Multi-dimensional expandable matrix operation system and control method thereof
Through a multi-dimensional scalable matrix computing system, the wavelength, space division and time division multiplexing of light is used to realize parallel processing and scale expansion of large-scale matrix computing, solving the problem that single-dimensional multiplexing in the existing technology cannot achieve parallel computing, and improving computing throughput and flexibility.
Patent Information
- Application Number
- CN202510377357.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-28
- Publication Date
- 2025-04-25
- Estimated Expiration
- 2045-03-28
AI Technical Summary
When the prior art realizes large-scale matrix operations, single-dimensional multiplexing cannot implement parallel computing, which limits the scale and number of devices, and the computing scale expansion depends on spatial combination, resulting in insufficient compactness and flexibility.
A multi-dimensional expansion-able matrix computing system is adopted to generate optical carrier signals of multiple carrier wavelengths through a tunable laser, and time division multiplexing and space division multiplexing are used for intensity modulators and micro-ring modulators to realize multi-dimensional expansion and parallel processing of matrix information.
It realizes parallel processing and scale expansion of matrix operations, without the need for secondary growth of components, reduces production costs, improves computing throughput and energy efficiency, and is suitable for application scenarios that require online training.
Smart Images

Figure CN119882923A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of optical computing, and more specifically, to a multi-dimensional scalable matrix operation system and a control method thereof. Background Art
[0002] With the continuous rise of artificial intelligence technology in recent years, the amount of data generated has grown exponentially, which has overburdened the transmission network, caused latency problems and increased power consumption. Therefore, it is necessary to be able to achieve flexible multi-mode data processing while having high computing throughput and energy efficiency. Integrated photonics offers unique potential advantages, including low interconnection losses and huge bandwidth, and is a promising platform for the development of optical computing accelerators. Photonic devices can utilize the different characteristics of light waves in time, wavelength and space to improve parallelism and data throughput. Advances in photonic linear processors have further improved data reuse efficiency and significantly reduced the energy overhead associated with optoelectronics and analog-to-digital conversion.
[0003] In actual operation, how to effectively utilize these light wave sizes to expand the scale of optical matrix processors has become a widely concerned solution. Research on single optical matrix operations based on spatial dimension multiplexing uses space division multiplexing (SDM) to achieve scalability, but it is difficult to solve large-scale problems because it requires a quadratic growth of components, which complicates manufacturing. In contrast, single wavelength division multiplexing (WDM) reduces component requirements but relies on bulky dispersive elements. Single time division multiplexing (TDM) cannot achieve low latency by serializing data into time waveforms, improving bandwidth utilization, and supporting real-time training. However, the computational scale expansion of these architectures usually relies on spatial combinations with different delays, so additional on-chip delay lines and dispersive elements are also required, limiting the compactness and flexibility of photonic processors. Summary of the invention
[0004] The purpose of the present invention is to overcome the shortcomings of the prior art that the single dimensional multiplexing cannot realize parallel large-scale matrix operations and at the same time limits the scale and number of devices, and to provide a multi-dimensional scalable matrix operation system and a control method thereof, which can perform large-scale matrix operations in parallel and realize the scalability of matrix operations.
[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: A multi-dimensional scalable matrix operation system is provided, comprising: Tunable laser: used to generate optical carrier signals of multiple carrier wavelengths; Intensity modulator: used to modulate the optical carrier signal according to the first linear sequence in the time domain, load the optical carrier signal on the light intensity, and obtain the light intensity signal carrying intensity information; Wavelength division multiplexer: used to multiplex the optical intensity signals of different wavelengths and input them into the micro-ring modulator; Micro-ring modulator: used to modulate the optical intensity signal according to the second linear sequence in the time domain to obtain a modulated signal; and to superimpose the modulated signals of different symbol periods through micro-ring feedback to obtain an output optical signal; the output optical signal carries the matrix multiplication and addition result of the first linear sequence and the second linear sequence; Demultiplexer: used to separate output optical signals of different wavelengths into individual channels; Photodetector: used to detect the intensity of the output optical signal on each channel.
[0006] The present invention provides a multi-dimensional scalable matrix operation system, which loads the input matrix on light waves of different wavelengths and utilizes the wavelength dimension of light to expand the scale of matrix operations. The light intensity signal after wavelength division multiplexing is input to different micro-ring modulators, which utilizes space division multiplexing to show the scalability of the matrix operation unit, converts the matrix information into a time series, and through time division multiplexing, improves the bandwidth utilization and expands the scale of matrix operations. At the same time, it supports real-time training and flexibly regulates the information of each matrix element. The present invention can utilize the various dimensions of light to flexibly control the input of matrix operations, can accelerate the processing of data in parallel, and quickly solve large-scale problems.
[0007] Furthermore, the first linear sequence is a time sequence formed by arranging input matrices in a corresponding multiplication order; multiple input matrices simultaneously form multiple parallel time sequences in the same order, i.e., multiple first linear sequences; multiple first linear sequences are encoded onto different optical carrier wavelengths by multiple intensity modulators; and the second linear sequence is a time sequence formed by arranging multiplication matrices in a multiplication order corresponding to the input matrix.
[0008] Furthermore, the input matrices are arranged in row order to form a first linear sequence; and the multiplication matrices are arranged in column order to form a second linear sequence.
[0009] Furthermore, the micro-ring modulator includes a 2×2 first multi-mode interferometer and a second multi-mode interferometer, and a Mach-Zehnder electro-optic modulator, the first multi-mode interferometer, the Mach-Zehnder electro-optic modulator, and the second multi-mode interferometer are cascaded, and the first input port of the first multi-mode interferometer is connected to the first output port of the second multi-mode interferometer to form a micro-ring; the second input port of the first multi-mode interferometer is connected to the output end of the wavelength division multiplexer, and the second output port of the second multi-mode interferometer is connected to the input port of the photodetector; the arm length of the Mach-Zehnder interferometer is controlled by regulating the electrode to perform intensity modulation.
[0010] Furthermore, the micro-ring modulator is integrated on a thin-film lithium niobate substrate. In the present invention, a dimensionally scalable matrix computing system can be integrated on a chip and can be integrated to expand computational parallelism or throughput, thereby obtaining a higher data volume.
[0011] Furthermore, the intensity modulator is a Mach-Zehnder electro-optic modulator.
[0012] The present invention also provides a control method for a multi-dimensional scalable matrix operation system, comprising the following steps: S1. Using a tunable laser to output multiple optical carrier signals of different wavelengths; S2. Arrange the input matrix into a time series in the time domain according to the corresponding multiplication sequence to obtain a first linear sequence, and use an intensity modulator to encode the first linear sequence on optical signals of different wavelengths to obtain an optical intensity signal carrying intensity information; S3. The optical intensity signals on different wavelength channels are loaded on the same channel through a wavelength division multiplexer and input to different micro-ring modulators; S4. Arrange the multiplication matrix in a multiplication order corresponding to the input matrix to form a time sequence in the time domain to obtain a second linear sequence; modulate the micro-ring modulator according to the second linear sequence and in a time division multiplexing manner to obtain a modulated signal; use the delay of the micro-ring to control the superposition of modulated signals of different symbol periods to obtain an output optical signal, and control the output optical signal to be output to the wavelength division multiplexer; S5. Separate the output optical signals of different wavelengths into various channels through the wavelength division multiplexer and control the output of the photodetector.
[0013] In the present invention, the multiple wavelengths output by the tunable laser are controlled to be the same as the resonant wavelength of the microring, and the multiple carrier wavelengths are aligned with their optical resonance through the precise dispersion management of the Mach-Zehnder electro-optic modulator and the ring waveguide in the microring modulator to maintain the same modulation response over the entire wavelength. The tunable laser is controlled to output multiple wavelengths, which are the same as the resonant wavelength of the microring, and the same modulation response is maintained over the entire wavelength, which realizes wavelength division multiplexing, improves the parallel processing capability of the matrix operation unit, and realizes the expansion of the scale of matrix operation.
[0014] In the present invention, by increasing the modulation rate of the micro-ring modulator, the micro-ring can have multiple modulation sub-signals of code element periods, so that the scale of the multiplication matrix can be expanded to a corresponding multiple. It is possible to achieve matrix multiplication operations of any scale by time division multiplexing without changing the number of device unit structures, expand the scale of matrix operations, eliminate the need for secondary growth of components, reduce production costs, and achieve flexible regulation.
[0015] In the present invention, multiple micro-ring modulators can be set up to allow real-time reprogramming and modification of the multiplication matrix, which can be combined with neural network linear operations to accelerate the neural network training process, and is very suitable for application scenarios that require online training.
[0016] Furthermore, in step S2, the input matrix X Generate time series by row order x , time series x The amplitude of the optical carrier is serially modulated by an external intensity modulator to obtain an optical signal carrying intensity information.
[0017] Furthermore, the symbol period is equal to one symbol period of the ring recursive delay ,Right now .
[0018] Further, in step S4, the matrix W is multiplied Sort by columns; to get the input matrix X The multiplication and addition results of each row and column of the multiplication matrix W need to be traversed, so it is necessary to repeatedly load a column vector of the multiplication matrix W until a time series x Modulation ends; when the next time series x When input again, continue to repeatedly load the next column vector of the multiplication matrix W until X All row and column multiplication and addition of W are completed.
[0019] Further, in step S4, the multiplication matrix W is re-edited and converted into The second linear arrangement sequence is formed and loaded on the micro-ring modulator, namely:
[0020] In the formula, The coupling matrix of the input micro-ring modulator is modulated on the optical intensity signal through the Mach-Zehnder electro-optic modulator; is the nth row and ith column element of the multiplication matrix W, The coupling matrix The element at row n and column i, is the last row and column i element of the multiplication matrix W, for In the coupling matrix Elements in subsequent rows of the same column.
[0021] Furthermore, the optical intensity signal is input into the micro-ring modulator through the first input port of the micro-ring modulator, and the multiplication operation of the input matrix and the multiplication matrix is realized by controlling the symbol period of the first linear sequence in the time domain to be equal to the symbol period of the second linear sequence loaded on the micro-ring modulator; the newly generated optical signal at the first output port of the micro-ring modulator returns to the second input port along the micro-ring waveguide, wherein the ring recursive delay matches the symbol period, and the next input symbol is just injected at the first input port; and then the accumulation is performed through coherent optical superposition, and finally the amplitude of the modulated signal in the micro-ring modulator satisfies:
[0022] In the formula, is the final result Y i Line j The elements of the column, is the element in the last column of the i-th row of the input matrix X, The coupling matrix The element in the last row and column j, is the i-th row of the input matrix X. Column elements, The coupling matrix No. The element in row and column j, for In the coupling matrix Elements in subsequent rows of the same column, is the element in the mth row and jth column of the multiplication matrix W.
[0023] Finally get No. i Line j The results of the columns also correspond to the input matrix X No. i Row and multiply matrix j The vector dot product of the column is obtained by the photodetector .
[0024] Furthermore, the amplitude-encoded signal in the micro-ring modulator Import to the second output port, and use the photodetector to read the result of the current operation from the second output port.
[0025] Compared with the prior art, the present invention has the following beneficial effects: A multi-dimensional expandable matrix operation system and a control method thereof of the present invention improve the parallel processing capability of matrix operations and realize the expansion of the scale of matrix operations; matrix multiplication operations of any scale can be realized by time division multiplexing without changing the number of device unit structures, thereby expanding the scale of matrix operations without the need for secondary growth of components, reducing manufacturing costs, and realizing flexible regulation; in addition, by setting up multiple micro-ring modulators, real-time reprogramming and modification of multiplication matrices are allowed, which can be combined with neural network linear operations to accelerate the neural network training process, and is very suitable for application scenarios requiring online training; and the system of the present invention can be integrated on a chip and can be integrated to expand computing parallelism or throughput, thereby obtaining a higher amount of data. BRIEF DESCRIPTION OF THE DRAWINGS
[0026] Figure 1 A schematic diagram of the structure of a multi-dimensional scalable matrix operation system invented; Figure 2 It is a schematic diagram of the structure of the micro-ring modulator of the present invention; Figure 3 It is a calculation diagram of the matrix operation system in Example 3; Figure 4 It is a calculation diagram of the matrix operation system in the fourth embodiment; Figure 5 It is a calculation diagram of the matrix operation system in Example 5; Figure 6 This is a graph showing the experimental results of extracting image edges using a multi-dimensional scalable matrix operation system in Example 6.
[0027] In the accompanying drawings: 100, tunable laser; 200, intensity modulator; 300, wavelength division multiplexer; 400, micro-ring modulator; 401, first multi-mode interferometer; 402, second multi-mode interferometer; 403, Mach-Zehnder electro-optic modulator; 404, first input port; 405, second input port; 406, first output port; 407, second output port; 500, demultiplexer; 600, photodetector. DETAILED DESCRIPTION
[0028] The present invention is further described below in conjunction with specific implementation methods. The accompanying drawings are only for illustrative purposes and are only schematic diagrams, not actual drawings, and should not be construed as limiting the present invention. In order to better illustrate the embodiments of the present invention, some parts of the accompanying drawings may be omitted, enlarged, or reduced, and do not represent the size of the actual product. It is understandable to those skilled in the art that some well-known structures and their descriptions in the accompanying drawings may be omitted.
[0029] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right" and the like indicate directions or positional relationships based on the directions or positional relationships shown in the drawings, it is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific direction, be constructed and operated in a specific direction. Therefore, the terms describing the positional relationship in the drawings are only used for illustrative purposes and cannot be understood as limitations on the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.
[0030] Embodiment 1 This embodiment is an embodiment of a multi-dimensional scalable matrix operation system. Figure 1 As shown, including: Tunable laser 100: used to generate optical carrier signals of multiple carrier wavelengths; Intensity modulator 200: used to modulate the optical carrier signal according to the first linear sequence in the time domain, load the optical carrier signal on the light intensity, and obtain a light intensity signal carrying intensity information; Wavelength division multiplexer 300: used to multiplex light intensity signals of different wavelengths and input them into micro-ring modulator 400; Micro-ring modulator 400: used to modulate the optical intensity signal according to the second linear sequence in the time domain to obtain a modulated signal; and to superimpose the modulated signals of different symbol periods through micro-ring feedback to obtain an output optical signal; the output optical signal carries the matrix multiplication and addition operation result of the first linear sequence and the second linear sequence; Demultiplexer 500: used to separate output optical signals of different wavelengths into various channels; Photodetector 600: used to detect the intensity of the output optical signal on each channel.
[0031] In this embodiment, the first linear sequence is a time sequence formed by arranging the input matrices in a corresponding multiplication order; multiple input matrices simultaneously form multiple parallel time sequences in the same order, that is, multiple first linear sequences; multiple first linear sequences are encoded onto different optical carrier wavelengths by multiple intensity modulators 200; the second linear sequence is a time sequence formed by arranging the multiplication matrices in a multiplication order corresponding to the input matrices.
[0032] In this embodiment, the input matrices are arranged in row order to form a first linear sequence; and the multiplication matrices are arranged in column order to form a second linear sequence.
[0033] In this embodiment, if Figure 2As shown, the micro-ring modulator 400 includes a 2×2 first multi-mode interferometer 401 and a second multi-mode interferometer 402, and a Mach-Zehnder electro-optic modulator 403. The first multi-mode interferometer 401, the Mach-Zehnder electro-optic modulator 403, and the second multi-mode interferometer 402 are cascaded, and the first input port 404 of the first multi-mode interferometer 401 is connected to the first output port 406 of the second multi-mode interferometer 402 to form a micro-ring; the second input port 405 of the first multi-mode interferometer 401 is connected to the output end of the wavelength division multiplexer 300, and the second output port 407 of the second multi-mode interferometer 402 is connected to the input port of the photodetector 600; the arm length of the Mach-Zehnder interferometer is controlled by regulating the electrode to perform intensity modulation.
[0034] In this embodiment, the micro-ring modulator 400 is integrated on a thin-film lithium niobate substrate. In the present invention, a dimensionally scalable matrix computing system can be integrated on a chip and can be integrated to expand computational parallelism or throughput, thereby obtaining a higher data volume.
[0035] In this embodiment, the intensity modulator 200 is a Mach-Zehnder electro-optic modulator 403 .
[0036] This embodiment provides a multi-dimensional scalable matrix operation system, which loads the input matrix onto light waves of different wavelengths and utilizes the wavelength dimension of light to expand the scale of matrix operations. The wavelength-division-multiplexed light intensity signal is input to different micro-ring modulators 400, which utilizes space division multiplexing to show the scalability of the matrix operation unit, converts the matrix information into a time series, and through time division multiplexing, improves the bandwidth utilization and expands the scale of matrix operations. At the same time, it supports real-time training and flexibly regulates the information of each matrix element. The present invention can utilize the various dimensions of light to flexibly control the input of matrix operations, can accelerate data processing in parallel, and quickly solve large-scale problems.
[0037] This embodiment provides a compact and flexible photonic matrix processing system with powerful multi-dimensional parallel scalability. It can expand matrix operations in the three dimensions of light, and use time division multiplexing to demonstrate outstanding capabilities in high computing throughput and operational versatility. By further integrating wavelength division multiplexing and space division multiplexing strategies, it can process higher data volumes and demonstrate powerful parallel processing capabilities.
[0038] Embodiment 2 This embodiment is a first embodiment of a control method for a multi-dimensional scalable matrix operation system. This embodiment uses the system provided in the first embodiment and specifically includes the following steps: The present invention also provides a control method for a multi-dimensional scalable matrix operation system, comprising the following steps: S1. Using a tunable laser 100 to output a plurality of optical carrier signals of different wavelengths; S2. Arrange the input matrix into a time series in the time domain according to the corresponding multiplication sequence to obtain a first linear sequence, and use the intensity modulator 200 to encode the first linear sequence on optical signals of different wavelengths to obtain an optical intensity signal carrying intensity information; S3. The optical intensity signals on different wavelength channels are loaded on the same channel through the wavelength division multiplexer 300 and input to different micro-ring modulators 400; S4. Arrange the multiplication matrix in a multiplication order corresponding to the input matrix to form a time sequence in the time domain to obtain a second linear sequence; modulate the micro-ring modulator 400 according to the second linear sequence and in a time division multiplexing manner to obtain a modulated signal; use the delay of the micro-ring to control the superposition of the modulated signals of different symbol periods to obtain an output optical signal, and control the output optical signal to be output to the wavelength division multiplexer 300; S5. The output optical signals of different wavelengths are separated into various channels through the wavelength division multiplexer 500, and the photodetector 600 is controlled to output.
[0039] In this embodiment, the process of matrix operation includes: In step S2, the input matrix X is arranged in row order to generate a time series x, and the time series x is serially modulated on the amplitude of the optical carrier by an external intensity modulator 200 to obtain an optical signal carrying intensity information.
[0040] The symbol period is equal to one symbol period of the ring recursive delay ,Right now .
[0041] In step S4, the matrix W is multiplied Sort by columns; to get the input matrix X The multiplication and addition results of each row and column of the multiplication matrix W need to be traversed, so it is necessary to repeatedly load a column vector of the multiplication matrix W until a time series x Modulation ends; when the next time series x When input again, continue to repeatedly load the next column vector of the multiplication matrix W until X All row and column multiplication and addition of W are completed.
[0042] Since the signal accumulated in the micro-ring is constantly affected by the newly modulated data loaded by the micro-ring modulator 400, in order to ensure that the light intensity signal read out after the final superposition is the multiplication of the corresponding row and column elements, the multiplication matrix W needs to be re-edited and converted into A second linear arrangement sequence is formed and loaded on the micro-ring modulator 400, namely:
[0043] In the formula, The coupling matrix of the input micro-ring modulator is modulated on the optical intensity signal through the Mach-Zehnder electro-optic modulator; is the nth row and ith column element of the multiplication matrix W, The coupling matrix The element at row n and column i, is the last row and column i element of the multiplication matrix W, for In the coupling matrix Elements in subsequent rows of the same column.
[0044] The optical intensity signal is input to the micro-ring modulator 400 through the first input port 404 of the micro-ring modulator 400, and the multiplication operation of the input matrix and the multiplication matrix is realized by controlling the symbol period of the first linear sequence in the time domain to be equal to the symbol period of the second linear sequence loaded on the micro-ring modulator 400; the newly generated optical signal at the first output port 406 of the micro-ring modulator 400 returns to the second input port 405 along the micro-ring waveguide, wherein the ring recursive delay matches the symbol period, and the next input symbol is just injected at the first input port 404; and then the accumulation is performed through coherent optical superposition (i.e., amplitude summation), and finally the amplitude of the modulated signal in the micro-ring modulator 400 satisfies:
[0045] In the formula, is the final result Y i Line j The elements of the column, is the element in the last column of the i-th row of the input matrix X, The coupling matrix The element in the last row and column j, is the i-th row of the input matrix X. Column elements, The coupling matrix No. The element in row and column j, for In the coupling matrix Elements in subsequent rows of the same column, is the element in the mth row and jth column of the multiplication matrix W.
[0046] Finally got No. i Line j The results of the columns also correspond to the input matrix X No. i Row and multiply matrix j The vector dot product of the column is obtained by the photodetector 600. .
[0047] To get the result It does not affect the next y value solution, and the signal encoded with amplitude in the micro-ring modulator 400 The result of the current operation is imported into the second output port 407, and the photoelectric detector is used to read the result of the current operation from the second output port 407.
[0048] In this embodiment, the multiple wavelengths output by the tunable laser 100 are controlled to be the same as the resonant wavelength of the microring, and the multiple carrier wavelengths are aligned with their optical resonance through the precise dispersion management of the Mach-Zehnder electro-optic modulator 403 and the ring waveguide in the microring modulator 400 to maintain the same modulation response over the entire wavelength. The tunable laser 100 is controlled to output multiple wavelengths that are the same as the resonant wavelength of the microring, and the same modulation response is maintained over the entire wavelength, which realizes wavelength division multiplexing, improves the parallel processing capability of the matrix operation unit, and realizes the expansion of the scale of matrix operation.
[0049] In this embodiment, by increasing the modulation rate of the micro-ring modulator 400, the micro-ring can have multiple modulation sub-signals of code element periods, so that the scale of the multiplication matrix can be expanded to a corresponding multiple. It is possible to achieve matrix multiplication operations of any scale by time division multiplexing without changing the number of device unit structures, expand the scale of matrix operations, eliminate the need for secondary growth of components, reduce manufacturing costs, and achieve flexible regulation.
[0050] In this embodiment, multiple micro-ring modulators 400 can also be set up to allow real-time reprogramming and modification of the multiplication matrix, which can be combined with neural network linear operations to accelerate the neural network training process, and is very suitable for application scenarios that require online training.
[0051] Embodiment 3 This embodiment is a second embodiment of a control method for a multi-dimensional scalable matrix operation system. This embodiment is similar to the second embodiment, except that, in this embodiment, the control of the scale-expanded matrix operation system is achieved through time division multiplexing of the micro-ring modulator 400. Figure 3 shown.
[0052] In this embodiment, the modulation rate of the micro-ring modulator 400 is increased by N times, so that the symbol period of the second linear sequence is , the attached figure shows that the micro-ring modulation rate is increased by 2 times.
[0053] In this embodiment, the second linear sequence is loaded onto the electrical input of the micro-ring modulator 400, which multiplies it with the input matrix and uses its feedback loop to accumulate the optical signal. Increasing the modulation rate of the micro-ring modulator 400 can perform time division multiplexing on the second linear sequence input.
[0054] In this embodiment, the symbol period of the second linear sequence is shortened, so that the first linear sequence can be multiplied with multiple data within one symbol, reducing the number of times the first linear sequence is repeatedly input and shortening the time used for multiplying two matrices.
[0055] In this embodiment, multiple multiplication matrices may also be cross-input to form a second linear arrangement sequence containing multiple matrix information, thereby achieving scale expansion of matrix operations.
[0056] Embodiment 4 This embodiment is a third embodiment of a control method for a multi-dimensional scalable matrix operation system. This embodiment is similar to the second embodiment, except that, in this embodiment, the control of the scale-expanded matrix operation system is achieved through wavelength division multiplexing, such as Figure 4 shown.
[0057] In this embodiment, the micro-ring modulator 400 is compatible with wavelength division multiplexing, and multiple carrier wavelengths are aligned with their optical resonances through precise dispersion management of Mach-Zehnder interferometers and micro-ring waveguides to maintain the same modulation response over the entire wavelength. The precise dispersion management mentioned here refers to the same symbol period.
[0058] The tunable laser 100 inputs M different wavelengths, serializes each row of the input matrix data to form different first linear sequences loaded on different wavelengths, combines different wavelength channels and inputs them to the micro-ring modulator 400, completes the matrix operation in parallel, and the output port of the micro-ring modulator 400 is divided into different wavelength channels, and the photodetector 600 is used to read out the optical signal of each wavelength. In this way, the matrix operation rate can be increased by M times.
[0059] Before combining different wavelength channels, multiple input matrices can be serialized to form multiple time series, which are encoded onto different optical carrier wavelengths using intensity modulator 200. Combining different wavelength channels and inputting into micro-ring modulator 400, multiple input matrix operations are completed in parallel, thus achieving the expansion of matrix scale.
[0060] Embodiment 5 This embodiment is a fourth embodiment of a control method for a multi-dimensional scalable matrix operation system. This embodiment is similar to the second embodiment, except that in this embodiment, multi-dimensional parallelism is used to realize the control of the scale-expanded matrix operation system, such as Figure 5 shown.
[0061] In this embodiment, the ring modulator can be compatible with wavelength division multiplexing by aligning multiple carrier wavelengths with their optical resonances to maintain the same modulation response across the wavelengths through precise dispersion management of Mach-Zehnder interferometers and micro-ring waveguides (micro-ring modulator 400).
[0062] In addition, the tunable laser 100 inputs M different wavelengths, serializes the data of the input matrix into different first linear sequences and loads them on different wavelengths. Further combining spatial parallelism, K micro-ring modulators 400 arrays are placed after combining different wavelength channels and share the combined input data.
[0063] By loading the data of each column of the multiplication matrix in parallel on the second linear sequence of the ring modulator, the multiplication operation of two matrices can be completed within one symbol period.
[0064] At the same time, M different wavelengths are used, and the M input matrices are serially formed into a time sequence, which is loaded onto different optical carriers using the intensity modulator 200, and further combined with spatial parallelism, K micro-ring modulator 400 arrays are placed after the combination of different wavelength channels, and the input is divided into wavelengths. The K multiplication matrices are serially formed into a sequence and loaded into the micro-ring modulator 400 array, and the computing throughput will increase from B using only time division multiplexing to BMK times using integrated wavelength division multiplexing and space division multiplexing, where B is the modulation symbol rate.
[0065] Embodiment 6 In this embodiment, if Figure 6 As shown, by using two ring modulators and two wavelengths for edge extraction, the extracted results are input into the computer's fully connected layer to build a neural network, and a handwritten digit classification experiment is carried out. The parallel scalability of the proposed scheme is verified experimentally. The TFLN micro-ring modulator used in the experiment has a bandwidth of more than 67 GHz, an area of 3.4mm×0.7 mm, and a free spectral range (FSR) of 0.148 nm. The selected wavelength corresponds to the optical resonance of the micro-ring modulator. In order to perform optical edge extraction, 300 handwritten digit images were selected, each with a resolution of 28×28 pixels, as shown in Figure 6 Due to the limitation of available test equipment, all 300 images were sequentially loaded to the first wavelength, convolved with the first kernel on the microring modulator, and then loaded to the next wavelength and convolved with the second kernel on the same microring modulator, as shown in (a). Figure 6 After convolution, the edge extraction results are as follows: Figure 6 During the experiment, the operating conditions of the microring modulator—including bias voltage, temperature, and electro-optical spectral response—remain unchanged for the input data at two wavelengths. Figure 6 The confusion matrix in (d) shows that the final classification accuracy is ~87%, which proves the feasibility of a multi-dimensional scalable matrix operation unit and its control method.
[0066] In the specific contents of the above-mentioned specific implementation methods, the various technical features can be combined in any non-contradictory manner. In order to make the description concise, not all possible combinations of the above-mentioned technical features are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0067] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.
Claims
1. A multi-dimensional scalable matrix operation system, characterized in that: include: Tunable laser (100): used to generate optical carrier signals of multiple carrier wavelengths; An intensity modulator (200): used to modulate an optical carrier signal according to a first linear sequence in the time domain, load the optical carrier signal onto the light intensity, and obtain an optical intensity signal carrying intensity information; A wavelength division multiplexer (300): used for multiplexing light intensity signals of different wavelengths and inputting them into a micro-ring modulator (400); A micro-ring modulator (400) is used to modulate the optical intensity signal according to a second linear sequence in the time domain to obtain a modulated signal; and to superimpose the modulated signals of different symbol periods through micro-ring feedback to obtain an output optical signal; the output optical signal carries the matrix multiplication and addition operation result of the first linear sequence and the second linear sequence; De-Wavelength Division Multiplexer (500): used to separate output optical signals of different wavelengths into various channels; Photodetector (600): used to detect the intensity of the output optical signal on each channel.
2. The multi-dimensional scalable matrix operation system according to claim 1, characterized in that: The first linear sequence is a time sequence formed by arranging the input matrices in a corresponding multiplication order; a plurality of input matrices simultaneously form a plurality of parallel time sequences in the same order, i.e., a plurality of first linear sequences; a plurality of the first linear sequences are encoded by a plurality of the intensity modulators (200) onto different optical carrier wavelengths; The second linear sequence is a time series formed by arranging the multiplication matrices in a multiplication order corresponding to the input matrix.
3. The multi-dimensional scalable matrix operation system according to claim 1, characterized in that: The micro-ring modulator (400) comprises a 2×2 first multi-mode interferometer (401) and a second multi-mode interferometer (402), and a Mach-Zehnder electro-optic modulator (403); the first multi-mode interferometer (401), the Mach-Zehnder electro-optic modulator (403), and the second multi-mode interferometer (402) are cascaded, and the first input port (404) of the first multi-mode interferometer (401) is connected to the first output port (406) of the second multi-mode interferometer (402) to form a micro-ring; the second input port (405) of the first multi-mode interferometer (401) is connected to the output end of the wavelength division multiplexer (300), and the second output port (407) of the second multi-mode interferometer (402) is connected to the input port of the photodetector (600); and the arm length of the Mach-Zehnder interferometer is controlled by regulating electrodes to perform intensity modulation.
4. A control method for a multi-dimensional scalable matrix operation system according to any one of claims 1 to 3, characterized in that: The following steps are involved: S1. Using a tunable laser (100) to output a plurality of optical carrier signals of different wavelengths; S2. Arranging the input matrix into a time series in the time domain according to the corresponding multiplication sequence to obtain a first linear sequence, and encoding the first linear sequence on optical signals of different wavelengths using an intensity modulator (200) to obtain an optical intensity signal carrying intensity information; S3. The optical intensity signals on different wavelength channels are loaded on the same channel through the wavelength division multiplexer (300) and input to different micro-ring modulators (400); S4. Arrange the multiplication matrix in a multiplication order corresponding to the input matrix to form a time sequence in the time domain, and obtain a second linear sequence; modulate the micro-ring modulator (400) according to the second linear sequence and in a time division multiplexing manner to obtain a modulated signal; use the delay of the micro-ring to control the superposition of modulated signals of different symbol periods to obtain an output optical signal, and control the output optical signal to be output to the wavelength division multiplexer (300); S5. The output optical signals of different wavelengths are separated into various channels through the wavelength division multiplexer (500), and the output of the photodetector (600) is controlled.
5. The control method of the multi-dimensional scalable matrix operation system according to claim 4, characterized in that: In step S2, the input matrix X Generate time series by row order x , time series x The amplitude of the optical carrier is serially modulated by an external intensity modulator (200) to obtain an optical signal carrying intensity information.
6. The control method of the multi-dimensional scalable matrix operation system according to claim 5, characterized in that: The symbol period is equal to one symbol period of the ring recursive delay ,Right now .
7. The control method of the multi-dimensional scalable matrix operation system according to claim 6, characterized in that: In step S4, the matrix W is multiplied Sort by columns; to get the input matrix X The multiplication and addition results of each row and column of the multiplication matrix W are traversed, and a column vector of the multiplication matrix W is repeatedly loaded until a time series x Modulation ends; when the next time series x When input again, continue to repeatedly load the next column vector of the multiplication matrix W until X All row and column multiplication and addition of W are completed.
8. The control method of the multi-dimensional scalable matrix operation system according to claim 7, characterized in that: In step S4, the multiplication matrix W is re-edited and converted into A second linear arrangement sequence is formed and loaded on the micro-ring modulator (400), namely: In the formula, The coupling matrix of the input micro-ring modulator is modulated on the optical intensity signal through the Mach-Zehnder electro-optic modulator; is the nth row and ith column element of the multiplication matrix W, The coupling matrix The element at row n and column i, is the last row and column i element of the multiplication matrix W, for In the coupling matrix Elements in subsequent rows of the same column.
9. The control method of the multi-dimensional scalable matrix operation system according to claim 8, characterized in that: The optical intensity signal is input into the micro-ring modulator (400) through the first input port (404) of the micro-ring modulator (400), and the multiplication operation of the input matrix and the multiplication matrix is realized by controlling the symbol period of the first linear sequence in the time domain to be equal to the symbol period of the second linear sequence loaded on the micro-ring modulator (400); the newly generated optical signal at the first output port (406) of the micro-ring modulator (400) returns to the second input port (405) along the micro-ring waveguide, wherein the ring recursive delay matches the symbol period, and the next input symbol is injected at the first input port (404); and then accumulation is performed through coherent optical superposition, and finally the amplitude of the modulated signal in the micro-ring modulator (400) satisfies: In the formula, is the final result Y i Line j The elements of the column, is the element in the last column of the i-th row of the input matrix X, The coupling matrix The last row and column j element, is the i-th row of the input matrix X. Column elements, The coupling matrix No. The element in row and column j, for In the coupling matrix Elements in subsequent rows of the same column, is the mth row and jth column element of the multiplication matrix W; Finally get No. i Line j The results of the columns also correspond to the input matrix X No. i Row and multiply matrix j The vector dot product of the columns is obtained by the photodetector (600). .
10. The control method of the multi-dimensional scalable matrix operation system according to claim 9, characterized in that: The signal encoded with amplitude in the micro-ring modulator (400) The result of the current operation is imported into the second output port (407) and read from the second output port (407) using the photoelectric detector (600).
Citation Information
Patent Citations
Photon tensor calculation acceleration method and device based on wavelength division multiplexing
CN114815959A
Photon two-dimensional convolution acceleration method and system based on time-wavelength interleaving
CN114819132A
Optoelectronic hybrid deep neural network operation device and operation method
CN116432726A
Photon matrix multiplication device and method for neural network
CN116484931A
Optical convolution accelerator and control method thereof
CN118607604A