An optical computing device, computing method, and computing system

Abstract

The application discloses an optical computing device, a computing method and a computing system. The optical computing device comprises a management unit, a first Ising unit and a second Ising unit. The management unit is connected to the first Ising unit and the second Ising unit. The first Ising unit and the second Ising unit receive a first group of signals. The first Ising unit generates a first group of feedback signals according to the first group of signals and a first problem sub-matrix. The second Ising unit generates a second group of feedback signals according to the first group of signals and a second problem sub-matrix. The first problem sub-matrix and the second problem sub-matrix are sub-matrices of a problem matrix, and the problem matrix indicates first data to be calculated. The management unit receives a first plurality of groups of feedback signals comprising the first group of feedback signals and the second group of feedback signals, and generates a first target feedback signal according to the first plurality of groups of feedback signals. The optical computing device provided by the application can increase the operation efficiency by using the parallel operation mode of the plurality of Ising units.

Description

Technical Field

[0001] This application relates to the field of information technology, and in particular to an optical computing device, computing method, and computing system. Background Technology

[0002] The Ising model describes a complex system with a large number of spin nodes, each with a spin state of +1 and -1. In this system, interactions exist between the spin nodes, altering their spin states. Based on these interactions, the Ising model progressively achieves an annealing process, meaning the hash rate of the system gradually decreases until convergence. Combinatorial optimization problems can be transformed into Ising models, utilizing the spin nodes and their interactions to characterize the parameters.

[0003] The optical Ising machine uses physical phenomena to simulate and solve the Ising model. It simulates the interaction between spin nodes using the interference between multiple optical signals. The phase of the optical signal represents the spin state of the spin node in the Ising model, and the interference between the optical signals can change the phase of the optical signal, thus simulating the change in the spin state of the spin node in the Ising model.

[0004] When the number of spin nodes in the Ising model increases, the number of optical signals in the optical Ising machine also needs to increase when simulating the Ising model. However, the number of optical signals that can be introduced into the optical Ising machine at one time by the current on-chip integration method is limited, and it cannot realize the simulation of Ising models with a large number of spin nodes. Summary of the Invention

[0005] This application provides an optical computing device, computing method, and computing system to enable the optical Ising machine to solve the Ising model with a large number of spin nodes.

[0006] In a first aspect, this application provides an optical computing device, which includes a management unit, a first Ising unit, and a second Ising unit. The management unit is connected to the first Ising unit and the second Ising unit. The optical computing device may include multiple Ising units (e.g., including m Ising units, where m is a positive integer not less than 2), and here only the first Ising unit and the second Ising unit included among the multiple Ising units are described as examples.

[0007] The first Ising unit and the second Ising unit can simultaneously receive the first set of signals. The first Ising unit generates a first set of feedback signals based on the first set of signals and the first problem submatrix. The second Ising unit generates a second set of feedback signals based on the first set of signals and the second problem submatrix, wherein the first problem submatrix and the second problem submatrix are different submatrices of the problem matrix, and the problem matrix is ​​used to indicate the first data to be calculated.

[0008] Each Ising unit can output a set of feedback signals; the management unit can receive a first set of multiple sets of feedback signals output by each Ising unit, the first set of multiple sets of feedback signals including a first set of feedback signals and a second set of feedback signals, and generate a first target feedback signal based on the first set of multiple sets of feedback signals. The first target feedback signal is used to instruct the optical computing device to perform Ising calculations on the first data to obtain a first intermediate result.

[0009] In the optical computing device provided in this application, the first Ising unit and the second Ising unit can operate in parallel. A set of feedback signals is generated based on the received first set of signals and the configured problem submatrix. When multiple Ising units operate in parallel, more optical signals can be received, supporting the simulation and solution of Ising models with a large number of spin nodes, and also increasing the computational efficiency.

[0010] In one possible design, the problem matrix may include a third and a fourth submatrix in addition to the first and second problem submatrixes. The third and fourth problem submatrixes are different from the first and second problem submatrixes.

[0011] The first Ising unit and the second Ising unit simultaneously receive the first set of signals. The first Ising unit can obtain the third set of feedback signals based on the first set of signals and the third problem submatrix; the second Ising unit generates the fourth set of feedback signals based on the first set of signals and the fourth problem submatrix; wherein, the third problem submatrix and the fourth problem submatrix are different submatrices of the problem matrix.

[0012] The management unit receives a second set of feedback signals, including a third set and a fourth set, and generates a second target feedback signal based on the second set of feedback signals. The second target feedback signal is used to instruct the optical computing device to perform a second intermediate result of the Ising calculation on the first data.

[0013] In the optical computing device provided in this application, the problem matrix is ​​decomposed into multiple sets of problem sub-matrices. For any set of problem sub-matrices, multiple problem sub-matrices (such as the first and second problem sub-matrices, the third problem matrix, and the fourth problem sub-matrices) can be pre-configured in multiple Ising units. Then, the multiple Ising units generate multiple sets of feedback signals in parallel based on the received first set of signals and the configured problem sub-matrices. The related operations for a problem matrix can be decomposed into multiple operation processes. For example, the calculation of multiple problem sub-matrices, including the first and second problem sub-matrices, by multiple Ising units is considered one operation; the calculation of multiple problem sub-matrices, including the third and fourth problem sub-matrices, by multiple Ising units is considered another operation. Correspondingly, multiple Ising units can also perform multiple parallel operations based on other problem sub-matrices. In this way, time-division multiplexing of Ising units is achieved during the operation of a problem matrix. The optical computing device completes one round of operation by performing these multiple operations. In this way, during the Ising calculation of the data to be calculated, not only are multiple Ising units used for parallel calculation, but also multiple Ising units can be used for time-sharing calculation (completing one calculation process in one time period and another calculation process in another time period). This can improve the utilization rate of Ising units, effectively increase the number of spin nodes in the simulated Ising model, and increase the computing efficiency of optical computing devices.

[0014] In one possible design, after performing one round of computation, the optical computing device can continue to perform multiple computations based on the target feedback signal generated during that round of computation (i.e., perform the next round of computation). The following is an example of the first computation in the subsequent multiple computations: the management unit can send the first target feedback signal to the first Ising unit and the second target feedback signal to the second Ising unit.

[0015] The first Ising unit can obtain the fifth set of feedback signals based on the first set of signals, the first target feedback signal, and the first problem submatrix; the second Ising unit can generate the sixth set of feedback signals based on the first set of signals, the second target feedback signal, and the second problem submatrix.

[0016] The management unit can receive a third set of multiple feedback signals, including a fifth set of feedback signals and a sixth set of feedback signals, and generate a third target feedback signal based on the third set of multiple feedback signals. The third target feedback signal is used to instruct the optical computing device to perform Ising calculations on the first data to obtain a third intermediate result.

[0017] In the optical computing device provided in this application, the optical computing device completes one round of computing process after completing multiple computing processes (such as m computing processes). The optical computing device can perform multiple rounds of computing processes to achieve iteration of multiple rounds of computing processes so as to accurately solve the Ising model.

[0018] In one possible design, the management unit can decompose the problem matrix and determine multiple sets of problem sub-matrices from the problem matrix. Each set of problem sub-matrices includes multiple problem sub-matrices, such as including a first problem matrix and a second problem sub-matrice, or including a third problem matrix and a fourth problem sub-matrice. The management unit can configure multiple problem sub-matrices in a set into multiple Ising units respectively.

[0019] In the optical computing device provided in this application, the management unit can decompose the problem matrix so that the optical computing device can simulate and solve the Ising model with a large number of nodes.

[0020] In one possible design, taking the first Ising unit as an example, the structures of the first and second Ising units are described below: The first Ising unit includes a spin signal generation module and a feedback calculation module. The spin signal generation module can generate a set of spin signals, and the feedback calculation module can generate a set of feedback signals based on the set of spin signals.

[0021] When the first Ising unit generates the first set of feedback signals, the spin signal generation module can obtain the first set of spin signals based on the first set of signals; the feedback calculation module can generate the first set of feedback signals based on the first set of spin signals and the first problem submatrix.

[0022] When the first Ising unit generates the third set of feedback signals, the spin signal generation module can obtain the second set of spin signals based on the first set of signals; the feedback calculation module can generate the third set of feedback signals based on the second set of spin signals and the third problem submatrix.

[0023] When the first Ising unit generates the fifth set of feedback signals, the spin signal generation module can obtain the third set of spin signals based on the first set of signals and the first target feedback signal; the feedback calculation module can generate the fifth set of feedback signals based on the third set of spin signals and the first problem submatrix.

[0024] In the optical computing device provided in this application, the spin signal generation module and the feedback calculation module in the first Ising unit can work together to generate a set of spin signals and a set of feedback signals, which simplifies the process of generating feedback signals.

[0025] In one possible design, the spin signal generation module can generate spin signals. The structure of the spin signal generation module is not limited here; for example, it can include a phase modulator array and an intensity modulator array. The spin signal generation module can first perform intensity modulation on a first set of signals, and then perform phase modulation. The intensity modulator array performs intensity modulation on the first set of signals to obtain a first set of modulated signals; then, the phase modulator array performs phase modulation on the first set of modulated signals to obtain a first set of spin signals. Alternatively, the spin signal generation module can first perform phase modulation on the first set of signals, and then perform intensity modulation. The phase modulator array performs phase modulation on the first set of signals to obtain a second set of modulated signals; then, the intensity modulator array performs intensity modulation on the second set of modulated signals to obtain a first set of spin signals.

[0026] In the optical computing device provided in this application, the spin signal generation module includes a phase modulator array and an intensity modulator array, which can specifically adjust the phase and intensity of the signal.

[0027] In one possible design, the intensity modulator array includes multiple intensity modulators, which can be Mach-Zehnder interferometers or electroabsorption modulators to suit different application scenarios.

[0028] In one possible design, the feedback calculation module includes multiple Mach-Zehnder interferometers to better load the problem submatrix (such as the first problem submatrix or the third problem submatrix).

[0029] In one possible design, after each round of computation by the optical computing device, the management unit calculates the Hamiltonian based on multiple target feedback signals. The Hamiltonian characterizes the system energy corresponding to the first data. When the Hamiltonian no longer decreases, i.e., when it converges, the optical computing device stops computation, and the management unit obtains the computation result of the first data based on multiple sets of spin signals generated in each Ising unit. When the Hamiltonian does not converge, the optical computing device continues computation until the Hamiltonian converges.

[0030] Secondly, this application provides a calculation method, the beneficial effects of which can be found in the relevant description of the first aspect, and will not be repeated here. This method is executed by an optical computing device, which includes a management unit, a first Ising unit, and a second Ising unit, with the management unit connected to the first and second Ising units.

[0031] The first Ising unit can obtain the first set of feedback signals based on the first set of signals and the first problem submatrix.

[0032] The second Ising unit can generate a second set of feedback signals based on the first set of signals and the second problem submatrix, wherein the first problem submatrix and the second problem submatrix are different submatrices of the problem matrix, and the problem matrix is ​​used to indicate the first data to be calculated.

[0033] The management unit receives a first set of multiple feedback signals, including a first set of feedback signals and a second set of feedback signals, and generates a first target feedback signal based on the first set of multiple feedback signals.

[0034] In one possible design, the first Ising unit can obtain the third set of feedback signals based on the first set of signals and the third problem submatrix;

[0035] The second Ising unit can generate a fourth set of feedback signals based on the first set of signals and the fourth problem submatrix; wherein the third problem submatrix and the fourth problem submatrix are different submatrices of the problem matrix;

[0036] The management unit can receive a second set of multiple feedback signals, including a third set of feedback signals and a fourth set of feedback signals, and generate a second target feedback signal based on the second set of multiple feedback signals.

[0037] In one possible design, the management unit can send the first target feedback signal to the first Ising unit and the second target feedback signal to the second Ising unit.

[0038] The first Ising unit can obtain the fifth set of feedback signals based on the first set of signals, the first target feedback signal, and the first problem submatrix.

[0039] The second Ising unit can generate a sixth set of feedback signals based on the first set of signals, the second target feedback signal, and the second problem submatrix.

[0040] The management unit can receive a third set of multiple feedback signals, including the fifth set of feedback signals and the sixth set of feedback signals, and generate a third target feedback signal based on the third set of multiple feedback signals.

[0041] Thirdly, this application provides a computing system that may include an optical computing device as described in the first aspect or any possible implementation thereof, and a laser for sending a first set of signals to the optical computing device.

[0042] Fourthly, this application also provides a computer program product, including program code, wherein the instructions included in the program code are executed by a computer to implement the computation method in the second aspect or any implementation thereof.

[0043] Fifthly, this application also provides a computer-readable storage medium for storing program code, the program code including instructions that are executed by a computer to implement the computation method in the second aspect or any implementation thereof. Attached Figure Description

[0044] Figure 1 This application provides a schematic diagram of the structure of an optical computing device;

[0045] Figure 2 A schematic diagram illustrating the decomposition of a problem submatrix provided in this application;

[0046] Figure 3 This application provides a schematic diagram of the structure of a spin signal generation module;

[0047] Figure 4 A schematic diagram of the structure of a feedback calculation module provided in this application;

[0048] Figure 5 A schematic diagram of a single operation process provided for this application;

[0049] Figure 6 A schematic diagram of a round of computation provided for this application;

[0050] Figure 7 A schematic diagram of a computing system provided in this application;

[0051] Figure 8 A schematic diagram of a calculation method provided in this application. Detailed Implementation

[0052] like Figure 1 As shown, an optical computing device 10 is provided in an embodiment of this application. The optical computing device 10 includes a management unit 100 and m Ising units 200, where m is an integer not less than 2. It should be noted that in this embodiment of the invention, the optical computing device 10 can be in the form of a circuit or a chip; the specific implementation of the optical computing device 10 is not limited here. For example... Figure 1 As shown, the management unit 100 can decompose the problem matrix into m groups of problem sub-matrices (in this embodiment, the problem sub-matrices are sub-matrices of the problem matrix). Each group of problem sub-matrices includes m problem sub-matrices. For any group of problem sub-matrices, the management unit 100 can configure the m problem sub-matrices of that group into m Ising units 200. Each Ising unit 200 is configured with one problem sub-matrice from that group of problem sub-matrices.

[0053] For any one of the m Ising elements 200, a set of feedback signals can be generated based on the received first set of signals and the configured problem submatrix. Specifically, the Ising element 200 can generate a set of spin signals based on the first set of signals (and the target feedback signal), and then generate a set of feedback signals using the configured problem submatrix and the set of spin signals.

[0054] The management unit 100 can also receive a set of feedback signals output from each Ising unit 200. The management unit 100 can generate a target feedback signal based on the acquired m sets of feedback signals. After generating m target feedback signals, the management unit 100 can send these m target feedback signals to the m Ising units 200 respectively. Each Ising unit 200 receives one target feedback signal from the m target feedback signals.

[0055] Specifically, within any Ising unit 200, each Ising unit 200 includes a spin signal generation module 210 and a feedback calculation module 220.

[0056] When the management unit 100 configures m problem submatrices of a set of problem submatrices into m Ising units 200, the m problem submatrices of this set of problem submatrices can be configured into feedback calculation modules 220 in the m Ising units 200. Each feedback calculation module 220 is configured with one problem submatric from this set of problem submatrices.

[0057] When the management unit 100 sends the m target feedback signals to the m Ising units 200 respectively, the m target feedback signals can be sent to the spin signal generation module 210 in each Ising unit 200. Each spin signal generation module 210 receives one target feedback signal from the m target feedback signals.

[0058] The spin signal generation module 210 is used to generate spin signals. The spin signal generation module 210 can generate a set of spin signals based on the received first set of signals. The set of spin signals can be transmitted to the feedback calculation module 220 in the Ising unit 200.

[0059] It should be noted that when the spin signal generation module 210 receives the target feedback signal, when generating a set of spin signals, the spin signal generation module 210 can generate a set of spin signals based on the first set of signals received and the target feedback signal.

[0060] The feedback calculation module 220 in the Ising unit 200 can generate a set of feedback signals based on the received set of spin signals and the configured problem submatrix.

[0061] The signal transmission process in the optical computing device 10 of this application embodiment is described below. In this application embodiment, since the problem sub-matrices configured simultaneously on the m feedback computing modules 220 are a set of problem sub-matrices decomposed from the problem matrix, the management unit 100 configures a set of problem sub-matrices for each of the m feedback computing modules 220. The m Ising units 200 can run together to generate m sets of feedback signals. For ease of explanation, the process of the management unit 100 configuring a set of problem sub-matrices for the m feedback computing modules 220 and the m Ising units 200 running to generate m sets of feedback signals is called one operation process. The process of the management unit 100 configuring m sets of problem sub-matrices (i.e., configuring the problem matrix) for the m feedback computing modules 220 and the m Ising units 200 running to generate m*m sets of feedback signals is an m-time operation process, which is called one round of operation. During each computation, the management unit 100 receives a set of feedback signals (a total of m sets of feedback signals) output by each Ising unit 200. It can generate a target feedback signal based on these m sets of feedback signals. When the optical computing device 10 performs m computations, the management unit 100 can generate a total of m target feedback signals. The management unit 100 can then send these m target feedback signals to the m Ising units 200 (spin signal generation module 210). Afterward, the optical computing device 10 can perform m more computations, i.e., another round of computation. In this way, the m target feedback signals generated in each round of computation are sent to the m Ising units 200 (spin signal generation module 210) for the next round of computation, thus achieving iterative multi-round computation.

[0062] The iterations of a single operation, a single round of operation, and multiple rounds of operation are explained below.

[0063] (1) One-time calculation process

[0064] The management unit 100 can first decompose the problem matrix. The method of decomposing the problem matrix by the management unit 100 is not limited here; any method that can decompose the problem matrix into m sub-matrices is applicable to the embodiments of this application. The embodiments of this application do not limit the structure of the management unit 100. For example, the management unit 100 can be a field-programmable gate array (FPGA) or a digital signal processing chip (DSP).

[0065] The problem matrix is ​​used to indicate the first data to be calculated. This first data can also be understood as the Ising model transformed from a combinatorial optimization problem. The problem matrix can be a matrix obtained after data simulation and extraction of the combinatorial optimization problem. Each element in the matrix represents the interaction between different spin nodes in the Ising model. The problem matrix can usually be a symmetric matrix.

[0066] like Figure 2 The diagram shown is a schematic diagram of a problem matrix decomposition provided in an embodiment of this application. The management unit 100 can decompose the problem matrix into m*m problem submatrices J. n*n Each problem submatrix J n*n It consists of n*n elements, with each m submatrix of problems divided into a group of submatrixes. Figure 2 The problem submatrix is ​​divided into columns, with each column forming a group of problem submatrixes. The management unit 100 can sequentially assign this group of problem submatrixes to m Ising units 200 (feedback calculation modules 220).

[0067] During a single operation, each Ising unit 200 can simultaneously receive a first set of signals, which includes at least one signal. In this embodiment, the example is given where the first set of signals includes n signals. This embodiment does not limit the type of signals in the first set; they can be optical signals or electrical signals. Each Ising unit 200 generates a set of feedback signals based on the received first set of signals and the configured problem submatrix.

[0068] The following example illustrates a calculation process involving the first and second Ising units among the multiple Ising units 200, using the management unit 100 configuring the first problem submatrix in the first set of problem submatrixes. The management unit 100 configures the first problem submatrix in the first set of problem submatrixes in the first Ising unit, and the management unit 100 configures the second problem submatrix in the first set of problem submatrixes in the second Ising unit.

[0069] The first Ising unit can obtain the first set of feedback signals based on the first set of signals and the first problem submatrix; the second Ising unit can generate the second set of feedback signals based on the first set of signals and the second problem submatrix.

[0070] The management unit 100 receives multiple sets of feedback signals output by multiple Ising units 200 (for ease of distinction, these multiple sets of feedback signals are referred to here as the first multiple sets of feedback signals), wherein the first multiple sets of feedback signals include a first set of feedback signals and a second set of feedback signals, and the management unit 100 generates a first target feedback signal based on the first multiple sets of feedback signals.

[0071] exist Figure 1In the optical computing device shown, each Ising unit 200 outputs a set of feedback signals. The management unit 100 can receive a set of feedback signals output by m Ising units 200, and can receive a total of m sets of feedback signals. These m sets of feedback signals are the first multiple sets of feedback signals. A first target feedback signal can be generated based on these m sets of feedback signals. The first target feedback signal is the first intermediate result of the optical computing device performing Ising calculation on the first data.

[0072] For any Ising unit 200, the spin signal generation module 210 in the Ising unit 200 can adjust the phase and intensity of each signal in the first set of signals to output a set of spin signals. The set of spin signals includes n spin signals.

[0073] This application embodiment does not limit the way the spin signal generation module 210 adjusts the phase and intensity of each signal in the first group of signals. For different types of signals, the way the spin signal generation module 210 adjusts the phase and intensity of each signal in the first group of signals is also different. The following uses an optical signal (such as a light pulse) in the first group of signals as an example to illustrate the way the spin signal generation module 210 adjusts the phase and intensity of each signal in the first group of signals:

[0074] See Figure 3 This is a schematic diagram of the structure of a spin signal generation module 210 provided in an embodiment of this application. The spin signal generation module 210 includes an intensity modulator array 211 and a phase modulator array 212.

[0075] The intensity modulator array 211 includes multiple intensity modulators capable of intensity modulating the received signal. The phase modulator array 212 includes multiple phase modulators capable of phase modulating the received signal.

[0076] The intensity modulator can be a Mach-Zehnder interferometer (MZI), an electro-absorption modulator (EAM), a semiconductor optical amplifier (SOA), or a variable optical attenuator (VOA). The phase modulator can be a waveguide.

[0077] The embodiments of this application do not limit the order in which intensity modulation and phase modulation are performed on the first group of signals. Intensity modulation can be performed first, followed by phase modulation (intensity modulator array 211 first receives the first group of signals, performs intensity modulation on the first group of signals, and outputs a modulated signal; phase modulator array 212 performs phase modulation on the modulated signal output by intensity modulator array 211, and outputs a spin signal); or phase modulation can be performed first, followed by intensity modulation (phase modulator array 212 first receives the first group of signals, performs phase modulation on the first group of signals, and intensity modulator array 211 performs intensity modulation on the signal output by phase modulator array 212, and outputs a spin signal).

[0078] If the signal in the first group of signals is an electrical signal, the spin signal generation module 210 can be an FPGA or a DSP to achieve intensity modulation of the electrical signal.

[0079] The structure of the spin signal generation module 210 described above is merely an example. The embodiments of this application do not limit the structure of the spin signal generation module 210. Any module that can generate a set of spin signals based on the first set of signals is applicable to the embodiments of this application.

[0080] After the spin signal generation module 210 outputs a set of spin signals, the feedback calculation module 220 receives the set of spin signals and generates a set of feedback signals based on the configured problem submatrix and the set of spin signals.

[0081] This application embodiment does not limit the way in which the feedback calculation module 220 generates a set of feedback signals based on the problem submatrix and a set of spin signals. For different types of signals, the way in which the feedback calculation module 220 generates a set of feedback signals based on the problem submatrix and a set of spin signals is also different. The following uses an optical signal in a set of spin signals as an example to illustrate the way in which the feedback calculation module 220 generates a set of feedback signals based on the problem submatrix and a set of spin signals.

[0082] See Figure 4 This is a schematic diagram of the structure of a feedback calculation module 220 provided in an embodiment of this application. The feedback calculation module 220 includes an interaction matrix 221 and a photodetector array 222.

[0083] Interaction matrix 221 includes multiple Mach-Zehnder interferometers (MZIs), the number of which corresponds to the problem submatrix J. n*n The number of elements in the matrix is ​​related to the number of elements. In this embodiment, the interaction matrix 221 includes n. 2 A Mach-Zehnder interferometer, utilizing n 2 The submatrix J of the Machzed interferometer construction problem n*nThe phase modulator unit 213 on each Mach-Zehnder interferometer can be implemented using thin-film lithium niobate, and the thin-film lithium niobate and the transmission waveguide are integrated together using heterogeneous integration technology. The J-phase interferometer is implemented using the electro-optic properties of thin-film lithium niobate. n*n Loading, using a transmission waveguide, can realize the problem submatrix J n*n The time-sharing refresh. 2 The submatrix J of the Machzed interferometer construction problem n*n When, the problem submatrix J can be... n*n The matrix is ​​decomposed into a chief matrix U, a diagonal matrix ∑, and a transpose chief matrix U. A set of spin signals (which can be viewed as column vectors) are transmitted through the interaction matrix 221, completing the matrix-vector multiplication operation and generating a set of feedback optical signals. The set of feedback optical signals includes n feedback optical signals.

[0084] The photodetector array 222 includes n photodetectors. The photodetectors can convert light signals into electrical signals. A set of feedback light signals passes through the photodetector array 222 to achieve photoelectric conversion and output a set of feedback signals, which includes n feedback signals.

[0085] See Figure 5 Taking the configuration of the first set of problem sub-matrices by the management unit 100 for multiple Ising units 200 as an example, the feedback calculation module 220 in the first Ising unit and the feedback calculation module 220 in the second Ising unit are used to explain a set of feedback signals. The first set of spin signals output by the spin signal generation module 210 in the first Ising unit is {σ1, σ2}, and the first problem sub-matrix loaded by the feedback calculation module 220 in the first Ising unit is... The first set of feedback signals output is The second set of spin signals output by the spin signal generation module 210 in the second Ising unit is {σ3, σ4}, and the second problem submatrix loaded by the feedback calculation module 220 in the second Ising unit is... The second set of feedback signals output is

[0086] and The calculation method is as follows:

[0087]

[0088]

[0089] Using J N*N The properties of symmetric matrices (J) ij =J ji The above formula can be transformed as follows:

[0090]

[0091]

[0092] The management unit 100 acquires m sets of feedback signals, sums the m sets of feedback signals, and generates the first target feedback signal.

[0093] Still with Figure 5 The first and second sets of feedback signals shown are used as examples for explanation. The management unit 100... and Summation generates the first target feedback signal.

[0094]

[0095]

[0096] Where N is the total number of spin signals (m*n) included in the m sets of spin signals output by the m spin signal generation modules 210. In the example above, m and n can be considered to be 2, and N is 4.

[0097] Based on the foregoing reasoning, any target feedback signal is a signal sequence comprising n signals, where the j-th signal f in the target feedback signal is... j as follows:

[0098]

[0099] Where N is the total number of spin signals (m*n) included in the m sets of spin signals output by the m spin signal generation modules 210, and σ i Let i be the i-th spin signal among all spin signals in the m-group of spin signals.

[0100] As can be seen from the above, a target feedback signal is the result of the operation between the spin signal and a set of problem submatrices in the problem matrix. In the embodiment of this application, the first target feedback signal is the result of the operation between the spin signal generated by m Ising units and the first set of problem submatrices during this operation.

[0101] (2) One round of calculation process

[0102] In one computational process, the m feedback computation modules 220 only load m sub-matrices from the problem matrix. Let t be the time taken for the optical computing device 10 to perform one computation, and t1 be the time taken for the first computation by the optical computing device 10. If the m feedback computation modules 220 load m*m sub-matrices from the problem matrix, then the required time is m*t.

[0103] The following explanation uses the aforementioned single operation as the first operation process to illustrate one operation process in the remaining m-1 operations of one round of operations:

[0104] Management unit 100 configures the m problem submatrices from another set of problem submatrices into m Ising units 200, with each Ising unit 200 configuring one problem submatric.

[0105] During this operation, each Ising unit 200 simultaneously receives the first set of signals, and each Ising unit 200 generates a set of feedback signals based on the received first set of signals and the configured problem submatrix.

[0106] Taking the configuration of the second set of problem submatrices by the management unit 100 for multiple Ising units 200 as an example, the following describes a calculation process involving the first and second Ising units among the multiple Ising units 200. The management unit 100 configures the third problem submatrix from the second set of problem submatrices in the first Ising unit, and configures the fourth problem submatrix from the second set of problem submatrices in the second Ising unit. The second set of problem submatrices is different from the first set of problem submatrices; that is, the second problem submatrices and the first problem submatrices, as well as the third and fourth problem submatrices, are different submatrices within the problem matrix.

[0107] The first Ising unit can obtain the third set of feedback signals based on the first set of signals and the third problem submatrix; the second Ising unit can generate the fourth set of feedback signals based on the first set of signals and the fourth problem submatrix.

[0108] The management unit 100 receives multiple sets of feedback signals output by multiple Ising units 200 (for ease of distinction, these multiple sets of feedback signals are referred to here as the second set of feedback signals), wherein the second set of feedback signals includes a third set of feedback signals and a fourth set of feedback signals, and the management unit 100 generates a second target feedback signal based on the second set of feedback signals.

[0109] exist Figure 1 In the optical computing device shown, each Ising unit 200 outputs a set of feedback signals. The management unit 100 can receive a set of feedback signals output by m Ising units 200, and can receive a total of m sets of feedback signals. These m sets of feedback signals are the second set of feedback signals. A second target feedback signal can be generated based on these m sets of feedback signals. The second target feedback signal is the second intermediate result of the optical computing device performing Ising calculation on the first data.

[0110] In this embodiment of the application, the second target feedback signal is the result of the operation between the spin signals generated by the m Ising units and the second set of problem sub-matrices during this operation.

[0111] For any Ising unit 200, the spin signal generation module 210 in the Ising unit 200 performs intensity modulation and phase modulation on the first set of signals, and outputs a set of spin signals. The manner in which the spin signal generation module 210 outputs a set of spin signals can be referred to the foregoing description, and will not be repeated here.

[0112] Each feedback calculation module 220 generates a set of feedback signals based on the configured problem submatrix and a set of spin signals. The method by which the feedback calculation module 220 generates a set of feedback signals based on the problem submatrix and a set of spin signals can be found in the foregoing description, and will not be repeated here.

[0113] The following example illustrates one round of computation by the optical computing device 10, where each computation is performed by the optical computing device 10, and a column of a submatrix of the problem matrix is ​​loaded into the m feedback computation modules 220.

[0114] like Figure 6 As shown, in one round of calculation of the optical computing device 10, during the m operations, the problem submatrix is ​​loaded in the m feedback computing modules 220 in each operation.

[0115] from Figure 6 During the first calculation of the optical computing device 10 within time t1, m feedback computing modules 220 load the first column of the problem submatrix from the problem matrix, and the m feedback computing modules 220 output m sets of feedback signals {f 11 f 21 , ...f m1}, where f 11 f is a set of feedback signals output by the first feedback calculation module 220 within time t1. 21 f is a set of feedback signals output by the second feedback calculation module 220 within time t1. m1 This refers to a set of feedback signals output by the m-th feedback calculation module 220 within time t1; the management unit 100 outputs the m sets of feedback signals {f 11 f 21 , ...f m1 Sum the results and output a target feedback signal f. 1 Target feedback signal f 1 Given a signal sequence consisting of n signals. 11 f 21 , ...f m1}

[0116] During time t2, as the optical computing device 10 performs its second calculation, m feedback computing modules 220 load the second column of the problem submatrix from the problem matrix, and the m feedback computing modules 220 output m sets of feedback signals {f 12 f 22, ...f m2}, where f 12 f is a set of feedback signals output by the first feedback calculation module 220 within time t2. 22 f is a set of feedback signals output by the second feedback calculation module 220 within time t2. m2 This refers to a set of feedback signals output by the m-th feedback calculation module 220 within time t2; the management unit 100 outputs the m sets of feedback signals {f 12 f 22 , ...f m2 Sum the results and output a target feedback signal f. 2 Target feedback signal f 2 It is a signal sequence that includes n signals.

[0117] During the time interval tm, during the m-th calculation of the optical computing device 10, the m feedback computing modules 220 load the m-th column of the problem sub-matrix from the problem matrix, and the m feedback computing modules 220 output m sets of feedback signals {f 1m f 2m , ...f mm}, where f 1m f is a set of feedback signals output by the first feedback calculation module 220 within time tm. 2m f is a set of feedback signals output by the second feedback calculation module 220 within the time interval tm. mm This refers to a set of feedback signals output by the m-th feedback calculation module 220 within a time period tm. The management unit 100 processes the m sets of feedback signals {f}. 1m f 2m , ...f mm Sum the results and output a target feedback signal f. m Target feedback signal f m It is a signal sequence that includes n signals.

[0118] After the m-th computation process of the optical computing device 10 is completed, the optical computing device 10 completes one round of computation. During this round of computation, the optical computing device 10 generates a total of m target feedback signals {f}. 1 f 2 , ...f m}

[0119] (3) Iteration of multi-round operation process

[0120] After the optical computing device 10 completes one round of computation, the management unit 100 will generate m target feedback signals {f}. 1 f 2 , ...f mIn order to perform the next round of calculation, the management unit 100 can send m target feedback signals {f} 1 f 2 , ...f m The signal is sent to m Ising units 200 respectively. Each Ising unit 200 receives a target feedback signal.

[0121] Specifically, management unit 100 can transmit feedback signals {f} from m targets. 1 f 2 , ...f m The signal is sent sequentially to each of the m Ising units 200 according to their arrangement. For example, the management unit 100 can send the target feedback signal f... 1 The signal is sent to the first Ising unit 200, and the management unit 100 can then send the target feedback signal f. 2 Send to the second Isin unit 200. Management unit 100 can then send the target feedback signal f. m Send to the m-th Ising unit 200.

[0122] When the management unit 100 sends a target feedback signal to an Ising unit 200, it can send the target feedback signal to the spin signal generation module 210 of the Ising unit 200.

[0123] The management unit 100 will send the target feedback signal f 1 Taking the method of sending to a spin signal generation module 210 as an example, the target feedback signal f 1 The system includes n signals. The management unit 100 loads the n signals onto the n intensity modulators of the intensity modulator array 211 in the spin signal generation module 210, with one signal from the n signals loaded onto each intensity modulator.

[0124] Management unit 100 sends feedback signals {f} from m targets. 1 f 2 , ...f m After the signal is sent to m Ising units 200 (spin signal generation module 210), the optical computing device 10 can perform m calculations, that is, one round of calculation.

[0125] The optical computing device 10 can perform a round of computation as described above. It should be noted that, since each spin signal generation module 210 is loaded with a target feedback signal, when each spin signal generation module 210 performs intensity modulation and phase modulation on the first group of signals during one computation, it can perform intensity modulation and phase modulation on the first group of signals based on the target feedback signal.

[0126] When the spin signal generation module 210 performs intensity modulation and phase modulation on the first group of signals based on the target feedback signal, if the signal in the first group of signals is an optical signal, the spin signal generation module 210 can perform intensity and phase modulation on the optical signal in the first group of signals according to the amplitude and phase of the target feedback signal. If the signal in the first group of signals is an electrical signal, the spin signal generation module 210 can perform intensity modulation on the electrical signal in the first group of signals according to the amplitude of the target feedback signal. Each Ising unit 200 generates a set of feedback signals based on the received first group of signals, the target feedback signal, and the configured problem submatrix.

[0127] The following example illustrates a calculation process involving the first and second Ising units among the multiple Ising units 200, using the configuration of the first set of problem sub-matrices by the management unit 100. The management unit 100 configures the first problem sub-matrix in the first set of problem sub-matrices in the first Ising unit, and configures the second problem sub-matrix in the first set of problem sub-matrices in the second Ising unit. The management unit 100 sends the first target feedback signal to the first Ising unit and the second target feedback signal to the second Ising unit.

[0128] The first Ising unit can obtain the fifth set of feedback signals based on the first set of signals, the first target feedback signal, and the first problem submatrix; the second Ising unit can generate the sixth set of feedback signals based on the first set of signals, the second target feedback signal, and the second problem submatrix.

[0129] The management unit 100 receives multiple sets of feedback signals output by multiple Ising units 200 (for ease of distinction, these multiple sets of feedback signals are referred to here as the third set of feedback signals), wherein the third set of feedback signals includes the fifth set of feedback signals and the sixth set of feedback signals, and the management unit 100 generates a third target feedback signal based on the third set of feedback signals.

[0130] For any Ising unit 200, the spin signal generation module 210 in the Ising unit 200 outputs a set of spin signals based on a set of signals and a target feedback signal.

[0131] The feedback calculation module 220 in the Ising unit 200 generates a set of feedback signals based on the problem submatrix and the set of spin signals;

[0132] The management unit 100 receives a set of feedback signals output by the feedback calculation module 220 in m Ising units 200, and can sum the m sets of feedback signals to generate a third target feedback signal.

[0133] exist Figure 1In the optical computing device shown, each Ising unit 200 outputs a set of feedback signals. The management unit 100 can receive a set of feedback signals output by m Ising units 200, and can receive a total of m sets of feedback signals. These m sets of feedback signals are the third set of feedback signals. A third target feedback signal can be generated based on these m sets of feedback signals. The third target feedback signal is the third intermediate result of the optical computing device performing Ising calculation on the first data.

[0134] In this embodiment, the third target feedback signal is the result of another operation between the spin signals generated by the m Ising units during this operation and the first set of problem submatrices. The spin signals here are generated based on the first target feedback signal.

[0135] After each round of calculation, the management unit 100 can use the m target feedback signals determined in that round to calculate the density H. Where, σ i and σ j It is the spin signal in a set of spin signals output by the spin signal generation module 210 during this round of calculation.

[0136] The Hafn density H indicates the system energy corresponding to the first data. When the system energy corresponding to the first data no longer decreases, that is, when the system energy of the Ising model no longer decreases, meaning the Hafn density H converges and stops decreasing, the Ising model solved by the optical computing device 10 has converged, and the optical computing device 10 can stop the operation. At this time, a set of spin signals output by each spin signal generation module 210 is the optimal solution. If the system energy corresponding to the first data continues to decrease, that is, when the system energy of the Ising model decreases, meaning the Hafn density H does not converge and continues to decrease, the optical computing device 10 still needs to continue the operation, that is, continue to perform another round of operation, until the Hafn density H converges.

[0137] like Figure 7 As shown, a computing system provided in an embodiment of this application is included in the computing system, which includes an optical computing device 10 and a laser 20. The laser 20 is capable of outputting a first set of signals so that the optical computing device 10 can receive the first set of signals and perform calculations.

[0138] To make the solution clearer, the following will refer to the previous embodiments, as follows: Figure 8 Taking the calculation method shown as an example, the workflow of the optical computing device provided in this embodiment of the invention will be summarized. Figure 8 As shown, during the operation: the management unit 100 decomposes the problem matrix and determines the first set of problem sub-matrices from it. The management unit 100 then configures the first problem sub-matrix in the first set of problem sub-matrices into the first Ising unit. The management unit 100 then configures the second problem sub-matrix in the first set of problem sub-matrices into the second Ising unit.

[0139] The first Ising unit receives the first set of signals (step 1) and generates the first set of feedback signals based on the received first set of signals and the first problem submatrix (step 2).

[0140] The first Ising unit receives the first set of signals (step 3) and generates the first set of feedback signals based on the received first set of signals and the second problem submatrix (step 4).

[0141] The management unit 100 receives the first multiple sets of feedback signals (including the first set of feedback signals and the second set of feedback signals) (step 5), and generates the first target feedback signal based on the first multiple sets of feedback signals (step 6).

[0142] The above process describes the execution of one computation in an optical computing device. The optical computing device can continue to execute m-1 computations, where one computation is as follows:

[0143] Management unit 100 configures the third problem submatrix in the second group of problem submatrixes into the first Ising unit, and management unit 100 configures the fourth problem submatrix in the second group of problem submatrixes into the second Ising unit.

[0144] The first Ising unit can obtain the third set of feedback signals based on the first set of signals and the third problem submatrix; the second Ising unit can generate the fourth set of feedback signals based on the first set of signals and the fourth problem submatrix.

[0145] The management unit 100 receives a second set of multiple feedback signals, wherein the second set of multiple feedback signals includes a third set of feedback signals and a third set of feedback signals. The management unit 100 generates a second target feedback signal based on the second set of multiple feedback signals.

[0146] In the foregoing description, the m operations performed by the optical computing device constitute one round of computation. The optical computing device can perform multiple rounds of computation in a similar manner. Taking the aforementioned process as the first round of computation, the following description uses the first operation in the next round of computation as an example to illustrate any operation in the next round of computation performed by the optical computing device:

[0147] The management unit 100 sends the first target feedback signal to the first Ising unit, and the management unit 100 sends the second target feedback signal to the second Ising unit.

[0148] The first Ising unit can obtain the fifth set of feedback signals based on the first set of signals, the first target feedback signal, and the first problem submatrix; the second Ising unit can generate the sixth set of feedback signals based on the first set of signals, the second target feedback signal, and the second problem submatrix.

[0149] The management unit 100 receives multiple sets of feedback signals output by multiple Ising units 200 (for ease of distinction, these multiple sets of feedback signals are referred to here as the third set of feedback signals), wherein the third set of feedback signals includes the fifth set of feedback signals and the sixth set of feedback signals, and the management unit 100 generates a third target feedback signal based on the third set of feedback signals.

[0150] The optical computing device provided in this invention decomposes the process of solving the Ising model into one or more rounds of computation. Each round of computation includes m computations (time-sharing computation), and each computation can be implemented using m Ising units in parallel (parallel computation). This enables the solution of Ising models with a large number of spin nodes, improving computational efficiency.

[0151] Furthermore, since the optical computing device provided in this embodiment of the invention has a simple structure and can be implemented on a chip, and the entire computing process is implemented through optical signals or electrical signals, the signal transmission speed is fast and the computing speed is greatly improved. Therefore, the computing device improved in this embodiment of the invention can be applied to neural network systems, for example, it can be used to implement feedback control in neural network systems.

[0152] It should be noted that the embodiments provided in this application are merely illustrative. Those skilled in the art will understand that, for the sake of convenience and brevity, the descriptions of each embodiment have different focuses, and parts not described in detail in a certain embodiment can be referred to in the relevant descriptions of other embodiments. The features disclosed in the embodiments, claims, and drawings of this invention can exist independently or in combination. Features described in hardware form in the embodiments of this invention can be executed by software, and vice versa. No limitations are imposed here.

Claims

1. An optical computing device, characterized in that, include: The first Ising unit is used to obtain a first set of feedback signals based on a first set of signals and a first problem submatrix, wherein the first set of signals is an optical signal or an electrical signal. The second Ising unit is used to generate a second set of feedback signals based on the first set of signals and the second problem submatrix, wherein the first problem submatrix and the second problem submatrix are different submatrixes of the problem matrix, and the problem matrix is ​​used to indicate the first data to be calculated. The management unit, connected to the first Ising unit and the second Ising unit, is used to receive a first set of multiple feedback signals including the first set of feedback signals and the second set of feedback signals, and to generate a first target feedback signal based on the first set of multiple feedback signals. The first target feedback signal is used to instruct the optical computing device to perform Ising calculation on the first data as a first intermediate result. The first Ising unit is also used to obtain a third set of feedback signals based on the first set of signals and the third problem submatrix; The second Ising unit is further configured to generate a fourth set of feedback signals based on the first set of signals and the fourth question submatrix; wherein the third question submatrix and the fourth question submatrix are different submatrices of the question matrix; The management unit is further configured to receive a second set of multiple feedback signals, including the third set of feedback signals and the fourth set of feedback signals, and generate a second target feedback signal based on the second set of multiple feedback signals.

2. The device as described in claim 1, characterized in that, The management unit is further configured to send the first target feedback signal to the first Ising unit and the second target feedback signal to the second Ising unit; The first Ising unit is further configured to obtain a fifth set of feedback signals based on the first set of signals, the first target feedback signal, and the first problem submatrix; The second Ising unit is used to generate a sixth set of feedback signals based on the first set of signals, the second target feedback signal, and the second problem submatrix; The management unit is configured to receive a third set of multiple feedback signals, including the fifth set of feedback signals and the sixth set of feedback signals, and generate a third target feedback signal based on the third set of multiple feedback signals.

3. The optical computing device according to any one of claims 1 to 2, characterized in that, The first Ising unit includes: A spin signal generation module is used to obtain a first set of spin signals based on the first set of signals; The feedback calculation module is used to generate a first set of feedback signals based on the first set of spin signals and the first problem submatrix.

4. The optical computing device according to claim 3, characterized in that, The spin signal generation module is used to perform intensity modulation and phase modulation on the first set of signals to obtain the first set of spin signals.

5. The device as described in claim 3, characterized in that, The spin signal generation module includes: An intensity modulator array is used to intensity modulate the first group of signals to obtain a first group of modulated signals; A phase modulator array is used to perform phase modulation on the first set of modulation signals to obtain the first set of spin signals.

6. The device as described in claim 4, characterized in that, The spin signal generation module includes: An intensity modulator array is used to intensity modulate the first group of signals to obtain a first group of modulated signals; A phase modulator array is used to perform phase modulation on the first set of modulation signals to obtain the first set of spin signals.

7. The device as described in claim 5 or 6, characterized in that, The intensity modulator array includes multiple Mach-Zehnder interferometers (MZI) or electroabsorption modulators (EAM).

8. The device as described in claim 5 or 6, characterized in that, The phase modulator array includes multiple waveguides.

9. The device as described in claim 3, characterized in that, The feedback calculation module includes multiple Mach-Zehnder interferometers (MZIs).

10. The device as claimed in claim 2, characterized in that, The management unit is also used for: The Hamiltonian is calculated based on the third target feedback signal, and the Hamiltonian is used to characterize the system energy corresponding to the first data. When the convergence of the Hamiltonian is determined, the calculation result of the first data is obtained based on the third set of spin signals.

11. A calculation method, characterized in that, The method is executed by an optical computing device, which includes a management unit, a first Ising unit, and a second Ising unit, wherein the management unit is connected to the first Ising unit and the second Ising unit; The first Ising unit obtains a first set of feedback signals based on a first set of signals and a first problem submatrix, wherein the first set of signals is an optical signal or an electrical signal; The second Ising unit generates a second set of feedback signals based on the first set of signals and the second problem submatrix, wherein the first problem submatrix and the second problem submatrix are different submatrixes of the problem matrix, and the problem matrix is ​​used to indicate the first data to be calculated; The management unit receives a first plurality of feedback signals including the first set of feedback signals and the second set of feedback signals, and generates a first target feedback signal based on the first plurality of feedback signals; The first Ising unit obtains the third set of feedback signals based on the first set of signals and the third problem submatrix; The second Ising unit generates a fourth set of feedback signals based on the first set of signals and the fourth question submatrix; wherein the third question submatrix and the fourth question submatrix are different submatrices of the question matrix; The management unit receives a second set of multiple feedback signals, including the third set of feedback signals and the fourth set of feedback signals, and generates a second target feedback signal based on the second set of multiple feedback signals.

12. The method as described in claim 11, characterized in that, The method also includes: The management unit sends the first target feedback signal to the first Ising unit and the second target feedback signal to the second Ising unit; The first Ising unit obtains the fifth set of feedback signals based on the first set of signals, the first target feedback signal, and the first problem submatrix; The second Ising unit generates a sixth set of feedback signals based on the first set of signals, the second target feedback signal, and the second problem submatrix; The management unit receives a third set of multiple feedback signals, including the fifth set of feedback signals and the sixth set of feedback signals, and generates a third target feedback signal based on the third set of multiple feedback signals.

13. A computing system, characterized in that, The system includes an optical computing device as described in any one of claims 1 to 10 and a laser, wherein the laser is used to send the first set of signals to the optical computing device.

Images