wavelength are grouped according to the preset number of channels, and each group corresponds to one first optical signal; the interval between adjacent wavelengths in the frequency domain of each first optical signal is the product of the interval between adjacent wavelengths in the multi-wavelength optical signal and the preset number of channels.

[0146] This embodiment provides a photonic tensor kernel that establishes a structured mapping relationship between signal channels by grouping modulated optical signals according to a preset number of channels. This ensures clear logical isolation between multiple signals. By setting the frequency domain interval between adjacent wavelengths in each signal to be the product of the original interval and the number of channels, the spectral guard band between channels is artificially expanded, effectively avoiding crosstalk caused by spectral overlap between different channels and solving the problem of isolation degradation. Furthermore, the photonic tensor calculation method in this embodiment, through a frequency interval expansion-based multiplexing approach, allows for linear expansion of calculation channels by increasing the preset number of channels without increasing the complexity of physical devices. This overcomes the channel number bottleneck limited by the radius of the micro-ring structure, thereby improving the parallel processing capability of the system while ensuring signal purity. Therefore, it achieves the effect of improving the accuracy and scalability of photonic tensor calculation.

[0147] In one embodiment, the beam splitting unit includes multiple beam splitters; the number of beam splitters corresponds to the number of preset channels; each beam splitter is used to split the corresponding first optical signal to obtain multiple copies of the first optical signal; the number of copies of the first optical signal is equal to the number of photon multiplication engines.

[0148] In this embodiment, the beam splitting unit may include multiple beam splitters. The input of each beam splitter may be connected to a parallel data loading optical engine, and the output may be connected to multiple photon multiplication engines. This allows for the distribution of a single optical signal energy across multiple paths, supporting parallel computation of the same input data and multiple sets of coefficient data. In one specific embodiment, the beam splitter may utilize a 1×P optical power divider to evenly distribute the input optical signal to multiple output ports with equal power. For example, the beam splitter may include one or more of a Y-branch waveguide beam splitter and a multimode interference coupler beam splitter.

[0149] This embodiment provides a photonic tensor kernel that uses multiple beamsplitters within a beam-splitting unit, with the number of beamsplitters matching the number of channels required by the system. Each beamsplitter replicates the input optical signal, generating multiple identical copies. This ensures that the number of optical signal copies matches the number of backend computing modules. By achieving precise one-to-many allocation of the optical signal, it ensures that each computing module receives the complete optical signal for parallel processing. This avoids computational errors caused by uneven signal distribution or channel mismatch, improving the system's parallel processing capability while ensuring signal integrity, thus achieving the technical effects of improving computational accuracy and scalability.

[0150] In one embodiment, both the first modulation module and the second modulation module employ one of the following: a Mach-Zehnder interference structure, a resonant micro-ring structure, and an electro-absorption modulation structure.

[0151] The Mach-Zehnder interferometer structure can achieve optical signal intensity modulation through the phase difference of the interferometer arms, providing a highly linear intensity control path. In a specific embodiment, the Mach-Zehnder interferometer structure can receive an input optical signal, split the input light into two paths, and modulate the light intensity by changing the phase of one or both paths, thereby obtaining a modulated optical signal.

[0152] An electro-absorption modulation structure can modulate an optical signal by changing the absorption coefficient of a material through an electric field. In this embodiment, the electro-absorption modulation structure can receive an input optical signal and, based on the Franz-Keldish effect or the quantum confinement Stark effect, change the semiconductor band structure by applying an external voltage, thereby changing the light absorption rate and obtaining a modulated optical signal.

[0153] The resonant microring structure can enhance the interaction between light and matter by utilizing a resonant cavity, achieving highly integrated wavelength-selective modulation, thus balancing integration and wavelength stability. For example, the resonant microring structure can receive an input optical signal, and by changing the refractive index of the microring, shift the resonant wavelength, thereby altering the transmittance of a specific wavelength of light to achieve intensity modulation, resulting in a modulated optical signal.

[0154] In one exemplary embodiment, for example in scenarios with specific requirements for modulation speed and thermal stability, the photonic tensor kernel can be deployed in edge computing nodes that require fast response and experience significant ambient temperature fluctuations, and both the first and second modulation modules employ electro-absorption modulation structures. Since electro-absorption modulation structures do not rely on thermo-optical effects, changes in ambient temperature do not cause significant fluctuations in light intensity; therefore, the tensor calculation results obtained by the photodetector module can exhibit higher stability. In another exemplary embodiment, for example in scenarios with higher integration requirements, a resonant microring structure can be used. In yet another exemplary embodiment, for example in scenarios with higher linearity requirements, a Mach-Zehnder interferometer structure can be used, thereby achieving flexible configuration according to scenario requirements.

[0155] This embodiment provides a photon tensor kernel that, by selecting one of the following structures for the first and second modulation modules—a Mach-Zehnder interference structure, a resonant microring structure, or an electroabsorption modulation structure—can reduce light intensity fluctuations and wavelength drift caused by changes in ambient temperature. It allows for a trade-off between accuracy and integration based on specific application scenarios, and avoids the channel limit problem caused by the radius constraint of a single microring structure. Therefore, it can improve the scalability and flexibility of the system while ensuring the accuracy of photon tensor calculation.

[0156] In one embodiment, the light source unit includes at least one of a superradiative diode light source, a multi-wavelength laser light source, and an optical frequency comb light source.

[0157] Among them, the superluminescent diode (SLD) is a broadband, high-brightness semiconductor light source that falls between ordinary light-emitting diodes (LEDs) and semiconductor lasers (LDs). When current is injected into the active region of the semiconductor, carrier recombination generates photons. These photons propagate in the waveguide structure and are continuously amplified through stimulated emission, thereby outputting high-power, low-coherence, and broadband incoherent light.

[0158] A multi-wavelength laser source can be a light source system capable of simultaneously or switchably outputting multiple discrete wavelengths (or frequencies) of laser light. For example, multi-segment gain region lasers, multi-wavelength fiber lasers, etc., can be used to achieve the output of multiple stable, narrow-linewidth laser wavelengths, with good wavelength spacing control between each wavelength.

[0159] An optical frequency comb source is a special multi-wavelength light source capable of generating a series of equally spaced, highly coherent, narrow-linewidth discrete laser frequencies (or wavelengths). Its output spectrum exhibits a regularly arranged "comb" structure in the frequency domain, with highly stable and precisely controllable frequency intervals between adjacent comb teeth. Taking a mode-locked laser as an example, when stable ultrashort pulse mode-locking operation is achieved within the laser cavity, the periodic pulse sequence in the time domain corresponds to a set of equally spaced frequency components in the frequency domain.

[0160] This embodiment provides a photon tensor kernel that, by employing a superradiative diode light source, a multi-wavelength laser light source, or an optical frequency comb light source, provides a highly stable and spectrally excellent optical signal generation capability. It can provide highly stable and equally spaced multi-wavelength signals, thereby reducing the interference of phase errors on the calculation results, improving calculation accuracy, and supporting a larger number of available channels. This achieves the technical effect of improving the accuracy and scalability of photon tensor calculations.

[0161] To more clearly illustrate the technical solution of this application, a detailed embodiment is also provided.

[0162] The increasing computational demands of current artificial intelligence (AI) models (such as Transformer and CNN) have led to a surge in the demand for computing power. Traditional electronic chips (GPUs / TPUs) based on the von Neumann architecture, with their separate in-memory and computational capabilities, are constrained by "memory walls" and "power walls," making it difficult to meet the energy efficiency requirements of real-time inference and large-scale training. Photonic computing, with its inherent parallelism (wavelength division multiplexing, WDM), high-speed transmission, low power consumption, and low latency, has become a revolutionary solution to overcome the computational bottleneck. However, it still faces challenges in integration, precision controllability, and optoelectronic synergy. Tensor operations are the mathematical foundation of AI computing. Their core role is to provide structured data representation and parallel computing architecture for neural networks, directly impacting model performance, hardware efficiency, and algorithm design. Over 90% of the computation in modern neural networks is concentrated in tensor operations, primarily matrix-vector multiplication (MVM). An MVM operation includes dot product of matrices and vectors, as well as multiplication-addition operations. Photonic tensor kernels used for AI computing acceleration primarily replace traditional electronic computing. They utilize photonic technology to achieve optical domain MVM operations, thereby solving the problem of high power consumption caused by data transfer in electronic computing, realizing the function of optical transmission as computing, and significantly improving the computing energy efficiency ratio.

[0163] Currently, the main technologies for realizing MVM on silicon photonics platforms include Mach-Zehnder interferometer (MZI) array architecture based on coherent light, microring resonator array architecture based on incoherent light using WDM technology, phase change material (PCM) + crossbar architecture, and arrayed waveguide grating router (AWGR) architecture. Due to the large thermo-optic coefficient of silicon (1.87 × 10⁻⁴ / K), changes in ambient temperature will cause phase fluctuations in MZI devices and wavelength shifts in microrings and AWGs. Simultaneously, due to limitations in fabrication processes, the actual size of silicon waveguides will deviate from the design value, causing phase inhomogeneity in the phase of each MZI unit under the MZI architecture, thus introducing phase errors, microring wavelength drift, and wavelength drift and degradation of adjacent isolation in AWGs. Furthermore, the arrayed microring architecture is limited by the microring radius, restricting the number of available channels, and the narrow spectral width of the microrings makes them more sensitive to changes in ambient temperature. All of these factors severely limit the scalability and computational accuracy of photonic tensor kernels. In addition, the write lifetime of phase change materials and the array fabrication yield are also major bottlenecks. Furthermore, the phase change materials introduce additional electro-optical conversion energy consumption when performing weight writing, which indirectly weakens their energy efficiency ratio. At the same time, optical crosstalk between adjacent PCM units in the cross waveguide array architecture limits the computational accuracy and scalability of the architecture.

[0164] To address the scalability and computational accuracy limitations of existing photonic tensor kernels and the urgent computational demands in AI computing, this embodiment proposes a novel on-chip photonic tensor kernel for hardware acceleration of AI computing. Parallel input data vectors are prepared using a broadband SLD light source or laser array, a 1×K loop-around AWG, and an MZI modulator array. A convolution kernel matrix is ​​prepared using a 1×1 loop-around AWG array and an MZI modulator array. Matrix-vector multiplication and addition operations are performed by cascading the preceding 1×K loop-around AWG for preparing the input data vector, the subsequent 1×1 loop-around AWG array for preparing the convolution kernel matrix, and a PD array, thereby achieving parallel optical computing.

[0165] In one embodiment, such as Figure 4 As shown, a novel high-precision scalable photonic tensor kernel for accelerating AI convolutional computation is provided, comprising a light source unit and a photonic integrated chip unit. The photonic integrated chip unit includes a parallel data loading optical engine, multiple beam splitters, a photonic dot product engine, and a photodetector array.

[0166] like Figure 5 As shown, the parallel data loading optical engine mainly consists of a 1×K (1 input port, K ​​output ports) loop AWG, which has N×K waveguides (corresponding to N×K wavelengths) and a first modulation module that loads the input data onto the optical wavelength, including N×K modulators.

[0167] like Figure 6As shown, the photonic dot product engine consists of a 1×1 (1 input port, 1 output port) loop AWG, which has N waveguides and a second modulation module that loads N convolution kernel elements onto the optical wavelength, including N modulators.

[0168] The photonic tensor kernel operates as follows: K sets of external data of length N are modulated to their corresponding optical wavelengths λ by a parallel data loading optical engine. （m-1）K+j Above, each group of N wavelength signals carrying N data points is split into P equal-power signals by a 1×P beam splitter. Each of the N wavelength signals carrying data then passes through a photonic dot product engine, which converts the N convolution kernel elements W... q，m Loaded to the corresponding light wavelength λ （m-1）K+j This allows for the control of wavelength λ. （m-1）K+j Data X carried j，m（t） and weighted data W q，m The multiplication, the output of the photon dot product engine, and a detector PD j， q This allows for the concatenation of N input data points with their corresponding N weight data points, thus completing a single operation of multiplying and summing them. .

[0169] The parallel data loading optical engine utilizes the waveguide spacing and position control at the AWG star coupler interface to divide the N×K wavelengths (spaced Δλ) in the input light into K groups and output them from K output waveguides. Each output waveguide contains N wavelengths, and the spacing between adjacent wavelengths is K times Δλ. Figure 7 The positional distribution of the AWG input and output waveguides and the loopback waveguide at the interface of the star coupler is given. For the sake of simplicity, the diagram is... Figure 5The two star couplers of the AWG are combined into one, and only the loopback waveguide and input / output waveguides are shown. To clearly understand the use of this loopback AWG, K loopback input waveguides are corresponding to each demultiplexed wavelength on the input waveguide side. In actual use, only one waveguide is selected. When one loopback input waveguide is selected, the wavelength will be output from one of the K specific output waveguides. For example, if the red loopback input waveguide is selected, the wavelength will be output from the red output waveguide on the right. If the green loopback input waveguide is selected, it will be output from the green output waveguide on the right. To distribute the N×K wavelengths evenly and at equal intervals among the K output waveguides, the N×K wavelengths are divided into N groups, numbered from group 1 to group N. Each group has K wavelengths, numbered from wavelength 1 to wavelength K. The loopback waveguide of each wavelength 1 in these N wavelength groups is connected to the red input waveguide on the left, the loopback waveguide of each wavelength 2 is connected to the green input waveguide on the left, and so on. The loopback waveguide of each wavelength K is connected to the purple input waveguide on the left. After transmission through the AWG, the N wavelengths input are looped back from the left red input waveguide, namely λ1, λ2, λ3, λ4, λ5, λ6, λ7, λ8, λ9, λ1, λ1, λ2 ...2, λ9, λ1, λ2, λ9, λ1, λ2, λ K+1 , λ 2K+1 , λ 3K+1 、…、λ （N-1）K+1 The wavelength group will be multiplexed out from the red output waveguide 1 on the right; the N wavelengths, namely λ2, λ3, and λ4, will be looped back from the green input waveguide on the left. K+2 , λ 2K+2 , λ 3K+2 、…、λ （N-1）K+2 The wavelength group will be multiplexed from the right-hand green output waveguide 2, and so on, looping back the N wavelengths, i.e., λ, from the left-hand purple input waveguide. K , λ K+K , λ 2K+K , λ 3K+K 、…、λ （N-1）K+K The wavelength group will be multiplexed from the purple output waveguide 1 on the right, such as... Figure 8 As shown. In order to achieve the above-mentioned functions, the AWG has N×K output waveguides ( Figure 7 The spacing of the black waveguide (top right) at the star coupler interface is set to D0, and each wavelength loop returns to the input side ( Figure 7 The upper left waveguide corresponds to K optional connecting waveguides. The spacing between these K waveguides at the star coupler interface is D0 / K. The spacing between waveguides of the same color corresponding to different wavelengths at the star coupler interface is D0. Figure 7 The schematic diagram shows that the spacing between the red and green waveguides is D0 / K, while the spacing between the red waveguides is D0; the spacing between the K output waveguides of the loopback AWG at the interface of the star coupler is also D0 / K. Figure 8 yes Figure 7The actual implementation structure is illustrated below. To achieve the goal of distributing N×K wavelengths (wavelength interval of Δλ, corresponding to the input star coupler, i.e., the input waveguide on the first star coupler) to K output waveguides after two transmissions via the AWG, each output waveguide contains N wavelengths with adjacent wavelengths spaced K×Δλ apart. Figure 8 The positional relationship between the input waveguide of the input star coupler and the output star coupler, i.e., the N×K output waveguides (black waveguides) on the second star coupler, at the star coupler interface satisfies: n si ×d a ×sinα+ n ai ×ΔL + n si ×d a ×sin θ i = M × λ i Where 1≤i≤N×K, n si and n ai These are the effective refractive indices of the free transmission region and the arrayed waveguide at the corresponding wavelength λi, respectively; α is the angle between the centerlines of the input waveguide and the input star coupler; and d... a θ is the center-to-center distance between adjacent array waveguides at the interface of the free transmission region, ΔL is the length difference between adjacent array waveguides, and θ is the center-to-center distance between adjacent array waveguides. i is the angle between the centerline of the i-th output waveguide and the centerline of the output star coupler, and M is an integer diffraction order. For N×K wavelengths, where λ1, λ2, ..., λ3 are numbered... K+1 , λ 2K+1 , λ 3K+1 、…、λ （N-1）K+1 The wavelengths, corresponding to the input waveguides 1-1, K+1-1, 2K+1-1, ..., (N-1)K+1-1 on the input star coupler, are transmitted via AWG and then multiplexed to the output waveguide #1 on the output star coupler; similarly, the wavelengths numbered λ... j , λ K+j , λ 2K+j , λ 3K+j 、…、λ （N-1）K+j Wavelength, 1≤j≤K, corresponds to input waveguides jj, K+jj, 2K+jj, ..., (N-1)K+jj on the input star coupler. After AWG transmission, these waveguides are multiplexed to the output waveguide #j on the output star coupler. On the loopback input waveguide of the input star coupler, for waveguide number ij (j is the remainder of integer i divided by integer K, i.e., j = mod (i, K)), the angle between its waveguide number ij and the centerline of the input free transmission region is γ. i-j The angle βj between the output waveguide #j and the center line of the output star coupler satisfies n si ×d a ×sin γi-j + n ai ×ΔL + n si ×d a ×sin β j = M × λ i Furthermore, the angle α between the centerlines of the input waveguide and the input star coupler and the angle β1 between the centerlines of the #1 output waveguide and the output star coupler satisfy the relationship α = β1.

[0170] Figure 9 Given Figure 2 The typical spectral response of the K output channels of the 1×K loopback AWG described above can be expressed as: AWG M (j) = I j (λ) （m-1）K+j ) is a wavelength λ （m-1）K+j The intensity curve of the Gaussian peak centered at the center; the spectrum of each output waveguide has N Gaussian-shaped peaks with a center wavelength interval of K×Δλ between adjacent peaks, where the wavelength interval Δλ is the wavelength difference between adjacent numbered center wavelengths, i.e., the wavelength difference between adjacent numbered wavelengths λ. （m-1）K+j and λ （m-1）K+j+1 The wavelength difference between them, where 1≤m≤N, 1≤j≤K.

[0171] Figure 10 Is Figure 2 This is a schematic diagram showing the addition of a modulator to the loopback waveguide. The modulator's function is to convert external data X... j，m (1≤m≤N) Loaded to the corresponding light wavelength λ （m-1）K+j The loading process involves controlling the wavelength λ of light. （m-1）K+j The light intensity can be controlled by using N×K modulators to load K sets of data of length N onto N×K wavelengths of the loop-through AWG. Each set of data is then output from K output waveguides after passing through the AWG. Figure 11 As shown.

[0172] Figure 12 yes Figure 10 Modulators are introduced into the N×K loop waveguides of the 1×K loop AWG described above to load N×K external data onto the wavelengths of the corresponding loop waveguides. That is, the transmittance of the data X is adjusted by the modulators at the corresponding wavelengths of each loop waveguide. j，m Loaded to optical wavelength λ （m-1）K+j Above. For K output waveguides, after data is loaded by the modulator, the typical spectral response in each waveguide is: Data(j) = ModX j，m (λ) （m-1）K+j() corresponds to the modulator in the loop waveguide and its effect on the propagating wavelength λ in the waveguide. （m-1）K+j The modulation curve. From Figure 12 It can be seen that after loading external data, the intensity values ​​corresponding to the N Gaussian peaks in each output channel are relatively higher than before the modulator was applied. Figure 9 The change has occurred, thus enabling the transfer of external data X j，m Loaded to the corresponding light wavelength λ （m-1）K+j superior.

[0173] for Figure 6 The key component for the photonic dot product engine to achieve its function is a 1×1 loop-loop AWG with a flat-top spectral response. Figure 13 A schematic diagram of the design principle of a 1×1 loop-back AWG with a flat-top spectral response is presented. It consists of an input unit based on an MZI structure, an arrayed waveguide grating, N loop-back waveguides, and one output waveguide. The purpose of the MZI-structured input unit is to achieve the desired response for each center wavelength λ of the N loop-back waveguides. （m-1）K+j Achieving flat-top spectral response characteristics such as Figure 15 As shown in the typical response spectrum, it generates an electric field distribution related to the input wavelength at the interface between the MMI and the star coupler by specially designing the MZI structure at the input end. This input structure can also be replaced with other specially designed structures that can achieve flat-top spectral response characteristics of the AWG output channel, such as a single MMI. Figure 14 It is Figure 13 The principle structure diagram is transformed into a specific implementation structure diagram. For Figure 14 The 1×1 loop-back AWG first demultiplexes the input light in the input waveguide to N loop-back output waveguides. The corresponding center wavelengths in each waveguide satisfy the following relationship: n s_（m-1）K+j × d a ×sinθ+n a_（m-1）K+j × ΔL' + n s_（m-1）K+j × d a ×sin φ m = M' × λ （m-1）K+j Where 1≤m≤N, 1≤j≤K, n s_（m-1）K+j and n a_（m-1）K+j These are the free transmission region and the corresponding wavelength λ in the arrayed waveguide, respectively. （m-1）K+j The effective refractive index, θ is the angle between the centerlines of the input waveguide and the input star coupler (i.e., the third star coupler), d a ΔL' is the center-to-center distance between adjacent arrayed waveguides at the interface of the free transmission region, ΔL' is the length difference between adjacent arrayed waveguides, and φ is the center-to-center distance between adjacent arrayed waveguides at the interface of the free transmission region. mM' is the angle between the centerline of the m-th output waveguide and the centerline of the output star coupler (i.e., the fourth star coupler), where M' is an integer diffraction order. The N loopback output waveguides loop back to the input star coupler, and after being transmitted through the AWG again, are multiplexed into the output waveguides for output. For the center wavelength λ in the loopback output waveguides... （m-1）K+j It also satisfies the relation: n s_（m-1）K+j × d a ×sinφ m + n a_（m-1）K+j × ΔL' + n s_（m-1）K+j × d a ×sin θ= M' × λ （m-1）K+j .

[0174] Figure 16 Is Figure 13 A schematic diagram showing the introduction of modulators into the N loop waveguides. Figure 17 This is a schematic diagram of an implementation structure of a 1×1 loop AWG with modulators introduced into each loop waveguide. The function of the modulator is to convert external data W... q，m (1≤m≤N, 1≤q≤P) Loaded to the corresponding light wavelength λ （m-1）K+j The loading process involves controlling the wavelength λ of light. （m-1）K+j The light intensity can be controlled, and data of length N can be loaded onto N wavelengths of the loopback AWG through N modulators, and then output from the output waveguide after transmission through the AWG. Figure 18 The typical output spectrum is shown below.

[0175] The modulator in the loop waveguide of a loop-loop AWG can be based on a Mach-Zehnder interferometer (MZI) structure, a resonant micro-ring structure, or an electro-absorption modulation structure, etc., with the aim of achieving intensity control of the optical signal. Different optical intensities correspond to different external data. In a 1×K loop-loop AWG, the modulation response function of the modulator to wavelength can be expressed as Mod_X. j，m (λ) （m-1）K+j In a 1×1 loop AWG, the modulation response function of the modulator to wavelength can be expressed as Mod_W. q，m (λ) （m-1）K+j ).

[0176] for Figure 21 The selection of the input light source needs to be combined with the requirements of the parallel data loading optical engine and the photon dot product engine, and its light source spectrum can be expressed as S(λ). The following is an example of an external light source connected to the input waveguide of a 1×K loop AWG, where... Figure 19 It is a process in which N×K single-wavelength lasers are combined into a multi-wavelength laser source with N×K wavelengths through a wavelength division multiplexer (WDM). Figure 20It utilizes a pump laser and a cascaded micro-ring based on silicon nitride waveguides to generate an optical frequency comb light source with multiple equally spaced frequency peaks (multi-wavelengths) by taking advantage of the nonlinear effect of silicon nitride. Figure 21 A typical broadband output spectrum of a superluminescent diode (SLD) is presented for connection to the input waveguide of a 1×K loop AWG to provide a broadband light source spectrum.

[0177] exist Figure 21 In this process, the light emitted from the input light source is transmitted through the parallel data loading optical engine and the photon dot product engine to multiply the external input data and weight data. Finally, it passes through the corresponding detectors PDj and q to complete the summation operation of the multiplication of N input data and the corresponding N weight data, which is the data domain operation. Transforming to optical domain operations, the formula is as follows:

[0178]

[0179] This embodiment provides a photonic tensor kernel, proposing different design architectures based on loop-loop AWGs to realize photonic tensor chips based on wavelength, time, and space multidimensional multiplexing. Its parallel data loading optical engine cleverly designs the positional differences of the input and output waveguides of a 1×K loop-loop AWG on a star coupler (Roland circle), recombinating multiple input optical wavelengths / narrowband peaks and distributing them to K output waveguides, each waveguide containing N optical wavelengths / narrowband peaks. Simultaneously, modulators on the loop-loop waveguides load external data onto the corresponding optical wavelengths / narrowband peaks, completing the loading of K groups of N wavelength data per group. Its photonic dot product engine uses a 1×1 loop-loop AWG and a modulator array to load N convolution kernel weight data onto N optical wavelengths / narrowband peaks. The cascaded parallel data loading optical engine, photonic dot product engine, and detector complete the dot product and summation operations on the external input data and convolution kernel weight data. The entire data loading and computation are completed on the photonic integrated chip. Compared to traditional parallel computing architectures based on WDM technology, which require external discrete WDM for wavelength demultiplexing and multiplexing, this architecture has significant advantages in miniaturization and low power consumption of photonic tensor kernels. Photonic tensor computation is essentially analog computation, and photonic devices are sensitive to process deviations and temperature variations, leading to decreased computational accuracy. This embodiment cleverly designs a loop-loop AWG, where wavelength demultiplexing and multiplexing functions share the same AWG, effectively eliminating wavelength mismatch problems caused by process deviations during demultiplexing and multiplexing. Simultaneously, a 1×K loop-loop AWG increases the spacing between N wavelengths in the same output waveguide to K times the spacing between adjacent wavelengths / narrowband peaks of the original input, ensuring that even with process deviations, the wavelengths / narrowband peaks of each output channel of the 1×K loop-loop AWG remain within the flat-top passband of the 1×1 loop-loop AWG in the photonic dot product engine, thereby reducing the impact of process deviations on the computational results. By employing the same waveguide structure design for all loop AWGs, all AWGs exhibit the same temperature dependence, achieving relative temperature insensitivity between AWGs. This ensures that the wavelengths of the 1×K loop AWGs and each 1×1 loop AWG are always aligned, reducing the impact of temperature variations or manufacturing deviations on the computational results. This embodiment addresses the scalability and computational accuracy issues of existing photonic tensor kernels and the urgent need for computing power in AI computation. It proposes a novel on-chip photonic tensor kernel for AI computing hardware acceleration. By integrating data loading and convolution kernel element loading units on-chip, it effectively reduces the size and power consumption of the computing chip. Furthermore, the innovative design effectively reduces the impact of process deviations and environmental temperature changes on computational accuracy, making it suitable for mass production.

[0180] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 22As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores data to be processed and coefficient data. The network interface communicates with external terminals via a network connection. When executed by the processor, the computer program implements a photon tensor computation method.

[0181] Those skilled in the art will understand that Figure 22 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0182] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the method as described in any of the above embodiments, including the following steps:

[0183] A multi-wavelength optical signal is generated, and each wavelength of the optical signal is modulated separately through a first loop array waveguide grating, and then multiplexed to obtain a multi-channel multi-wavelength first optical signal; wherein each channel of the first optical signal is isolated from each other in the spectrum; a first modulation module is coupled on the loop path of the first loop array waveguide grating, which is used to modulate the optical signal of each wavelength according to the input data to be processed;

[0184] Each of the first optical signals is split to obtain multiple copies of the first optical signal. The optical signals of each wavelength in each copy of the first optical signal are modulated by multiple second loop array waveguide gratings and multiplexed to obtain a single second optical signal. A second modulation module is coupled on the loop path of the second loop array waveguide grating to modulate the optical signals of each wavelength according to the input coefficient data.

[0185] The second optical signal is converted into photoelectric signal to obtain tensor calculation results.

[0186] In one embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor includes a photonic tensor kernel as described in any of the above embodiments, the photonic tensor kernel comprising:

[0187] The light source unit is used to generate multi-wavelength light signals;

[0188] The parallel data loading optical engine includes a first loop-back array waveguide grating, which is used to modulate the optical signals of each wavelength in the multi-wavelength optical signals and then multiplex them to obtain multiple multi-wavelength first optical signals; wherein each first optical signal is isolated from each other in the spectrum; a first modulation module is coupled on the loop-back path of the first loop-back array waveguide grating, which is used to modulate the optical signals of each wavelength according to the input data to be processed;

[0189] The beam splitting unit is used to split each of the first optical signals to obtain multiple copies of the first optical signal;

[0190] The photonic dot product engine includes a second loop array waveguide grating, which is used to modulate the optical signals of each wavelength in a single copy of the first optical signal and multiplex them to obtain a single second optical signal; wherein, a second modulation module is coupled on the loop path of the second loop array waveguide grating, which is used to modulate the optical signals of each wavelength according to the input coefficient data;

[0191] The photoelectric detection module is used to perform photoelectric conversion on each of the second optical signals to obtain tensor calculation results.

[0192] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0193] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0194] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0195] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for calculating photon tensors, characterized in that, The method for calculating the photon tensor includes: A multi-wavelength optical signal is generated, and each wavelength of the optical signal is modulated separately through a first loop array waveguide grating, and then multiplexed to obtain a multi-channel multi-wavelength first optical signal; wherein each channel of the first optical signal is isolated from each other in the spectrum; a first modulation module is coupled on the loop path of the first loop array waveguide grating, which is used to modulate the optical signal of each wavelength according to the input data to be processed; Each of the first optical signals is split to obtain multiple copies of the first optical signal. The optical signals of each wavelength in each copy of the first optical signal are modulated by multiple second loop array waveguide gratings and multiplexed to obtain a single second optical signal. A second modulation module is coupled on the loop path of the second loop array waveguide grating to modulate the optical signals of each wavelength according to the input coefficient data. The second optical signal is converted into photoelectric signal to obtain tensor calculation results.

2. The photon tensor calculation method according to claim 1, characterized in that, The multiplexing process yields multiple, multi-wavelength first optical signals, including: The modulated optical signals of each wavelength are grouped according to a preset number of channels, with each group corresponding to one of the first optical signals; the frequency domain spacing between adjacent wavelengths in each of the first optical signals is the product of the frequency domain spacing between adjacent wavelengths in the multi-wavelength optical signals and the preset number of channels.

3. The photon tensor calculation method according to claim 1, characterized in that, Before modulating the optical signals of each wavelength in a single copy of the first optical signal, the method further includes: Spectral shaping is performed on a single copy of the first optical signal so that each wavelength component in the copy of the first optical signal falls within a preset flat-top passband.

4. The photon tensor calculation method according to claim 1, characterized in that, The data to be processed includes neural network activation values; the coefficient data includes convolution kernel weights; and the tensor calculation results include the output vector of matrix-vector multiplication.

5. A photon tensor kernel, characterized in that, The photon tensor kernel includes: The light source unit is used to generate multi-wavelength light signals; The parallel data loading optical engine includes a first loop-back array waveguide grating, which is used to modulate the optical signals of each wavelength in the multi-wavelength optical signals and then multiplex them to obtain multiple multi-wavelength first optical signals; wherein each first optical signal is isolated from each other in the spectrum; a first modulation module is coupled on the loop-back path of the first loop-back array waveguide grating, which is used to modulate the optical signals of each wavelength according to the input data to be processed; The beam splitting unit is used to split each of the first optical signals to obtain multiple copies of the first optical signal; The photonic dot product engine includes a second loop array waveguide grating, which is used to modulate the optical signals of each wavelength in a single copy of the first optical signal and multiplex them to obtain a single second optical signal; wherein, a second modulation module is coupled on the loop path of the second loop array waveguide grating, which is used to modulate the optical signals of each wavelength according to the input coefficient data; The photoelectric detection module is used to perform photoelectric conversion on each of the second optical signals to obtain tensor calculation results.

6. The photon tensor kernel according to claim 5, characterized in that, The first loopback array waveguide grating includes a first star coupler, a second star coupler, an array waveguide connecting the first star coupler and the second star coupler, and a first loopback path connecting the output side of the second star coupler and the input side of the first star coupler; The first modulation module is coupled to the first loopback path and is used to modulate the optical signal of each wavelength according to the input data to be processed, and send the modulated optical signal of each wavelength back to the first star coupler.

7. The photon tensor kernel according to claim 5, characterized in that, The second loop array waveguide grating includes a third star coupler, a fourth star coupler, an array waveguide connecting the third star coupler and the fourth star coupler, and a second loop path connecting the output side of the fourth star coupler and the input side of the third star coupler; The second modulation module is coupled to the second loop path and is used to modulate the optical signal of each wavelength according to the input coefficient data, and send the modulated optical signal of each wavelength back to the third star coupler.

8. The photon tensor kernel according to claim 7, characterized in that, The second loop array waveguide grating also includes a flat-top spectral response input unit; The third star coupler acquires a copy of the first optical signal after spectral shaping through the flat-top spectral response input unit; each wavelength component in the first optical signal copy falls within a preset flat-top passband.

9. The photon tensor kernel according to claim 8, characterized in that, The flat-top spectral response input unit employs a Mach-Zehnder interferometer or a multimode interferometer.

10. The photon tensor kernel according to claim 5, characterized in that, The multiplexing process yields multiple, multi-wavelength first optical signals, including: The modulated optical signals of each wavelength are grouped according to a preset number of channels, with each group corresponding to one of the first optical signals; the frequency domain spacing between adjacent wavelengths in each of the first optical signals is the product of the frequency domain spacing between adjacent wavelengths in the multi-wavelength optical signals and the preset number of channels.

11. The photon tensor kernel according to claim 10, characterized in that, The beam splitting unit includes multiple beam splitters; the number of beam splitters corresponds to the number of preset channels; each beam splitter is used to split the corresponding first optical signal to obtain multiple copies of the first optical signal; the number of copies of the first optical signal is equal to the number of photon multiplication engines.

12. The photon tensor kernel according to claim 5, characterized in that, Both the first modulation module and the second modulation module adopt one of the following: Mach-Zehnder interference structure, resonant micro-ring structure, and electro-absorption modulation structure.

13. The photon tensor kernel according to claim 5, characterized in that, The light source unit includes at least one of the following: superradiative diode light source, multi-wavelength laser light source, and optical frequency comb light source.

14. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, The processor includes a photonic tensor core as described in any one of claims 5 to 13.

15. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 4.

Images