Optical computing system and electronic device

Abstract

This application discloses an optical computing system and electronic device, applied in the fields of optical computing and artificial intelligence technology. To solve the problems of low architecture integration, high overall power consumption, and easy generation of operational crosstalk, an optical frequency comb light source is used to output optical signals of various wavelengths; a micro-ring modulator array is used to modulate and encode the received optical signals of various wavelengths to determine the optical signals corresponding to each element in the input matrix; an arrayed waveguide grating router is used to synchronously transmit the optical signals corresponding to each element in the input matrix; a Mach-Zehnder interferometer array is used to encode the weight matrix, and based on the encoded weight matrix and the optical signals corresponding to each element in the input matrix, the multiplication operation between the weights and the matrix elements is performed to obtain the corresponding multiplication results; a photodetector array is used to perform photoelectric conversion on the optical signals after optical domain accumulation based on each multiplication result to obtain the output result; the system features high integration, low power consumption, and low crosstalk.

Description

Technical Field

[0001] This application relates to the fields of optical computing and artificial intelligence technology, and in particular to an optical computing system and electronic device. Background Technology

[0002] With the rapid development of artificial intelligence technology, pre-trained language models such as convolutional neural networks and Transformers are widely used. Matrix multiplication, as its core fundamental operation, directly impacts the processing speed and energy consumption of tasks. Optical computing, based on the inherent parallelism, low loss, and low energy consumption of photons, has become the preferred solution for achieving high-speed matrix operations. Among them, optical computing architectures based on Mach-Zehnder interferometers (MZIs) have outstanding speed and energy consumption advantages in matrix operations and image processing. However, optical computing matrix multiplication in related technologies typically implements matrix multiplication operations by performing multiple sets of matrix-vector multiplications in parallel. This architecture suffers from low integration, high overall power consumption, and is prone to computational crosstalk. Summary of the Invention

[0003] This application provides an optical computing system and electronic device to at least solve the problems of low integration, high power consumption, and easy generation of computational crosstalk in related technologies.

[0004] This application provides an optical computing system, including: An optical frequency comb light source is used to output optical signals of various wavelengths. A micro-ring modulator array connected to an optical frequency comb light source is used to modulate and encode various received optical signals of different wavelengths in order to determine the optical signal corresponding to each element in the input matrix. An arrayed waveguide grating router connected to a micro-ring modulator array is used to synchronously transmit the optical signals corresponding to each element in the input matrix. The Mach-Zehnder interferometer array connected to the arrayed waveguide grating router is used to encode the weight matrix, and to perform multiplication operations between the weights and matrix elements based on the encoded weight matrix and the optical signals corresponding to each element in the input matrix, so as to obtain the corresponding multiplication results. The photoelectric detection array connected to the Mach-Zehnder interferometer array is used to perform photoelectric conversion on the optical signal after optical domain accumulation based on each multiplication result, and obtain the output result.

[0005] This application also provides a photon matrix calculation method applied to the aforementioned optical computing system, the method comprising: A light source with an optical frequency comb can be used to output optical signals of various wavelengths. A micro-ring modulator array is used to modulate and encode the received optical signals of various wavelengths in order to determine the optical signal corresponding to each element in the input matrix. An arrayed waveguide grating router is used to synchronously transmit the optical signals corresponding to each element in the input matrix. The weight matrix is ​​encoded using a Mach-Zehnder interferometer array, and the multiplication operation between the weight and the matrix element is performed based on the encoded weight matrix and the optical signal corresponding to each element in the input matrix to obtain the corresponding multiplication result. The optical signal, after optical domain accumulation based on each multiplication result, is converted into an optical signal using a photoelectric detection array to obtain the output result.

[0006] This application also provides an electronic device, including the optical computing system described above.

[0007] As can be seen from the above technical solution, the beneficial effects of this application are as follows: This application provides an optical computing system, comprising: an optical frequency comb light source for outputting optical signals of various wavelengths; a micro-ring modulator array for modulating and encoding the received optical signals of various wavelengths to determine the optical signals corresponding to each element in an input matrix; an arrayed waveguide grating router for synchronously transmitting the optical signals corresponding to each element in the input matrix; a Mach-Zehnder interferometer array for encoding a weight matrix and performing multiplication operations between the weights and matrix elements based on the encoded weight matrix and the optical signals corresponding to each element in the input matrix to obtain the corresponding multiplication results; and a photodetector array for photoelectric conversion of the optical signals accumulated in the optical domain based on the multiplication results to obtain the output results.

[0008] The optical computing system provided in this application outputs multiple optical signals of different wavelengths through an optical frequency comb light source. Combined with a micro-ring modulator array, each wavelength optical signal is independently modulated and encoded to determine the optical signal corresponding to each element in the input matrix. This achieves high-density parallel encoding of the input matrix elements, avoiding the problems of device redundancy and increased crosstalk caused by repeated stacking of optical paths in parallel architectures of multiple matrix-vector multiplications. This application uses an arrayed waveguide grating router to synchronously transmit the optical signals corresponding to each element in the input matrix. Utilizing its passive wavelength routing characteristics, it achieves low-loss, low-crosstalk parallel transmission of multi-wavelength optical signals, eliminating the need for additional tuning devices and reducing system power consumption. A Mach-Zehnder interferometer array directly performs weight matrix encoding and multiplication operations in the optical domain, fully utilizing the natural parallel computing capabilities of photons and significantly improving the speed of matrix multiplication. Finally, a photoelectric detector array performs photoelectric conversion on the accumulated optical signals in the optical domain to obtain the output result, achieving accurate reading of the optical computing results. Therefore, this application effectively solves the problems of low integration, high power consumption, and easy generation of computational crosstalk in optical computing architectures in related technologies, and realizes photonic matrix multiplication operations with high integration, low power consumption, low crosstalk, and high speed.

[0009] In addition, this application also provides an electronic device with corresponding advantages. Attached Figure Description

[0010] To more clearly illustrate the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a schematic diagram of the structure of an optical computing system provided in an embodiment of this application; Figure 2 This is a schematic diagram of another optical computing system provided in an embodiment of this application; Figure 3 An architecture diagram of an optical computing system provided in this application embodiment; Figure 4 This is a schematic diagram of the structure of a silicon-based MZI modulator provided in an embodiment of this application; Figure 5 A schematic diagram of a silicon-based MZI modulator array corresponding to a 4×4 weight matrix is ​​provided in an embodiment of this application; Figure 6 A flowchart illustrating a photon matrix calculation method provided in this application embodiment; Figure 7 This is a structural block diagram of a photonic matrix computing device provided in an embodiment of this application. Detailed Implementation

[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of this application.

[0013] It should be noted that, in the description of this application, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. The terms "first," "second," etc., in this application are used to distinguish similar objects and are not used to describe a specific order or sequence.

[0014] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0015] It should be noted that optical computing, with its inherent parallel propagation of photons and low energy consumption, far surpasses traditional electronic computing in terms of computational speed and energy density, providing a new approach for high-speed matrix multiplication in the optical domain. In particular, optical computing architectures based on Mach-Zehnder interferometers have proven to have high computational efficiency and low energy consumption in matrix-vector multiplication and convolution operations. This allows for rapid matrix transformation and filtering of acoustic features in speech recognition, increasing response speed several times compared to traditional processors and enabling real-time feature extraction from high-resolution images. Inference latency on the ImageNet dataset (a large visualization database for visual object recognition software research) is reduced to microseconds, with energy consumption only one-thousandth that of a GPU. However, optical computing solutions for matrix multiplication still have several drawbacks: the architectures mainly rely on multiple sets of matrix-vector multiplications in parallel, resulting in large size, high power consumption, and increased crosstalk; the use of microring resonators to assist MZI in parallel matrix multiplication limits stability and scalability, and does not fully utilize multidimensional multiplexing technology, leading to device redundancy and low wavelength utilization.

[0016] In view of this, this application provides a highly integrated, low-power, low-crosstalk, and high-speed optical computing system. This application addresses matrix multiplication operations involved in pre-trained language models such as neural networks, convolutional neural networks, and Transformers. The optical computing system first uses an optical frequency comb source to output optical signals of different wavelengths. Next, the optical signals enter a microring modulator (MRM) array, where each wavelength is modulated via a thermo-optical effect to encode the elements of the input matrix X. Then, the optical signals are passively routed through an arrayed waveguide grating router (AWGR), synchronously transmitting the encoded optical signals to a Mach-Zehnder interferometer (MZI) array used to encode the weight matrix W. Upon receiving the optical signals, the MZI array adjusts the splitting ratio of each MZI modulator to perform multiplication operations between the weights and the corresponding input elements. Finally, the multiplication results are summed in the optical domain and output as matrix multiplication results after photoelectric conversion by a photodetector array. Y = W × X This optical computing system performs matrix multiplication operations in parallel through wavelength multiplexing, MRM-encoded input matrices, AWGR passive device routing of propagation wavelengths, and MZI-encoded weight matrices. Compared to related technologies, this invention solves the problems of low integration, high power consumption, and weak scalability. By leveraging the inherent parallelism and low-loss characteristics of photons, it achieves a balance between high precision, high speed, and low power consumption.

[0017] The technical solution will be explained and introduced in detail below. Please refer to... Figure 1 The diagram shown illustrates the structure of an optical computing system, which includes: Optical frequency comb light source 1 is used to output optical signals of various wavelengths; The micro-ring modulator array 2, connected to the optical frequency comb light source 1, is used to modulate and encode the received optical signals of various wavelengths in order to determine the optical signals corresponding to each element in the input matrix. The arrayed waveguide grating router 3, connected to the micro-ring modulator array 2, is used to synchronously transmit the optical signals corresponding to each element in the input matrix. The Mach-Zehnder interferometer array 4, connected to the arrayed waveguide grating router 3, is used to encode the weight matrix and perform multiplication operations between the weights and matrix elements based on the encoded weight matrix and the optical signals corresponding to each element in the input matrix, to obtain the corresponding multiplication results. The photoelectric detector array 5, connected to the Mach-Zehnder interferometer array 4, is used to perform photoelectric conversion on the optical signal after optical domain accumulation based on each multiplication result, to obtain the output result.

[0018] It should be noted that the optical frequency comb light source 1 in this embodiment can generate optical signals of various wavelengths. This optical frequency comb light source 1 is a technique that arranges optical frequencies at uniform intervals, based on the principle of ultrashort pulse laser generation and frequency control. In practical applications, the optical frequency comb light source 1 can be an electro-optic modulated optical frequency comb, for example, an electro-optic modulated optical frequency comb with a center wavelength of 1550nm. Furthermore, the various optical signals of different wavelengths generated by the optical frequency comb light source 1 are co-originating and phase-synchronized optical signals, and each optical signal supports parallel computation to support coherent accumulation operations in the optical domain.

[0019] In this embodiment, the optical frequency comb light source 1 transmits various optical signals of different wavelengths to the micro-ring modulator array 2. The micro-ring modulator array 2 modulates the received optical signals of different wavelengths and encodes each element in the input matrix to determine the optical signal corresponding to each element in the input matrix. The MRM is an optical modulation device based on the micro-ring resonance effect. The core of the MRM consists of a ring waveguide coupled with a straight waveguide, and it mainly modulates the light intensity based on the resonant enhancement and suppression characteristics of light.

[0020] Furthermore, in this embodiment, the micro-ring modulator array 2 transmits the modulated and coded optical signal to the arrayed waveguide grating router 3. The arrayed waveguide grating router 3 can synchronously transmit optical signals of different wavelengths to the Mach-Zehnder interferometer array 4. The arrayed waveguide grating router 3 is the core router component in the photonic computing system. It utilizes the optical path difference of the arrayed waveguide to achieve a deterministic mapping between wavelength and port, essentially a wavelength-selective router based on diffraction and interference, capable of simultaneously separating, routing, and transmitting multiple wavelength signals in parallel. In practical applications, the arrayed waveguide grating router 3 is a passive device, requiring no additional tuning components. It can synchronously transmit optical signals of different wavelengths to the input of the Mach-Zehnder interferometer array 4 with low crosstalk and low loss. The routing characteristics of the arrayed waveguide grating router 3 effectively avoid the exponential attenuation and cascading loss of optical power as the matrix dimension increases, effectively improving parallelism.

[0021] In this embodiment, the Mach-Zehnder interferometer array 4 receives optical signals transmitted from the arrayed waveguide grating router 3, each corresponding to an element in the input matrix. The Mach-Zehnder interferometer array 4 can encode a weight matrix and perform multiplication operations between the weights and matrix elements based on the encoded weight matrix and the optical signals corresponding to each element in the input matrix, obtaining the corresponding multiplication results. Multiple multiplication results can be obtained. The MZI in this embodiment is an optical silicon-based device based on the principle of optical interference, such as an optical switch or modulator. It has wide applications in optical computing and optical communication, enabling precise control of the amplitude and phase of optical signals to perform complex mathematical operations. In practical applications, multiple MZI modulators can be combined into an MZI array and their phases can be electro-optically or thermo-optically modulated to correspond to the elements of the weight matrix W. Furthermore, when the Mach-Zehnder interferometer array 4 receives the optical signals transmitted from the arrayed waveguide grating router 3, the weights can be multiplied by the input elements in the optical domain through the beam splitting ratio.

[0022] It is understood that the photoelectric detection array 5 in this embodiment is connected to the output end of the Mach-Zehnder interferometer array 4, and can perform photoelectric conversion on the optical signal after optical domain accumulation based on each multiplication result to obtain the output result, wherein the final output result is the column summation result.

[0023] Therefore, in this embodiment, multiple optical signals of different wavelengths are output from an optical frequency comb light source, and each wavelength optical signal is independently modulated and encoded by a micro-ring modulator array to determine the optical signal corresponding to each element in the input matrix. This achieves high-density parallel encoding of the input matrix elements, avoiding the problem of device redundancy and increased crosstalk caused by repeated stacking of optical paths in the parallel architecture of multiple matrix vector multiplications. This application synchronously transmits the optical signals corresponding to each element in the input matrix through an arrayed waveguide grating router, utilizing its passive wavelength routing characteristics to achieve low-loss, low-crosstalk parallel transmission of multi-wavelength optical signals without the need for additional tuning devices, thus reducing system power consumption. The weight matrix encoding and multiplication operations are directly completed in the optical domain through a Mach-Zehnder interferometer array, making full use of the natural parallel computing capability of photons and significantly improving the operation speed of matrix multiplication. Finally, the output result is obtained by photoelectric conversion of the accumulated optical signal in the optical domain through a photodetector array, realizing accurate reading of the optical calculation result. This application achieves high integration, low power consumption, low crosstalk, and high speed photonic matrix multiplication operations.

[0024] Please refer to the following. Figure 2 and Figure 3 The following are structural diagrams of another optical computing system. The embodiments of this application will further explain and optimize the technical solution.

[0025] In one embodiment, the optical computing system may further include an optical beam splitter 6 disposed between the optical frequency comb light source 1 and the micro-ring modulator array 2. The optical beam splitter 6 is used to split the received optical signals of various wavelengths into multiple beams; each beam-splitter includes optical signals of all wavelengths. The micro-ring modulator array 2 is used to modulate and encode the optical signals of various wavelengths in each beam split optical signal to determine the optical signals corresponding to each element in the input matrix.

[0026] It should be noted that in this embodiment, the optical beam splitter 6 is connected between the output end of the optical frequency comb light source 1 and the input end of the micro-ring modulator array 2. The optical signals of different wavelengths output by the optical frequency comb light source 1 first enter the optical beam splitter 6. The optical signals of different wavelengths can be output to the optical beam splitter 6 in a single transmission manner. The optical beam splitter 6 evenly splits these combined optical signals containing all wavelengths into multiple optical signals, so that each optical signal contains optical components of all wavelengths, and the power of each optical signal remains consistent.

[0027] Understandably, the micro-ring modulator array 2 receives the multi-beam split optical signals transmitted by the optical beamsplitter 6. Each split optical signal contains optical signals of all wavelengths. In practical applications, the micro-ring modulator array 2 can include multiple micro-ring modulators, so each micro-ring modulator in the array 2 can simultaneously receive optical signals of all wavelengths. Each micro-ring modulator can independently modulate and encode optical signals of different wavelengths, thereby encoding the element values ​​of the corresponding rows and columns in the input matrix onto the optical signals of the corresponding wavelengths. Since the split optical signals are simultaneously input to each micro-ring modulator, and each micro-ring modulator can process optical signals of different wavelengths in parallel, all elements in the input matrix can be encoded simultaneously, achieving highly parallel data loading.

[0028] In practical applications, the optical beam splitter 6 in this embodiment can adopt a passive beam splitting structure, which does not require additional control circuitry, thus reducing system power consumption and complexity.

[0029] In other words, in this embodiment, by setting an optical beamsplitter 6 between the optical frequency comb light source 1 and the micro-ring modulator array 2, the combined optical signal containing all wavelengths is uniformly split into multiple identical optical signals, which are then input to each micro-ring modulator in the micro-ring modulator array 2. Therefore, each micro-ring modulator can receive an optical signal with the same power and containing all wavelengths, ensuring that the encoding reference of each row of the input matrix is ​​consistent and improving the encoding accuracy. In addition, the parallel input of each optical signal to each micro-ring modulator allows all elements of the input matrix to be encoded simultaneously, fully leveraging the parallel advantages of optical computing.

[0030] In one embodiment, the optical computing system may further include a wavelength division multiplexer 7 disposed between the optical frequency comb light source 1 and the optical beam splitter 6. Wavelength division multiplexer 7 is used to combine multiple optical signals of different wavelengths output from optical frequency comb light source 1 into the same optical path and transmit them to optical beam splitter 6; The optical beam splitter 6 is used to split optical signals of different wavelengths in the combined optical signal to obtain multi-beam split optical signals.

[0031] It should be noted that the optical frequency comb light source 1 in this embodiment can output multiple optical signals of different wavelengths. These optical signals can be multiple independent optical signals, each carrying a specific wavelength of light. In practical applications, the optical comb teeth generating different wavelengths in the optical frequency comb light source 1 may be spatially separated, or need to be output separately through different optical fibers or waveguides. In this embodiment, in order to effectively transmit these different wavelength optical signals to the subsequent optical beam splitter 6 and micro-ring modulator array 2, these different wavelength optical signals can be combined into the same optical path first.

[0032] It is understood that in this embodiment, a wavelength division multiplexer 7 can be provided between the optical frequency comb light source 1 and the optical beam splitter 6. The wavelength division multiplexer 7 enables the use of optical elements such as gratings or thin-film filters to combine multiple optical signals of different wavelengths from different input ports to the same output port, allowing multiple optical signals of different wavelengths to be transmitted in parallel in the same optical fiber or the same waveguide without interference. Since optical signals of different wavelengths have different frequencies physically, they do not experience coherent crosstalk when propagating in the medium. Therefore, the wavelength division multiplexer 7 can efficiently combine multiple optical signals of different wavelengths into a single combined optical signal.

[0033] Furthermore, in this embodiment, the combined optical signal after wavelength division multiplexing (WDM) by the wavelength division multiplexer 7 contains optical components of all wavelengths. This combined optical signal is transmitted to the optical beamsplitter 6. After receiving the combined optical signal, the optical beamsplitter 6 uses the beam splitting characteristics of the combined optical signal to uniformly split the combined optical signal into multiple optical signals. Since the combined optical signal already contains optical components of all wavelengths, each split optical signal completely contains optical signals of all wavelengths, and the power of each optical signal remains consistent. The split multiple optical signals are transmitted to the micro-ring modulator array 2. In practical applications, the micro-ring modulator array 2 may include multiple micro-ring modulators. Therefore, the split multiple optical signals can be input to each micro-ring modulator in the micro-ring modulator array 2 for subsequent modulation and encoding processing.

[0034] It should also be noted that in this embodiment, the wavelength division multiplexer 7 and the optical beamsplitter 6 cooperate to first combine multiple optical signals of different wavelengths, and then demultiplex the combined optical signal. In other words, the wavelength division multiplexer 7 combines multiple independent wavelength optical signals into a single combined optical signal, reducing the number of optical paths; the optical beamsplitter 6 then re-divides this combined optical signal into multiple optical signals containing all wavelengths, providing them to each micro-ring modulator. Compared to directly transmitting multiple independent wavelength signals to each micro-ring modulator, the implementation in this embodiment reduces the number of optical fibers or waveguides between the light source and the modulator array, simplifies system wiring, ensures power balance and phase consistency among the optical signals, improves coding accuracy, and facilitates system integration and packaging.

[0035] This embodiment uses a 4×4 input matrix as an example for illustration. When the input matrix is ​​a 4×4 matrix, the optical frequency comb light source 1 generates four different optical signals with wavelengths λ1, λ2, λ3, and λ4. The optical frequency comb light source 1 transmits the generated four different wavelength optical signals λ1, λ2, λ3, and λ4 to the wavelength division multiplexer 7. The wavelength division multiplexer 7 performs multiplexing processing on the four different wavelength optical signals λ1, λ2, λ3, and λ4 to obtain a multiplexed optical signal (λ1, λ2, λ3, λ4). This multiplexed optical signal is then propagated in parallel to the optical beam splitter 6 through an optical fiber, thereby reducing the number of optical paths. Accordingly, the optical beam splitter 6 in this embodiment can be a 1×4 beam splitter. The 1×4 beam splitter evenly divides the received combined optical signal into four identical optical signals, which are output to the four micro-ring modulators in the micro-ring modulator array 2 respectively. That is, the optical signal received by each micro-ring modulator contains four wavelengths λ1, λ2, λ3, and λ4.

[0036] In one embodiment, the microring modulator array 2 includes a plurality of microring modulators 21, the number of which is the same as the number of rows of the input matrix. Each microring modulator 21 has a plurality of independent channels, the number of which is the same as the number of columns of the input matrix. Each channel is used for resonant modulation of an optical signal of one wavelength.

[0037] It should be noted that the micro-ring modulator array 2 in this embodiment can be composed of multiple micro-ring modulators 21, and the number of micro-ring modulators 21 is strictly equal to the number of rows of the input matrix. For example, when the input matrix X is an N×N (e.g., 4×4) matrix, the number of rows of the input matrix is ​​N. The micro-ring modulator array 2 can include N micro-ring modulators 21, and the N micro-ring modulators 21 are used to encode the elements of the first row to the Nth row of the input matrix, so that each row of the input matrix corresponds to an independent micro-ring modulator 21, realizing row-by-row parallel encoding.

[0038] Each micro-ring modulator 21 in this embodiment has multiple independent channels. The number of channels is equal to the number of columns in the input matrix. If the input matrix is ​​an N×N matrix, then the number of channels is N. Taking a 4×4 matrix as an example, the number of columns in the input matrix is ​​4, so each micro-ring modulator 21 has 4 independent channels. Each channel is used for resonant modulation of an optical signal of a specific wavelength. For example, for each micro-ring modulator MRM... m Where m = 1, 2, 3, 4, and receives optical signals containing four wavelengths λ1, λ2, λ3, and λ4, where any MRM m In the diagram, channel A corresponds to wavelength λ1, channel B corresponds to wavelength λ2, channel C corresponds to wavelength λ3, and channel D corresponds to wavelength λ4. That is to say, channel A resonates only with optical signals of wavelength λ1, channel B resonates only with optical signals of wavelength λ2, channel C resonates only with optical signals of wavelength λ3, and channel D resonates only with optical signals of wavelength λ4.

[0039] It is understood that in the embodiments of this application, each channel is resonantly modulated to correspond to the optical signal of the corresponding wavelength, which can ensure that each channel can independently encode an element in the input matrix.

[0040] It should also be noted that the micro-ring modulator 21 can be constructed by coupling a ring waveguide and a straight waveguide. When an optical signal propagates through the straight waveguide, only optical signals whose wavelength matches the resonant wavelength of the micro-ring will be coupled into the micro-ring and undergo resonant enhancement, while optical signals of other wavelengths will pass through unaffected. Therefore, in practical applications, multiple independent ring waveguides can be set in each micro-ring modulator 21 as multiple independent channels, and the resonant wavelength of each channel can be tuned to different target wavelengths, thereby achieving independent modulation of optical signals of different wavelengths.

[0041] Furthermore, since each micro-ring modulator 21 is independent of the others, and the channels within each micro-ring modulator 21 are also independent of each other, all channels of all micro-ring modulators 21 can be modulated simultaneously. In other words, all elements of the input matrix can be quickly encoded into the corresponding wavelength optical signal in parallel, without serial loading, fully leveraging the parallel processing advantages of optical computing. For example, for an N×N input matrix, the four micro-ring modulators 21 simultaneously start modulation to independently encode the 16 elements of the input matrix X, without affecting each other.

[0042] It is understood that in this embodiment, the number of micro-ring modulators 21 is the same as the number of rows in the input matrix, and the number of channels in each micro-ring modulator 21 is set to be the same as the number of columns in the input matrix. A unique correspondence between channels and wavelengths is established, thereby enabling a deterministic, one-to-one mapping between input matrix elements and optical signals. The optical computing system provided in this embodiment enables all elements of the input matrix to be encoded in parallel onto optical signals of different wavelengths, greatly improving data loading speed. Moreover, each channel is independent and does not interfere with each other, effectively avoiding crosstalk problems that may occur during parallel encoding, ensuring encoding accuracy and reliability. Combined with the advantages of low power consumption and fast modulation speed of the micro-ring modulator 21, this embodiment is beneficial to improving the system's integration and energy efficiency ratio.

[0043] In one implementation, the optical beam splitter 6 is a 1×N beam splitter, where N is equal to the number of rows in the input matrix; The optical beam splitter 6 is used to uniformly split the received optical signal into N identical optical signals, and input the N identical optical signals to the N micro-ring modulators in the micro-ring modulator array 2 respectively.

[0044] It should be noted that the optical beamsplitter 6 in this embodiment adopts a 1×N beamsplitter structure, with 1 input port and N output ports, where N equals the number of rows in the input matrix X. For example, when the input matrix X is a 4×4 matrix, the number of rows in the input matrix is ​​4. Therefore, the optical beamsplitter 6 adopts a 1×4 beamsplitter, that is, it has 1 input port and 4 output ports. The combined optical signal received by the optical beamsplitter 6 contains optical components of all wavelengths (e.g., λ1, λ2, λ3, λ4). The optical beamsplitter 6 distributes the input optical power to each output port in an equal proportion, so that the optical signal power output by each output port is equal, which is 1 / N of the input optical power. Thus, the received optical signal can be uniformly split into N identical optical signals.

[0045] It should also be noted that each output port of the optical beamsplitter 6 is connected to a micro-ring modulator 21 in the micro-ring modulator array 2. In this embodiment, an example with a 4×4 input matrix is ​​used. Assume that the first output port of the optical beamsplitter 6 is connected to the first micro-ring modulator MRM1, the second output port to the second micro-ring modulator MRM2, the third output port to the third micro-ring modulator MRM3, and the fourth output port to the fourth micro-ring modulator MRM4. Since the optical signals at each output port are completely identical in power and all contain optical components of all wavelengths, the optical signal reference received by each micro-ring modulator 21 is exactly the same. That is, in this embodiment, the optical beamsplitter 6 can simultaneously distribute the combined optical signal to all micro-ring modulators 21, allowing all micro-ring modulators 21 to modulate their received optical signals in parallel, thereby achieving simultaneous encoding of all elements of the input matrix.

[0046] In this embodiment, on the one hand, the combined optical signal is evenly distributed to the N micro-ring modulators 21, ensuring that the optical signal power received by each micro-ring modulator 21 is consistent and eliminating encoding errors caused by differences in optical power. On the other hand, since the optical beam splitter 6 is a passive device, it does not consume additional power, and the optical signals after beam splitting arrive at each micro-ring modulator 21 simultaneously, parallel encoding of all elements of the input matrix is ​​achieved, effectively avoiding the delay overhead caused by serial loading.

[0047] In one embodiment, the micro-ring modulator 21 includes a heater and a modulation electrode; The heater is used to tune the resonant wavelength of each channel to the target wavelength so that each channel resonates only with the optical signal of the target wavelength among various wavelength optical signals; The modulation electrode is used to adjust the light intensity of the optical signal of the corresponding wavelength according to the value of each element in the input matrix.

[0048] It is understood that the micro-ring modulator 21 in this embodiment integrates a heater and a modulation electrode, and wavelength selection and data encoding are achieved through the heater and the modulation electrode, respectively.

[0049] In practical applications, the heater in this embodiment can be a titanium nitride heater, which can be integrated near the ring waveguide (i.e., channel) of the micro-ring modulator 21. By applying a specific voltage or current to the heater, heat is generated, and the temperature of the ring waveguide rises accordingly, thereby changing the refractive index of the waveguide material (i.e., thermo-optic effect). The change in refractive index causes the resonant wavelength of the micro-ring to drift. Therefore, the resonant wavelength of each channel can be tuned to a preset target wavelength by controlling the electrical power applied to the heater. For example, during the system initialization phase, an appropriate voltage is applied to the heater of the first channel of the first micro-ring modulator MRM1 to lock its resonant wavelength at λ1; another voltage is applied to the heater of the second channel of MRM1 to lock its resonant wavelength at λ2; another voltage is applied to the heater of the third channel of MRM1 to lock its resonant wavelength at λ3; and another voltage is applied to the heater of the third channel of MRM1 to lock its resonant wavelength at λ4. In this way, each channel can generate a resonant response only to a specific wavelength of optical signal, while having essentially no response to other wavelengths of optical signal.

[0050] Furthermore, the modulation electrode in this embodiment can be used to implement data encoding. That is, after the resonant wavelength of each channel has been locked to the target wavelength by the heater, the modulation electrode can dynamically adjust the effective refractive index of the microring by applying an external voltage, thereby controlling the transmission efficiency (i.e., the degree of light intensity attenuation) of a specific wavelength optical signal. For example, when each MRM k When k=1,2,3,4, optical signals containing four wavelengths λ1, λ2, λ3, and λ4 are received, resonance occurs in the micro-ring tuning section. That is, channel A only resonates with optical signals of wavelength λ1. The column elements x of the independent encoding input matrix X are used to encode the input matrix X. k1 Channel B independently encodes the column elements x of the input matrix X for optical signals with wavelength λ2 only. k2 Channel C resonates only with optical signals of wavelength λ3, independently encoding the column elements x of the input matrix X. k3 Channel D independently encodes the column elements x of the input matrix X for optical signals with wavelength λ4 only. k4 .

[0051] In practical applications, the values ​​of each element in the input matrix X are converted into corresponding driving voltages and applied to the modulation electrodes of the corresponding channels. The values ​​of each element are normalized and mapped to the [0,1] interval. This driving voltage determines the degree of modulation of the optical signal at the resonant wavelength by that channel. For example, when it is necessary to encode the input matrix element x... 11 When the voltage is 0.3, the system applies a corresponding voltage to the modulation electrode of the first channel of the first micro-ring modulator MRM1, so that the transmittance of the channel to the optical signal with wavelength λ1 is 0.3.

[0052] The micro-ring modulator 21 in this embodiment includes a heater and a modulation electrode, enabling the separation of wavelength selection and data encoding functions. The heater is used for wavelength configuration during system initialization, locking each channel to the target wavelength and ensuring a one-to-one correspondence between channels and wavelengths. The modulation electrode is used for real-time data encoding, rapidly modulating the light intensity based on the values ​​of elements in the input matrix. This embodiment ensures both the stability and accuracy of wavelength selection and high-speed data loading, without interference between the two.

[0053] In one embodiment, the micro-ring modulator array 2 further includes a multimode interference combiner 22 that is connected one-to-one with each micro-ring modulator 21; The multimode interference combiner 22 is used to combine the optical signals output from each channel of the micro-ring modulator 21, which carry different wavelengths and have been encoded into the input matrix elements, into a single combined optical signal, and transmit the combined optical signal to the corresponding input terminal of the arrayed waveguide grating router 3.

[0054] It should be noted that the multimode interference combiner 22 in this embodiment can be a passive optical device based on the multimode interference effect. Each microring modulator 21 (corresponding to one row of the input matrix) is equipped with one multimode interference combiner 22, and the multimode interference combiner 22 is connected between the output of the microring modulator 21 and the input of the arrayed waveguide grating router 3. This embodiment uses a 4×4 matrix as an example. The microring modulator array 2 includes four microring modulators 21, each with four independent channels, thus requiring four multimode interference combiners 22. It can be understood that after each channel of the microring modulator 21 completes modulation encoding, taking the first microring modulator MRM1 as an example, its four channels output optical signals carrying wavelengths λ1, λ2, λ3, and λ4 respectively, with light intensities corresponding to the four elements x in the first row of the input matrix. 11 x 12 x 13 x 14 These four optical signals enter the four input ports of the multimode interferometric combiner 22 respectively. The multimode interferometric combiner 22 uses the coherent superposition and mode matching characteristics of multimode interference to combine the four optical signals into one combined optical signal without crosstalk. The combined optical signal is then transmitted to the corresponding input port of the arrayed waveguide grating router 3 with low loss, so that the data encoding information carried by different wavelengths is completely and synchronously sent into the subsequent arrayed waveguide grating router 3, realizing wavelength multiplexing and parallel delivery of matrix row elements.

[0055] In this embodiment, the combined optical signal contains optical components of all four wavelengths, and the intensity of each wavelength carries information about the corresponding input matrix element. Since the optical signals of different wavelengths generated by the optical frequency comb source 1 are from the same source and are phase-synchronized, the coherence between each wavelength is predictable. By configuring the waveguide structure, the multimode interferometric combiner 22 can effectively couple optical signals of different wavelengths into the output waveguide without generating significant crosstalk or energy loss.

[0056] It should also be noted that the combined optical signal output by the multimode interferometric multiplexer 22 is transmitted to the corresponding input port of the arrayed waveguide grating router 3. For example, the multimode interferometric multiplexer 22 corresponding to the first micro-ring modulator MRM1 outputs the combined optical signal to the first input port of the arrayed waveguide grating router 3, the multiplexer 22 corresponding to the second micro-ring modulator MRM2 outputs to the second input port, and so on. The combined optical signal received at each input port completely contains the information of the corresponding input matrix row elements, and different row elements are transmitted in parallel to the arrayed waveguide grating router 3, distinguished by wavelength.

[0057] In other words, by setting a one-to-one corresponding multimode interferometric combiner 22 at the output end of each micro-ring modulator 21, optical signals carrying different wavelengths output from different channels (corresponding to different columns of the matrix) within the same row (same micro-ring modulator) can be combined into a single combined optical signal. This allows parallel optical signals from multiple channels in the same row to be channeled into the same waveguide, significantly reducing the number of optical paths between the encoding array and the routing array. Furthermore, optical signals of different wavelengths do not experience crosstalk during the combining process, and their respective encoded information is fully preserved. Taking an N×N matrix as an example, if parallel transmission is used, N×N independent optical paths are required, while this embodiment only requires N optical paths, greatly simplifying product wiring and packaging complexity. In addition, the multimode interferometric combiner 22 has the advantages of low loss and low crosstalk, ensuring the integrity of the encoded optical signal during the combining process. Moreover, it is a passive device that does not consume additional power, which is beneficial to ensuring the overall energy efficiency of the optical computing system.

[0058] In one embodiment, the number of wavelength types in the multiple different wavelength optical signals output by the optical frequency comb light source 1 is equal to the number of columns in the input matrix; The number of independent channels in each micro-ring modulator 21 is equal to and corresponds one-to-one with the number of wavelength types. Each channel is used to resonate and modulate an optical signal of a corresponding wavelength.

[0059] It should be noted that in this embodiment, the number of wavelength types output by the optical frequency comb light source 1 is equal to the number of columns in the input matrix X. For example, when the input matrix X is a 4×4 matrix, it has 4 columns. Therefore, the optical signal output by the optical frequency comb light source 1 contains 4 different wavelengths, denoted as λ1, λ2, λ3, and λ4 respectively. Each wavelength corresponds to one column in the input matrix X, that is, wavelength λ1 corresponds to all elements in the first column (x...). 11 x 21 x 31 x 41 ), wavelength λ2 corresponds to all elements (x) in the second column. 12 x 22 x 32 x 42 And so on. The number of independent channels in each micro-ring modulator 21 is equal to and corresponds one-to-one with the number of wavelength types. For example, channel A corresponds to wavelength λ1, channel B to wavelength λ2, channel C to wavelength λ3, and channel D to wavelength λ4. This one-to-one correspondence in this embodiment is achieved through wavelength tuning of the micro-ring modulator 21. As mentioned above, during the system initialization phase, the heater of each channel can be used to tune the resonant wavelength of that channel to the corresponding target wavelength, so that channel A only resonates with λ1, channel B only resonates with λ2, channel C only resonates with λ3, and channel D only resonates with λ4. When a combined optical signal containing all wavelengths enters the micro-ring modulator 21, each channel automatically selects the wavelength corresponding to its own channel for modulation, while optical signals of other wavelengths pass through unaffected, achieving natural allocation of wavelength resources without the need for additional wavelength separation devices.

[0060] It should also be noted that, since each micro-ring modulator 21 has the same number of channels with the same wavelength correspondence, all elements of the input matrix can be encoded in parallel according to a two-dimensional row × column structure. For example, the element x in the i-th row and j-th column... ij Encoding is performed by the j-th channel of the i-th micro-ring modulator 21, which encodes wavelength λ. j The optical signal is modulated, and the intensity of the modulated light is proportional to x. ij The value of , and all channels of all micro-ring modulators 21 work simultaneously, so that the N×N elements of the entire input matrix X are encoded in the same time period.

[0061] In this embodiment, the number of wavelength types equals the number of columns in the input matrix, and the number of independent channels in each micro-ring modulator 21 corresponds one-to-one with the number of wavelength types, enabling the column dimension information of the input matrix to be encoded into the wavelength dimension. Furthermore, since different wavelengths can be physically modulated and transmitted independently, elements in different columns are naturally separated in the optical domain, avoiding crosstalk. Moreover, based on the scheme provided in this embodiment, when the number of columns in the input matrix increases, only the number of wavelength types and the number of channels in each micro-ring modulator 21 need to be increased accordingly, without changing the basic system architecture, facilitating matrix operations of different scales.

[0062] In one embodiment, a programmable optical filter 8 is further included, which is disposed between the optical frequency comb light source 1 and the micro-ring modulator array 2. The programmable optical filter 8 is used to filter out a preset number of optical signals of different wavelengths from the optical signal output by the optical frequency comb light source 1, wherein the preset number is equal to the number of columns of the input matrix.

[0063] It should be noted that the programmable optical filter 8 in this embodiment is a wavelength selection device, and its passband characteristics can be configured according to actual needs. The optical frequency comb generated by the optical frequency comb light source 1 is a wide-spectrum optical signal that may contain multiple equally spaced comb teeth (i.e., optical signals of different wavelengths). However, in actual matrix operations, the number of columns in the input matrix (i.e., the number of wavelength types required) is fixed. For example, for a matrix operation with a 4×4 input matrix, only 4 wavelengths are needed to meet the requirements. If all wavelength signals are sent to the subsequent system, it will waste wavelength resources and introduce unnecessary crosstalk and noise. Therefore, in this embodiment, a programmable optical filter 8 can be set between the output end of the optical frequency comb light source 1 and the input end of the micro-ring modulator array 2. When the programmable optical filter 8 is working, its passband characteristics can be dynamically configured according to the number of columns in the input matrix, enabling it to select the required number of specific wavelength optical signals from the wide-spectrum optical signal and filter out other wavelength optical signals. For example, for matrix operations with a 4×4 input matrix, it can be configured to allow only four wavelengths, λ1, λ2, λ3, and λ4, to pass through; for matrix operations with an 8×8 input matrix, it can be configured to allow eight wavelengths to pass through.

[0064] In this embodiment, by setting a programmable optical filter 8 between the optical frequency comb light source 1 and the micro-ring modulator array 2, flexible selection of the output wavelength of the optical frequency comb can be achieved. This allows for the selection of different numbers of wavelengths according to different application scenarios, adapting to matrix operations of different dimensions, and enhancing the system's versatility and scalability. By filtering out unwanted wavelength components, the programmable optical filter 8 can reduce background noise and potential crosstalk in subsequent optical paths, improving signal quality.

[0065] In one implementation, the arrayed waveguide grating router 3 is an N×N passive router device with N input ports and N output ports, where N is equal to the number of rows in the input matrix; The arrayed waveguide grating router 3 is used to receive multiple optical signals carrying different wavelengths from the output of the i-th micro-ring modulator through the i-th input port, and passively route and synchronously transmit each optical signal from the i-th input port to the i-th output port, so as to realize the row-to-row parallel mapping of the optical signal corresponding to the element of the i-th row in the input matrix to the Mach-Zehnder interferometer array 4.

[0066] It should be noted that the arrayed waveguide grating router 3 in this embodiment achieves a deterministic mapping between wavelength and port by utilizing the optical path difference of the arrayed waveguide, and is a wavelength-selective router based on diffraction and interference. When the input matrix X is an N×N matrix, the arrayed waveguide grating router 3 can employ an N×N passive router device, which has N input ports and N output ports, where N is the number of rows in the input matrix. For example, when the input matrix X is a 4×4 matrix, the number of rows is 4, therefore the arrayed waveguide grating router 3 uses a 4×4 AWGR, that is, it has 4 input ports and 4 output ports. The arrayed waveguide grating router 3 can simultaneously perform the separation, routing, and parallel transmission of multi-wavelength signals, and the entire process requires no additional tuning devices. Being a passive device, it can synchronously transmit optical signals of different wavelengths to the input end of the Mach-Zehnder interferometer array 4 with low crosstalk and low loss.

[0067] It is understandable that the input terminals of the arrayed waveguide grating router 3 are connected one-to-one with the output terminals of the multimode interferometric combiner 22. For example, for an input matrix of 4×4, the first input port of the arrayed waveguide grating router 3 receives the combined optical signal carrying different wavelengths (λ1, λ2, λ3, λ4) output by the first micro-ring modulator MRM1 (corresponding to the first row of the input matrix). The light intensity of each wavelength in the combined optical signal encodes the four elements x of the first row of the input matrix. 11 x 12 x 13 x 14 (That is, the light intensity encoding element x at wavelength λ1) 11 The light intensity encoding element x at wavelength λ2 12 The light intensity encoding element x at wavelength λ3 13 The light intensity encoding element x at wavelength λ4 14 The second input port receives the combined optical signals carrying different wavelengths (λ1, λ2, λ3, λ4) output from the second micro-ring modulator MRM2, which encodes the element x in the second row. 21 x 22 x 23x 24 The third input port receives the combined optical signals carrying different wavelengths (λ1, λ2, λ3, λ4) output from the third micro-ring modulator MRM3, which encodes the element x in the third row. 31 x 32 x 33 x 34 The fourth input port receives the combined optical signals carrying different wavelengths (λ1, λ2, λ3, λ4) output from the fourth micro-ring modulator MRM4, which encodes the element x in the fourth row. 41 x 42 x 43 x 44 .

[0068] Furthermore, in this embodiment, all wavelengths of optical signals received at the i-th input port of the arrayed waveguide grating router 3 are passively routed and synchronously transmitted to the i-th output port. That is, the combined optical signal at the first input port (containing λ1~λ4, carrying x...) 11 ~x 14 The combined optical signal (carrying x) is transmitted to the first output port; the combined optical signal (carrying x) at the second input port is transmitted to the first output port. 21 ~x 24 The combined optical signal (carrying x) is transmitted to the second output port; the combined optical signal (carrying x) at the third input port is transmitted to the second output port. 31 ~x 34 The combined optical signal (carrying x) is transmitted to the third output port; the combined optical signal (carrying x) at the fourth input port is transmitted to the third output port. 41 ~x 44 The optical signal is transmitted to the fourth output port, so that the optical signal corresponding to each row element of the input matrix is ​​completely and synchronously transmitted to the corresponding output port, realizing row-to-row parallel mapping.

[0069] Furthermore, in this embodiment, the output port of the arrayed waveguide grating router 3 transmits the output signal to the corresponding input port of the Mach-Zehnder interferometer array 4. That is, the first output port of the arrayed waveguide grating router 3 is connected to the MZI modulator array in the Mach-Zehnder interferometer array 4 used to encode the first row of the weight matrix, the second output port is connected to the MZI modulator array used to encode the second row of the weight matrix, and so on. Since the arrayed waveguide grating router 3 is a passive routing device, the entire routing process does not consume additional power, and all signals arrive synchronously, ensuring that the data in each row of the input matrix is ​​aligned with the data in each row of the weight matrix in the Mach-Zehnder interferometer array 4, providing correct time synchronization and spatial mapping for subsequent matrix multiplication operations.

[0070] It's important to note that MZI (Mechanical-Zenith Imaging) can be used to implement matrix-vector multiplication, a fundamental operation in various computational tasks such as neural network inference and signal processing. By combining multiple MZI modulators into an optical network, the input optical signal can be encoded as a vector, and the phase modulation of the MZI corresponds to the weight values ​​of the matrix. When the optical signal passes through the MZI network, the interference effect of the light naturally performs the matrix-vector multiplication operation, and the result is presented in the form of an output optical signal. Due to the high speed and parallelism of light, this optical implementation is more efficient than traditional electronic computing when processing large-scale data.

[0071] like Figure 4 As shown, each silicon-based MZI modulator consists of two couplers and two phase shifters, where both the input and output couplers can be 2×2 couplers. In practical applications, such as when there is only one input signal, the input coupler splits the light from the single light source into two beams. Each beam passes through two different phase shift arms (i.e., paths connecting different phase shifters), resulting in a certain phase difference. Finally, these two beams are combined again by a beam combiner to form an interference signal. Based on the phase difference, constructive or destructive interference of the light amplitude can be achieved. MZI has advantages such as simple working principle, mature manufacturing process, and high reliability. Furthermore, its manufacturing process is compatible with CMOS (Complementary Metal-Oxide-Semiconductor) technology, making it a key basic unit in optoelectronic integrated systems.

[0072] The second-order transmission unitary matrix of the silicon-based MZI modulator can be written as follows: , ; Among them, E in1 E represents the input signal at the first input terminal of the input coupler. in2 E represents the input signal at the second input terminal of the input coupler. o1 E represents the coupling signal at the first output terminal of the output coupler. o2 This represents the coupling signal at the second output terminal of the output coupler. Indicates phase difference, Indicates the phase of the first phase shifter. This represents the phase of the second phase shifter. In this relationship, only E... in1 Port input optical signal, E in2 =0, at which point it is possible to calculate from E in1 Enter into E o1 Output optical power transfer function for: ; In other words, for any real-valued matrix W∈R (N×N) A mathematical scheme (such as min-max normalization) is needed to scale each element to the range [0,1], so that each element in the preprocessed matrix W corresponds to the optical power transfer function T of a silicon-based MZI modulator. MZI That is, (W) il =T (i,l) MZI Dynamic control of matrix elements can be achieved by adjusting the phase difference.

[0073] In one embodiment, the Mach-Zehnder interferometer array 4 includes a Mach-Zehnder modulator array consisting of multiple cascaded Mach-Zehnder modulators, which is used to encode a weight matrix. Each Mach-Zehnder modulator controls the splitting ratio by adjusting the phase difference between its two interferometer arms. The splitting ratio is equal to the value of the corresponding element in the preprocessed weight matrix, so that the Mach-Zehnder interferometer array 4 can perform the multiplication operation of the weight and the input matrix element in the optical domain.

[0074] It should be noted that in practical applications, the elements of the original weight matrix W0 can be any real value (including negative numbers or numbers greater than 1), while the splitting ratio of a Mach-Zehnder modulator can only represent values ​​within the range [0,1]. Therefore, the original weight matrix needs to be preprocessed, for example, by using methods such as min-max normalization to scale each element of the weight matrix W0 to the range [0,1], resulting in the preprocessed weight matrix W. Then, each element in W is mapped to the corresponding optical power transfer function (i.e., splitting ratio) of the Mach-Zehnder modulator, i.e., (W). il =T (i,l) MZI Based on this, the splitting ratio of each Mach-Zehnder modulator is equal to the value of the corresponding element in the preprocessed weight matrix, and can be adjusted by changing the phase difference Δ. It can dynamically control the splitting ratio, thereby enabling dynamic configuration of weighting elements.

[0075] Furthermore, in practical applications, multiple Mach-Zehnder modulators can be cascaded according to a specific topology to form a Mach-Zehnder modulator array, used to encode the complete weight matrix. When the optical signal output from the arrayed waveguide grating router 3 (with encoded input matrix elements) enters the Mach-Zehnder modulator array, each Mach-Zehnder modulator attenuates or allows the input optical signal to pass through according to its preset splitting ratio, thereby realizing the multiplication operation between the weights and the input matrix elements. Let the input matrix element x... kl (element in row k, column l) and weight element w ij Taking the multiplication of (the element in the i-th row and j-th column) as an example, when carrying x klThe optical signal of the information is split by a ratio of w ij When using an MZI modulator, the intensity of the output optical signal is w. ij ×x kl Since the transmission and processing of optical signals in the MZI modulator are purely optical, the entire process requires no photoelectric conversion, resulting in high processing speed and low energy consumption.

[0076] In this embodiment, a weight matrix is ​​encoded by a Mach-Zehnder modulator array composed of multiple cascaded Mach-Zehnder modulators. The splitting ratio of each MZI modulator is equal to the value of the corresponding element in the preprocessed weight matrix, thus realizing the multiplication operation between the weights and the elements of the input matrix in the optical domain. This optical domain multiplication method in this embodiment fully utilizes the inherent high-speed parallel characteristics of photons, avoiding the delay and energy consumption caused by frequent data transfer and element-by-element serial multiplication in related technologies. Simultaneously, by preprocessing, any real-valued weight matrix is ​​mapped to the [0,1] splitting ratio range achievable by MZI, enabling the optical computing system to process general weight matrices and exhibiting good versatility.

[0077] In one implementation, the number of Mach-Zehnder modulators in the Mach-Zehnder modulator array is N. 2 , where N equals the number of rows in the input matrix; the Mach-Zehnder modulator array is used to encode the N×N weight matrix; Each Mach-Zehnder modulator corresponds to one element in the weight matrix. The Mach-Zehnder modulator adjusts the phase difference between its two interferometer arms through thermo-optic or electro-optic effects to control the splitting ratio.

[0078] It should be noted that for an N×N weight matrix W, N² Mach-Zehnder modulators are required to completely encode each element in the matrix. Therefore, the number of Mach-Zehnder modulators in the Mach-Zehnder modulator array in this embodiment is N. 2 For example, for a 4×4 weight matrix, N=4, therefore 16 Mach-Zehnder modulators (such as...) are needed. Figure 5 As shown, the 16 Mach-Zehnder modulators are arranged in a cascaded manner, with each modulator having a specific position in the array corresponding to a corresponding element in the weight matrix. For example, the four Mach-Zehnder modulators in the first row correspond to the four elements w in the first row of the weight matrix. 11 w 12 w 13 w 14 The four Mach-Zehnder modulators in the second row correspond to the four elements w in the second row. 21 w 22 w 23 w 24Similarly, the one-to-one correspondence structure in the embodiments of this application makes the encoding of the weight matrix more intuitive and easier to control.

[0079] It should also be noted that each Mach-Zehnder modulator adjusts the phase difference between its two interferometer arms through either a thermo-optical effect or an electro-optical effect, thereby controlling the beam splitting ratio. The thermo-optical effect involves applying current to the heater on the MZI phase-shifting arm, raising the waveguide temperature and thus changing the refractive index, causing a phase shift. The electro-optical effect, on the other hand, changes the phase by applying voltage to alter the refractive index of the waveguide material. In practical applications, the appropriate method can be selected based on actual needs, or a combination of both methods can be used.

[0080] Furthermore, during system initialization, based on the trained weight matrix W0, after preprocessing to obtain W, the system converts each element in W into a corresponding control voltage or current, which is then applied to the corresponding Mach-Zehnder modulator to adjust their phase difference, ensuring the splitting ratio is precisely equal to the target value. Once configured, the splitting ratio of the MZI modulator array is fixed, ready for the arrival of the input matrix's optical signal to perform the corresponding calculations, thereby improving overall efficiency. When the input matrix is ​​updated, if the weight matrix remains unchanged, the MZI modulator array does not need to be reconfigured; however, if the weight matrix needs updating (e.g., different layers of a neural network have different weights), the system can recalculate the control signal and reconfigure the splitting ratio of each Mach-Zehnder modulator.

[0081] Furthermore, in actual optical signal transmission, the arrayed waveguide grating router 3, as a passive routing device, introduces a fixed insertion loss, affecting the optical signal strength reaching the input of the MZI modulator. In practical applications, to ensure the accuracy of multiplication operations, each element x of the input matrix X can be accurately recovered from the optical signal after AWGR routing. kl The original values ​​are obtained. Since the AWGR loss is fixed, the entire system can be calibrated once and reused. For example, when the AWGR insertion loss is 0.98dB, its corresponding power transmission efficiency is approximately 0.80. In this case, only a splitting ratio of 1.25 needs to be adapted at the MZI input to restore the attenuated optical signal to its original value. This calibration process can be completed during system initialization and does not need to be repeated during subsequent operation. After restoring the original values ​​of the input matrix elements, the system uses the MZI phase shifter voltage to control the weighted encoding splitting ratio to achieve the splitting of each element x of the extracted input matrix X. kl The original value is multiplied by the corresponding element of the weight matrix W, where the splitting ratio is the value of the corresponding element in the weight matrix. Taking the first wavelength λ1 in the first micro-ring modulator MRM1 as an example, the optical signal corresponding to this wavelength encodes the input matrix element x. 11 The value when the optical signal passes through a beam splitter with a splitting ratio of w 11When using an MZI modulator, the intensity of the output optical signal is the product w of the weight elements and the input elements. 11 ×x 11 The other MZI modulators in this embodiment operate in the same way as described above, with each of the other MZI modulators performing multiplication operations on the corresponding element.

[0082] It should also be noted that during the multiplication operation, when carrying the input matrix element x... kl Optical signals (in wavelength λ) l (Indicated) After entering the MZI array, the optical signal will pass through the Mach-Zehnder modulators in the corresponding rows and columns. That is, for the element y in the i-th row and j-th column of the output matrix Y... ij It is necessary to calculate the element y in the i-th row and j-th column of the output matrix Y. ij The value of y ij The value of is equal to the sum of k from 1 to N, that is, the element w in the i-th row and k-th column of the weight matrix W. ik The element x in the k-th row and j-th column of the input matrix X kj The sum of the products. In the optical domain, this calculation can be achieved through wavelength routing and intensity accumulation; that is, all encoded x kj The optical signal (k=1…N, i.e., the elements of each row in the j-th column) is multiplied by w when passing through the MZI modulator in the i-th row. ik Then, y is obtained by accumulating in the optical domain. ij For a 4×4 matrix, 16 Mach-Zehnder modulators simultaneously perform 16 multiplication operations. Each multiplication corresponds to the product of a weight element and an input matrix element. All multiplication operations are performed concurrently to achieve the product of the weight matrix W with each element of the input matrix X, without the need for serial waiting. The result of these 16 multiplications can be output as an optical signal to the subsequent accumulation bus for optical domain column summation.

[0083] In this embodiment, an array of N² Mach-Zehnder modulators is configured, corresponding one-to-one with the elements of an N×N weight matrix, enabling complete encoding of the weight matrix. Each MZI modulator is independently controlled, and the splitting ratio can be dynamically configured, allowing the system to flexibly adapt to different weight matrices. Furthermore, this embodiment utilizes the parallel operation of the N² MZI modulators, simultaneously performing N² multiplications and N accumulations in the optical domain. This fully leverages the parallelism of photons, achieving matrix multiplication operations and improving overall performance.

[0084] In one embodiment, the photoelectric detection array 5 includes a balanced photoelectric detector, a positive weight column bus, and a negative weight column bus; The positive weight column bus is used to passively coherently accumulate the optical signals of the same wavelength corresponding to all positive weight multiplication results in the same column, and input the coherent accumulation result into the positive input terminal of the balanced photodetector. The negative weight column bus is used to passively superimpose the optical signals carried by the free spectral range (FSR) of the optical frequency comb corresponding to all negative weight multiplication results in the same column, and input the passive superposition result into the negative input terminal of the balanced photodetector. A balanced photodetector is used to perform photoelectric conversion and differential operations on the optical signals received at the positive and negative input terminals to obtain a column summation result.

[0085] It should be noted that the photodetector array 5 in this embodiment includes a balanced photodetector, a positive weight column bus, and a negative weight column bus. After the Mach-Zehnder interferometer array 4 completes the multiplication operation between the weights and the elements of the input matrix, the multiplication results are output to the photodetector array 5 in the form of optical signals. Specifically, for the matrix multiplication Y = W × X, the element y in the i-th row and j-th column of the output matrix Y is... ij This is equivalent to summing k from 1 to N, i.e., w ik ×x kj For example, for The matrix, the weight matrix W, multiplied by the input matrix X, is: ; For the j-th row of vector y , where j,l=1,2,3,4.

[0086] In the optical domain, this summation operation can be implemented through a column bus structure. That is, all multiplication results located in the same column are coupled to the positive weight column bus in the form of coherent optical signals of the same wavelength. With the help of the natural coherent superposition property of light, these optical signals of the same wavelength will automatically superimpose their amplitudes to complete the passive accumulation of the positive weight sum without any additional adders or control circuits.

[0087] Understandably, for a weight matrix containing only positive real numbers (such as a 4×4 positive real number matrix), all multiplication results are positive, and accumulation can be directly performed using the positive weight column bus. Taking a 4×4 matrix as an example, the bus output of the first column of the MZI array corresponds to the element y in the first row of the output matrix. 11 ~y 14 The bus output of the second column of the MZI array corresponds to y 21 ~y 24 The third column of the MZI array corresponds to y 31 ~y 34 The fourth column of the MZI array corresponds to y 41 ~y 44Furthermore, when the weight matrix contains negative weights, the result of the negative weight multiplication is carried by the independent free spectral range (FSR) of the optical frequency comb light source 1. After being precisely separated by a micro-ring filter, it is connected to the negative weight column bus, synchronously completing the passive superposition of the sum of negative weights. In this embodiment, by accumulating the positive and negative weights on different buses, the mutual cancellation of positive and negative signals in the optical domain can be effectively avoided.

[0088] Furthermore, the balanced photodetector in this embodiment is a high-sensitivity differential detection device with a positive input terminal and a negative input terminal. The accumulation result of the positive weight column bus (corresponding to the sum of positive weights) is input to the positive input terminal of the balanced photodetector, and the superposition result of the negative weight column bus (corresponding to the sum of negative weights) is input to the negative input terminal of the balanced photodetector. The balanced photodetector performs photoelectric conversion on the optical signals at the positive and negative input terminals and performs electrical domain differential operation, that is, the output electrical signal is proportional to the light intensity at the positive input terminal minus the light intensity at the negative input terminal. Through this differential operation, the system can cancel the negative weight component, and at the same time, it can also eliminate the background noise and crosstalk signals common to the positive and negative input terminals, thereby obtaining a more accurate column summation result. For the positive real number weight matrix, the negative weight column bus can be idle or grounded, and the balanced photodetector only needs to perform photoelectric conversion on the positive input terminal.

[0089] It should also be noted that, in the actual reading of the output results, combined with timing control and signal addressing logic, the balanced photodetector sequentially collects and extracts the optical signals of each element of the output matrix Y according to the row and column order. For an N×N matrix, the system sequentially reads N columns of buses, with each column outputting N summation results, for a total of N×N elements. For example, for a 4×4 matrix, the balanced photodetector sequentially samples the optical signals of the four column buses (i.e., the optical signals corresponding to the 16 elements), converting the results of the parallel summation in the optical domain into readable and subsequently processed digital signals in the electrical domain, achieving a one-to-one correspondence between each output matrix element and the superimposed signal in the optical domain. Since the columns of buses operate in parallel, all column summation results are generated simultaneously in the optical domain, and the balanced photodetector only needs to read them sequentially without waiting for the calculation process.

[0090] In this embodiment, passive accumulation is achieved using the natural coherent superposition characteristics of light through a positive weighted bus, eliminating the need for additional adders or control circuits and realizing zero-power optical domain summation. A negative weighted bus carries the negative weight component, which is automatically canceled out by differential operations of a balanced photodetector, simultaneously eliminating background noise and crosstalk, thus improving the accuracy and stability of the output signal. By accumulating the positive and negative weights on different buses, mutual cancellation of positive and negative signals in the optical domain can be effectively avoided. Furthermore, the balanced photodetector, based on a high-sensitivity differential detection principle, can efficiently convert weak optical signals into electrical signals while canceling noise through differential operations. Combined with timing control and signal addressing logic, the balanced photodetector reads the output of each bus column sequentially, achieving a one-to-one correspondence between the output matrix elements and the superimposed optical domain signal, resulting in more accurate results.

[0091] In one implementation, the optical computing system can be used to perform matrix multiplication operations in a neural network; The optical computing system is configured to: convert the data to be processed into an input matrix, set the weights of at least one layer of the neural network as a weight matrix, perform matrix multiplication operations through the optical computing system, and output classification or recognition results based on the operation results.

[0092] It should be noted that the optical computing system provided in this application embodiment can be widely used in neural network inference tasks, such as matrix multiplication operations in pre-trained language models like convolutional neural networks and Transformers. The forward propagation of any layer in a neural network is essentially a linear transformation between the input matrix and the weight matrix, i.e., Y = W × X, where X is the input feature matrix, W is the weight matrix of that layer, and Y is the output feature matrix. The optical computing system in this application embodiment can perform this matrix multiplication operation at high speed in the optical domain, providing hardware acceleration for neural network inference.

[0093] It should also be noted that in practical applications, the electrical signal data to be processed (such as pixel values ​​of an image or sensor data) can be converted into an input matrix X. For example, in image classification tasks, the pixel values ​​of the image can be preprocessed (such as normalization and / or resizing) and organized into a matrix form suitable for the input layer dimension of a neural network to obtain the input matrix X. In sensor data processing tasks, the data collected by sensors (such as time-series signals such as temperature, pressure, and vibration) can be arranged into a matrix form in a preset manner (such as time windows or spatial locations) to obtain the input matrix X. It is understandable that after obtaining the input matrix X, this input matrix X can be used to input into an optical computing system for matrix multiplication operations.

[0094] Furthermore, before performing matrix multiplication, the weight matrix of the current layer in the neural network can be preprocessed using the method described above. That is, each real-valued element in the weight matrix is ​​normalized and mapped to the range [0,1]. Then, the splitting ratio of each MZI modulator is set to the value of the corresponding element in the preprocessed weight matrix through the phase shifter voltage control of the MZI modulator. For multi-layer neural networks, configuration can be performed layer by layer: after the matrix multiplication of the first layer is completed, the output result is used as the input matrix of the second layer. The system reconfigures the weight matrix of the second layer and continues to perform matrix multiplication operations, and so on, until the calculation of all layers is completed. Since the configuration frequency of the weight matrix is ​​low (only once per layer), and the configuration of the MZI modulator can be quickly completed through thermo-optic or electro-optic effects, the system can efficiently support the inference task of multi-layer neural networks.

[0095] It should also be noted that after obtaining the output result based on the optical computing system provided in this application embodiment, the system can determine the final classification or recognition result based on the output electrical signal. For example, in an image classification task, the output result can be a probability distribution vector, and the category with the highest probability can be determined based on the output result, and the category with the highest probability can be used as the classification result (e.g., cat or dog). As another example, in an anomaly detection task, the output result can be a binary signal or an anomaly score, and the system can determine whether the output result corresponds to normal data or abnormal data based on a preset threshold.

[0096] In this embodiment of the application, by applying the optical computing system provided in the above embodiments of the application to matrix multiplication operations in neural networks, it is beneficial to significantly improve the speed and energy efficiency of neural network inference.

[0097] Based on the above embodiments, this application also provides an electronic device, including an optical computing system as described in any of the above embodiments.

[0098] It should be noted that the electronic device in this application embodiment has the same beneficial effects as the optical computing system provided in the above embodiments, and for a detailed description of the optical computing system involved in this application embodiment, please refer to the above embodiments, which will not be repeated here.

[0099] Based on the above embodiments, this application also provides a photon matrix calculation method, applied to the optical computing system described above, such as... Figure 6 As shown, the method includes: S110: Utilizes an optical frequency comb light source to output optical signals of various wavelengths; S120: The micro-ring modulator array is used to modulate and encode the received optical signals of various wavelengths in order to determine the optical signal corresponding to each element in the input matrix. S130: The arrayed waveguide grating router is used to synchronously transmit the optical signals corresponding to each element in the input matrix. S140: Encode the weight matrix using a Mach-Zehnder interferometer array, and perform multiplication operations between the weights and matrix elements based on the encoded weight matrix and the optical signals corresponding to each element in the input matrix to obtain the corresponding multiplication results; S150: The optical signal after optical domain accumulation based on each multiplication result is converted into an optical signal by a photoelectric detection array to obtain the output result.

[0100] For a description of the features in the embodiment corresponding to the photon matrix calculation method, please refer to the relevant description of the embodiment corresponding to the optical computing system, which will not be repeated here.

[0101] Based on the above embodiments, this application also provides a photonic matrix computing device, applied to the optical computing system described above, such as... Figure 7 As shown, the device includes: Generation module 11 is used to output optical signals of various wavelengths using an optical frequency comb light source; The modulation and coding module 12 is used to modulate and encode the received optical signals of various wavelengths using a micro-ring modulator array, so as to determine the optical signals corresponding to each element in the input matrix. Transmission module 13 is used to synchronously transmit the optical signals corresponding to each element in the input matrix using an arrayed waveguide grating router; The calculation module 14 is used to encode the weight matrix using the Mach-Zehnder interferometer array, and to perform multiplication operations between the weights and matrix elements based on the encoded weight matrix and the optical signals corresponding to each element in the input matrix, so as to obtain the corresponding multiplication results. The conversion module 15 is used to perform photoelectric conversion on the optical signal after optical domain accumulation based on each multiplication result using a photoelectric detection array to obtain the output result.

[0102] For a description of the features in the embodiment corresponding to the photonic matrix computing device, please refer to the relevant description of the embodiment corresponding to the optical computing system, which will not be repeated here.

[0103] Any of the components, modules, units, parts, methods, and operations described herein can be implemented using software, firmware, hardware (e.g., fixed logic circuitry), manual processing, or any combination thereof. Alternatively or additionally, any functionality described herein can be executed at least in part by one or more hardware logic components, such as, but not limited to, a central processing unit (CPU), a field-programmable gate array (FPGA), an application-specific integrated circuit (ASIC), an application-specific standard product (ASSP), a system-on-a-chip (SoC), a complex programmable logic device (CPLD), a microprocessor (MCU), etc. The terms "system," "computing device," or "apparatus" as used herein encompass various means, devices, and machines for processing data, including, for example, one or more programmable processors, computers, SoCs, or combinations thereof. The apparatus may also include code that creates an execution environment for the computer program in question, such as code constituting processor firmware, a protocol stack, a database management system, an operating system, a cross-platform runtime environment, a virtual machine, or one or more combinations thereof. The aforementioned computer program (also known as a program, software, software application, app, script, or code) can be written in any form of programming language, including compiled or interpreted languages, declarative or procedural languages, and can be deployed in any form, including as a standalone program or as a module, component, subroutine, object, or other unit suitable for a computing environment.

[0104] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0105] The optical computing system and electronic device provided in this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this application. It should be noted that those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of the claims of this application.

Claims

1. An optical computing system, characterized in that, include: An optical frequency comb light source is used to output optical signals of various wavelengths. The micro-ring modulator array connected to the optical frequency comb light source is used to modulate and encode the received optical signals of various wavelengths in order to determine the optical signals corresponding to each element in the input matrix. An arrayed waveguide grating router connected to the micro-ring modulator array is used to synchronously transmit the optical signals corresponding to each element in the input matrix. The Mach-Zehnder interferometer array connected to the arrayed waveguide grating router is used to encode the weight matrix, and to perform multiplication operations between the weights and matrix elements based on the encoded weight matrix and the optical signals corresponding to each element in the input matrix, so as to obtain the corresponding multiplication results. The photoelectric detection array connected to the Mach-Zehnder interferometer array is used to perform photoelectric conversion on the optical signal after optical domain accumulation based on each of the multiplication results to obtain the output result.

2. The optical computing system according to claim 1, characterized in that, It also includes an optical beam splitter disposed between the optical frequency comb light source and the micro-ring modulator array; The optical beam splitter is used to split the received optical signals of various wavelengths to obtain multiple split optical signals; wherein each split optical signal includes optical signals of all wavelengths. The micro-ring modulator array is used to modulate and encode the optical signals of various wavelengths in each of the split optical signals to determine the optical signals corresponding to each element in the input matrix.

3. The optical computing system according to claim 2, characterized in that, It also includes a wavelength division multiplexer disposed between the optical frequency comb light source and the optical beam splitter; The wavelength division multiplexer is used to combine multiple optical signals of different wavelengths output by the optical frequency comb light source into the same optical path and transmit them to the optical beam splitter. The optical beam splitter is used to split optical signals of different wavelengths in a combined optical signal to obtain multi-beam split optical signals.

4. The optical computing system according to claim 1, characterized in that, The microring modulator array includes multiple microring modulators, the number of which is the same as the number of rows in the input matrix. Each microring modulator has multiple independent channels, the number of which is the same as the number of columns in the input matrix. Each channel is used for resonant modulation of an optical signal of one wavelength.

5. The optical computing system according to claim 4, characterized in that, The micro-ring modulator includes a heater and a modulation electrode; The heater is used to tune the resonant wavelength of each of the channels to the target wavelength, so that each channel resonates only with the optical signal corresponding to the target wavelength among the various optical signals of the wavelengths; The modulation electrode is used to adjust the light intensity of the optical signal of the corresponding wavelength according to the value of each element in the input matrix.

6. The optical computing system according to claim 4, characterized in that, The micro-ring modulator array also includes a multimode interference combiner connected to each of the micro-ring modulators in a one-to-one correspondence. The multimode interference combiner is used to combine the optical signals output from each channel of the micro-ring modulator, which carry different wavelengths and have been encoded into input matrix elements, into a single combined optical signal, and transmit the combined optical signal to the corresponding input terminal of the arrayed waveguide grating router.

7. The optical computing system according to claim 4, characterized in that, The number of wavelength types in the various optical signals of different wavelengths output by the optical frequency comb light source is equal to the number of columns in the input matrix; The number of independent channels in each of the micro-ring modulators is equal to and corresponds one-to-one with the number of wavelength types, and each channel is used to resonate and modulate an optical signal of a corresponding wavelength.

8. The optical computing system according to claim 7, characterized in that, It also includes a programmable optical filter disposed between the optical frequency comb light source and the micro-ring modulator array; The programmable optical filter is used to filter out a preset number of optical signals of different wavelengths from the optical signals output by the optical frequency comb light source.

9. The optical computing system according to claim 7, characterized in that, The various optical signals of different wavelengths are from the same source and are phase-synchronized to support coherent accumulation operations in the optical domain.

10. The optical computing system according to claim 4, characterized in that, The arrayed waveguide grating router is an N×N passive router device with N input ports and N output ports, where N is equal to the number of rows in the input matrix; The arrayed waveguide grating router is used to receive multiple optical signals carrying different wavelengths from the output of the i-th micro-ring modulator through the i-th input port, and passively route and synchronously transmit each optical signal from the i-th input port to the i-th output port, so as to realize the row-to-row parallel mapping of the optical signal corresponding to the element of the i-th row in the input matrix to the Mach-Zehnder interferometer array.

11. The optical computing system according to claim 2, characterized in that, The optical beam splitter is a 1×N beam splitter, where N is equal to the number of rows in the input matrix; The optical beam splitter is used to uniformly split the received optical signal into N identical optical signals, and input the N identical optical signals to the N micro-ring modulators in the micro-ring modulator array respectively.

12. The optical computing system according to claim 1, characterized in that, The photoelectric detection array includes a balanced photoelectric detector, a positive weighted column bus, and a negative weighted column bus; The positive weight column bus is used to passively coherently accumulate the optical signals of the same wavelength corresponding to all positive weight multiplication results in the same column, and input the coherent accumulation result into the positive input terminal of the balanced photodetector. The negative weight column bus is used to passively superimpose the optical signals corresponding to the independent free spectral range of the optical frequency comb, which are all the negative weight multiplication results located in the same column, and input the passive superposition result into the negative input terminal of the balanced photodetector. The balanced photodetector is used to perform photoelectric conversion and differential operation on the optical signals received at the positive and negative input terminals to obtain a column summation result.

13. The optical computing system according to any one of claims 1 to 12, characterized in that, The Mach-Zehnder interferometer array includes a Mach-Zehnder modulator array composed of multiple cascaded Mach-Zehnder modulators, which is used to encode the weight matrix. Each Mach-Zehnder modulator controls the splitting ratio by adjusting the phase difference between its two interferometer arms. The splitting ratio is equal to the value of the corresponding element in the preprocessed weight matrix, so that the Mach-Zehnder interferometer array can perform the multiplication operation of the weights and the elements of the input matrix in the optical domain.

14. The optical computing system according to claim 13, characterized in that, The number of Mach-Zehnder modulators in the Mach-Zehnder modulator array is N. 2 Where N equals the number of rows in the input matrix; the Mach-Zehnder modulator array is used to encode an N×N weight matrix; Each of the Mach-Zehnder modulators corresponds to an element in the weighting matrix. The Mach-Zehnder modulator adjusts the phase difference between its two interferometer arms through thermo-optic or electro-optic effects to control the beam splitting ratio.

15. An electronic device, characterized in that, Includes the optical computing system as described in any one of claims 1 to 14.

Images