Optical computing device

Abstract

Provided is an optical computing device capable of utilizing the spatial, wavelength, and temporal parallelism of light. Specifically, provided is an optical computing device comprising: a light source that transmits optical signals of a plurality of wavelengths; a first light modulator array that modulates the optical signals in accordance with an input matrix of wavelength-domain rows and time-domain columns; an optical splitting means that spatially distributes the modulated optical signals of the plurality of wavelengths; a second light modulator array that modulates the spatially parallelized optical signals of the plurality of wavelengths into wavelength collective optical signals in accordance with a weighting matrix of time-domain rows and spatial-domain columns; a wavelength splitting means that separates output optical signals into a plurality of respective wavelengths; a photodetector array that detects the separated optical signals; and a time accumulation means that accumulates the optical signals of the plurality of wavelengths detected by the time domain for each wavelength domain and spatial domain and generates a prescribed matrix-matrix product signal.

Description

Optical computing device 【0001】 The present disclosure relates to an optical computing device, and more particularly to an optical computing device that can utilize the spatial, wavelength, and time parallelism of light. 【0002】 Attention has been focused on information processing by machine learning using a deep neural network (DNN). The DNN is a large-scale non-linear network in which a large number of neurons that perform non-linear responses are connected by synapses, and in particular, deep learning techniques using a network in which neurons are arranged in multiple layers have been widely applied. By using this technology, excellent performance has been reported in a wide range of fields such as image and voice recognition, robot control, and artificial data generation. 【0003】 On the other hand, as the number of layers and the number of neurons (e.g., the number of nodes) increase, the scale of the required matrix operations (Y = A·X) increases explosively, and the increase in the time and power required for inference and learning has become a major issue. In recent years, as a technique for solving such problems, attention has been focused on an information processing technique that uses an optical circuit as a matrix operation element. FIG. 1 shows the principles of the main conventional calculation configurations, the space division multiplexing (SDM) type and the wavelength division multiplexing (WDM) type. 【0004】 In the SDM type, for example, matrix calculation is executed using a configuration in which Mach-Zehnder interferometers (MZIs) having N-port input and output are arranged in a mesh shape (see Non-Patent Document 1). In this configuration, the input vector X ∈ C N is input from each optical port, and the weight matrix A ∈ C represented by the branching ratio and phase shift of the MZI N×NUsing this method, it is possible to perform operations at a rate of OPS to N²B operations per second per unit time. Here, OPS (Operations Per Second) is the number of operations per unit time, and B is the baud rate of the optical signal. In the WDM type, by using optical filters of a scale corresponding to the number of wavelengths, such as a micro-ring resonator (MRR), and utilizing wavelength parallelism and spatial parallelism, it is possible to perform convolution operations at a rate of wavelength parallelism (M), spatial parallelism (N), and baud rate (B) of OPS to MNB operations (see Non-Patent Literature 2). 【0005】 The above configurations are all significantly limited in terms of achievable computation speed and processing depending on the size of the implementable devices. For example, the SDM type requires ~N2 MZIs and the WDM type requires ~MN MRRs, but there are technical limitations on the size of the matrix that can be implemented from the standpoint of ease of manufacturing. For example, in the case of MZI, N=16 is reported (see Non-Patent Literature 3), and in the case of MRR arrays, M=N=4 is reported (see Non-Patent Literature 2). This indicates that, per unit device, only calculations of a scale of about 16x16 (or 4x4) can be performed in a single operation. Although it is possible to apply to larger matrix operations by parallelizing spatially small optical computers, like systolic arrays used in existing electronic computers, the typical size of the above unit operation devices (MZI, MRR) is several orders of magnitude larger than that of electronic devices, reaching 10-100 μm square, so there are technical limitations to spatial parallelization, and it is difficult to perform large-scale matrix multiplication calculations such as those used in current large-scale language models. Furthermore, the number of operations per unit time (OPS), a key performance indicator for computer hardware, is also constrained by the N and M that can be implemented on the chip. For example, in the above case, assuming a baud rate of around 30 GBAud for the input signal to the computer, the computation speed is limited to approximately 10 tera operations per second. 【0006】 U.S. Patent Publication No. 11,054,575 【0007】Shen, Y. et al. "Deep learning with coherent nanophotonic circuits." Nature photonics 11, 441-446 (2017).Feldmann, J. et al. "Parallel convolutional processing using an integrated photonic tensor core." Nature 589, 52-58 (2021).Kita, S., et al. "Demonstration of a Clements-type 16× 16 photonic analog matrix processor based on silicon photonics." Conference on Lasers and Electro-Optics / Pacific Rim., 2022.Texas Instruments, Integrator Circuit, DESIGN RESOURCES, [online], [Searched on November 29, 2024], Internet<URL:https: / / www.ti.com / lit / an / sboa275b / sboa275b.pdf?ts=1732838232329> 【0008】 This disclosure addresses the aforementioned challenges and provides a configuration for an optical computing device that can perform large-scale matrix calculations on an optical chip by applying the spatial, wavelength, and time parallelism of light. The configuration of the optical computing device facilitates the execution of large-dimensional matrix calculations on optical devices by utilizing time-direction multiplexing. Furthermore, this configuration is easily expandable by connecting optical fibers or optical waveguide structures, and boasts excellent scalability. 【0009】According to an embodiment for achieving this objective, an optical computing device is provided that can utilize the spatial, wavelength, and time parallelism of light, comprising: a light source that emits optical signals of multiple wavelengths; a first optical modulator array that modulates optical signals of multiple wavelengths according to an input matrix, wherein the input matrix of the first optical modulator array consists of rows in the wavelength domain and columns in the time domain; an optical branching means for spatially distributing the modulated optical signals of multiple wavelengths; a second optical modulator array that modulates the spatially parallelized optical signals of multiple wavelengths into a wavelength-collectively optical signal according to a weight matrix, wherein the weight matrix of the second optical modulator array consists of rows in the time domain and columns in the spatial domain; a wavelength branching means for separating the optical signals output from the second optical modulator array into their respective multiple wavelengths; an optical detector array that detects each of the separated multiple wavelength optical signals; and a time integration means for integrating the multiple wavelength optical signals detected in the time domain in the wavelength domain and the spatial domain to generate a predetermined matrix and matrix product signal. 【0010】 Furthermore, according to another embodiment, there is an optical computing device that can utilize the spatial, wavelength, and time parallelism of light, and further comprises: a memory means configured to store an input matrix and a weight matrix; a DAC array that converts the input matrix and the weight matrix from digital signals to analog signals; and an ADC array that converts optical signals of multiple wavelengths detected by an optical detector array from analog signals to digital signals. 【0011】 Furthermore, according to another embodiment, there is an optical computing device that can utilize the spatial, wavelength, and time parallelism of light, and further comprises a loss compensation mechanism configured to compensate for the loss of the matrix and matrix product signals output from the time integration means. 【0012】The first diagram shows a conventional calculation method, with (a) showing a spatial division multiplexing (SDM) type and (b) showing a wavelength division multiplexing (WDM) type calculation method. The second diagram shows the operating principle and overall configuration of an optical computing device in one operational configuration, with (a) being a schematic diagram of how signals in the wavelength, space, and time directions are calculated, and (b) being an overview diagram showing the components of the optical computing device. The third diagram shows a light source in an optical computing device in one operational configuration, with (a) and (b) showing the use of individual lasers, and (c) and (d) showing the use of a wavenumber comb light source. The fourth diagram shows a wavelength division multiplexing optical modulator array in an optical computing device in one operational configuration, with (a) showing a configuration with a wavelength multiplexer after each modulator, (b) showing the case using a direct modulation LD, (c) showing a configuration with a wavelength demultiplexer and a wavelength multiplexer before and after the modulator, (d) showing the case consisting only of modulators, (e) showing a configuration with a wavelength demultiplexer before each modulator, and (f) showing the case where a ring resonator is used as the modulator. This figure shows an optical splitting means in an optical computing device according to one embodiment, where (a) shows an example of using a 1:N optical splitter and (b) shows an example of using an M:N optical splitter. This figure shows a wavelength-batch optical modulator array in an optical computing device according to one embodiment. This figure shows a wavelength splitting means and an optical detector array in an optical computing device according to one embodiment, where (a) shows a case where a wavelength demultiplexer is used as the wavelength splitting means and a PD is used as the optical detector array, and (b) shows a case where a diffraction grating is used as the wavelength splitting means and a two-dimensional photodetector array is used as the optical detector array. This figure shows an embodiment of an optical computing device according to one embodiment. This figure shows a configuration in which scalability is improved by partially integrating an optical computing device according to one embodiment onto multiple chips and combining each chip. This figure compares the computing power and scale of an optical computing device between an optical computing device according to one embodiment and an optical computing device according to conventional technology. This figure shows estimated OPS and power efficiency values ​​for an optical computing device according to one embodiment. 【0013】(Overall Configuration and Operating Principle of the Optical Computing Device) One embodiment of the optical computing device in this disclosure will be described with reference to Figure 2. The configuration of the optical computing device in this embodiment is characterized by calculating an output matrix Y, which is the "matrix-matrix product" of an input matrix X (referred to as "input optical matrix X" when the input matrix X is described in the optical domain in this specification) and a weight matrix A (referred to as "weight optical matrix A" when the weight matrix A is described in the optical domain in this specification), using optical signals multiplexed in the wavelength, space, and time domains (see Figure 2(a)). The main conventional methods described in Figure 1 correspond to the "matrix-vector product" using one of the above degrees of freedom, so this point is fundamentally different. Furthermore, in the conventional methods, the matrix size that could be implemented was limited by the spatial arrangement of elemental devices such as MZI and MRR, but this method is also characterized by the use of time-division multiplexing as one of the degrees of freedom to represent the matrix. Since time-division multiplexing allows for arbitrary increases in matrix size regardless of the spatial arrangement of optical computing devices, this method has the excellent advantage of significantly increasing the size of representable matrices. 【0014】Specific configuration examples and operation of the optical computing device will be described with reference to Figures 2(a) and (b). The optical computing device 100 according to this embodiment comprises a light source 111 that emits at least two or more wavelengths (for example, M wavelengths), a wavelength division multiplexing optical modulator array 112 composed of M optical modulators, each modulating optical signals of M wavelengths, an optical branching means 113 that spatially distributes the modulated signals of each wavelength in the N direction, a wavelength collectively optical modulator array 114 composed of N optical modulators that modulate N spatially parallelized signals into a wavelength collectively signal, a wavelength branching means 115 that separates the modulated wavelength division multiplexed optical signals into M wavelengths, and an M × N optical detector array 116 that receives these wavelength-separated optical signals. Here, the total number of wavelengths M' emitted from the light source 111 is greater than M (M' = pM, p ≥ 1), and a portion of M may be extracted and used by a wavelength filter or the like. The system also includes a memory means 117 for storing matrices A and X to be operated on, digital-to-analog signal converter (DAC) arrays 118 and 119 for converting the digital signals stored on the memory means 117 into analog signals, and an analog-to-digital converter (ADC) array 120 for converting the output analog signals back into digital signals. 【0015】 It is desirable that the output power of the light source 111 be uniform for each wavelength, or that it be wavelength-filtered to be uniform. Furthermore, it is desirable that the wavelength-batch optical modulator array 114 be composed of modulator elements that are wavelength-independent (or have little wavelength dependence) in order to modulate the wavelengths collectively. In addition, a high-frequency amplifier may be appropriately provided in one of the high-frequency electronic wirings to adjust the input / output voltage ranges of each DAC array 118, 119 and ADC array 120, as well as the input / output voltage ranges of each wavelength-division multiplexer optical modulator array and optical detector array 116. Similarly, an optical amplifier may be provided in either the optical wiring or waveguide to compensate for optical losses in the optical computing device 100 and principle losses associated with optical branching. Examples of optical amplifiers include Er-doped fiber amplifiers (EDFAs) and semiconductor optical amplifiers (SOAs). 【0016】As will be described later, the output signal of the optical detector array 116 needs to be integrated for K time slots. Therefore, it is desirable to have this time integration means 121 in either the optical domain, the analog electronic domain, or the digital domain. Furthermore, the output signal Y obtained from the matrix operation will have variations due to the principle branching loss in the spatial branching of the optical circuit of the optical computing device 100, the insertion loss of the optical circuit and / or its wavelength dependence, etc. It is desirable to equalize these using loss compensation means 122 that compensates for these optically or electronically. In Figure 2(b), the time integration means 121 and the loss compensation means 122 are shown as an example of being inserted after ADC reception, but as will be described later, these mechanisms may be realized with analog electronic devices or optical devices (filters and amplifiers). Therefore, they may be inserted in the preceding stage before ADC reception. Possible specific embodiments of these will be described later, and the claims of this application are not limited to these order or forms of implementation. 【0017】 The following describes the specific signal processing operation of the optical computing device 100 according to this embodiment, with further reference to Figure 2(b). Initially, the optical signals with multiple wavelengths output from the light source 111 are introduced into the wavelength division multiplexing optical modulator array 112. The M x K input matrix X stored in the memory means 117 is introduced into the wavelength division multiplexing optical modulator array 112 via the input signal DAC array 118. Here, the K-dimensional signals in the row direction are represented in each time slot using time division multiplexing, and the M-dimensional signals in the column direction are represented by M modulators corresponding to different wavelengths, thereby generating the input optical matrix X in the wavelength-time domain. Here, the input optical matrix X may be represented as a real number using light intensity modulation, or it may be represented in the complex domain using an IQ modulator. In the former case, it has the advantage of being compatible with real number calculations used in general machine learning, and it is not sensitive to phase changes of light within the device, thus improving the ease of device manufacturing. In the latter case, superior features emerge, such as the ability to extend the calculation target to complex matrices and map signals to the IQ space, thereby suppressing the required signal-to-noise ratio. 【0018】The optical signal modulated by the wavelength-time domain input matrix X is distributed in the spatial direction by an optical branching means 113 with a branching number N. This means that the optical signal modulated by the input matrix X is spatially multiplied by 1 / αi and copied. Here, αi (where i = 1, 2, ..., N) is the branching ratio to each path. The branching ratio αi can be arbitrary as it can be corrected in subsequent processing, but it is desirable to distribute with equal branching ratios (αi = N) to reduce the processing load. The case where αi = N is described below as an example, but this does not limit the scope of the claims of this application. For the sake of understanding, a schematic diagram of the wavelength-space-time signal of the signal generated in the optical domain at this point is shown in "I" of Figure 2(a). This corresponds to the optical signal modulated by the wavelength-time domain input matrix X having its intensity multiplied by 1 / N in the spatial direction and being copied. 【0019】 The branched optical signals are each inserted into the wavelength-batch optical modulator array 114. A K x N column weight matrix A, stored in the memory means 117, is introduced to the optical modulators constituting the wavelength-batch optical modulator array 114 via a separate weight signal DAC 119. Here, the N-dimensional signals in the row direction are represented by N modulators corresponding to different spatial channels, and the K-dimensional signals in the column direction are represented in each time slot using time-division multiplexing, generating the weight optical matrix A in the spatial and temporal domains. Here, the optical signals introduced into the wavelength-batch optical modulator array 114 are signals modulated in wavelength and temporal space, but since all wavelength signals are modulated with the same weight, it corresponds to performing calculation processing with similar weights in the wavelength direction. Schematic diagrams of the wavelength, spatial, and temporal signal configuration of the weight matrix A generated by the wavelength-batch optical modulator array 114, and schematic diagrams of the output signals obtained as a result of calculations with the signals introduced into the optical modulators are shown in "II" and "III" of Figure 2(a). 【0020】 【0021】 【0022】Furthermore, the received signal undergoes additional attenuation β due to insertion loss in the optical circuit and loss dependence on wavelength and space. β is a coefficient that depends on the wavelength and spatial channel. These can also be compensated for by amplifying the signal by 1 / β using the loss compensation mechanism 122 described above. Moreover, by using a coherent detector in the optical detector array 116, it is possible to acquire the output signal Y in the complex number domain. 【0023】 The following describes specific embodiments of each component that constitutes the optical computing device 100 in this disclosure. 【0024】 (Light Source) Hereinafter, embodiments of the light source 111 will be described with reference to Figure 3. As a first embodiment of the light source 111, there may be an array of laser light sources as shown in Figure 3(a). The light source 111 comprises a plurality of lasers (e.g., LDs) 311 to 31M and a plurality of waveguides (or optical fibers) 321 to 32M connected to each laser. In this embodiment, any number of wavelength parallelisms M can be realized by increasing the number of lasers according to the number of wavelengths required for the calculation (e.g., M wavelengths). Furthermore, when using a wavelength demultiplexer (WDM-DEMUX) type, which will be described later, as a means of optical modulation, a DEMUX element is not required because the downstream stage of the light source is already in a wavelength-separated state, and excellent features such as reduced principle loss and footprint can be obtained. 【0025】 The laser light source arrays shown in Figures 3(a) and 3(b) have superior advantages compared to optical comb light sources described later, such as generally higher stability of optical power and center frequency, and lower calculation errors caused by fluctuations in these frequencies. In addition, a light source array with multiple lasers has the advantage of being able to set the output power from each laser 311-31M sufficiently higher than the noise level, thereby reducing the number of optical loss compensation mechanisms 122 described later. This not only reduces the number of components but also reduces the impact of noise added due to signal amplification, thus contributing to the reduction of calculation errors. Furthermore, by adopting a direct laser modulation method in the modulation method described later (shown in Figure 4(b)), it is possible to realize the superior advantage of integrating the light source 111 and the wavelength division multiplexing optical modulator array 112 as a single unit. 【0026】 As a modified embodiment, the output signal from the LD may be combined into a single output port using a wavelength multiplexer (WDM-MUX) as shown in Figure 3(b). In this case, excellent compatibility with the ring modulator described later is achieved, and fiber connection is simplified when the light source and modulator chips are separated. The WDM-MUX and WDM-DEMUX can be constructed using any optical filter, such as an array-type waveguide grating (AWG), an MRR array, or a delay interferometer array. The total number of wavelengths M' emitted from the light source 111 is greater than M (M' = pM, p≧1), and a portion of M may be extracted and used using a wavelength filter or the like. 【0027】 A second embodiment of the light source 111 is a frequency comb light source 350 as shown in Figure 3(c). In this case, input light having M wavelengths can be generated simultaneously, which has the excellent feature of enabling simplification and integration of the device. Existing technologies such as optical microring resonators and fiber ring resonators can be used to generate optical frequency combs. The spacing of the oscillating frequency combs, output intensity, and device size differ depending on the method. In the embodiments of this disclosure, any technology can be selected as long as it does not violate the wavelength spacing constraints described later. From the viewpoint of integrating computing devices, it is desirable to reduce the device footprint. From the viewpoint of the above constraints, it is desirable to use a so-called microring optical resonator having a ring diameter of about 10-1000 μm as a comb light source. 【0028】 Furthermore, the output signal from the frequency comb light source 350 may be split into different output ports 342 to 34M for each wavelength using a wavelength demultiplexer 330 (WDM-DEMUX) as shown in Figure 3(d). In this case, it has the advantage of excellent connectivity with MZ-type and field absorption type modulators, which will be described later. Moreover, the total number of wavelengths M' oscillated from the frequency comb light source is greater than M (M' = pM, p≧1), and a portion of M may be extracted and used with a wavelength filter or the like. 【0029】The following explains the constraints on the frequency (wavelength) spacing of light sources. As mentioned above, since the signals of each wavelength are used as their respective matrix elements, independence (no wavelength crosstalk) is required. An optical signal modulated at a baud rate B has an effective bandwidth B', and therefore the modulated optical signal has a spread of B' in the frequency domain (the reciprocal of the wavelength domain). However, B' depends on the generated impulse response, and B' ≥ B. The value of B is typically on the order of 10–100 GHz, which is the bandwidth of fast optical modulation. The equality sign indicates the Nyquist limit, meaning that a perfectly rectangular waveform can be formed in the frequency domain. This means that if the adjacent optical frequency spacing W is less than B' (≥ B), crosstalk will occur between the signals spread in the frequency domain. Therefore, W > B' (≥ B) gives the lower limit of the frequency (wavelength) spacing of light sources. The upper limit of the wavelength spacing is constrained by the ratio of the wavelength bandwidth F (e.g., C-band, 1530-1570 nm, approximately 5 THz) supported by the device being used and the required number of wavelengths M. This is because, when M wavelengths of light are input with a wavelength spacing W, if the constraint WM ≤ F is exceeded, it becomes impossible to arrange each wavelength signal within the supported bandwidth F. Therefore, it is desirable to select the wavelength spacing W within the constraint B ≤ B' < W ≤ (F / M). 【0030】 (First Optical Modulator Array) Hereinafter, an embodiment of the wavelength division multiplexing optical modulator array 112 (first optical modulator array) will be described with reference to Figure 4. As one embodiment of the wavelength division multiplexing optical modulator array 112, waveguide ports 411 to 414 and modulators 421 to 424 are independently arrayed for each wavelength of an optical signal having an M wavelength, and the wavelength signals are bundled by a subsequent wavelength multiplexer (WDM-MUX) 430. This method is called the individual wavelength input type. The wavelength division multiplexing optical modulator array 112 can use, for example, a Mach-Zehnder modulator (MZM) array, an electric field absorption (EA) type modulator, or an MRR modulator array. Alternatively, the input signal may be represented in the complex domain by using an IQ modulator realized by connecting multiple MZMs in stages. 【0031】The above method assumes that optical signals of each wavelength are introduced from different input ports; therefore, it is desirable to use the light source configuration shown in Figures 3(a) and 3(d) as the light source 111. When integrating with the light source 111 configuration in Figure 3(a), a direct modulation method using direct modulation LDs 441 to 444, which can modulate the laser output intensity with the input signal, may be used, as shown in Figure 4(b). In this case, an excellent effect is achieved, such as the ability to integrate the light source 111 and the wavelength division multiplexer optical modulator array 112. Furthermore, as shown in Figure 4(c), by adding a WDM-DEMUX 450 further upstream, it is possible to accommodate configurations where wavelength division multiplexed signals are supplied from a single input port, as shown in Figures 3(b) and (c). Moreover, as shown in Figure 4(d) or (e), a configuration without using the WDM-MUX 430 downstream of the wavelength division multiplexer optical modulator array 112 is also possible. In this case, a 1:N splitter cannot be used in the optical branching means 113 described later, but connection to the next stage is possible by using an M:N optical splitter. This configuration has excellent effects such as suppressing the increase in device size and optical loss associated with WDM-MUX. Furthermore, the wavelength division multiplexed optical modulator array 112 can also be realized by configuring multiple wavelength-selective optical modulators 461 to 46M (e.g., MRR modulators) in the direction of optical signal propagation, as shown in Figure 4(f). This configuration has the excellent advantage that optical modulation is possible without introducing WDM-DEMUX when the wavelength division multiplexed signal is supplied from a single input port. 【0032】 (Optical Splitting Means) Hereinafter, an embodiment of the optical splitting means 113 will be described with reference to Figure 5. When the optical signals from the wavelength division multiplexer optical modulator array 112 are bundled and input into a single fiber 521, the optical splitting means 113 can be configured using a 1:N optical splitter 510 as shown in Figure 5(a). This configuration can be, for example, a method of configuring multiple stages of a 1x2 multimode interferometer (MMI) or 2x2 directional coupler integrated using optical waveguide technology, or a 1:N multimode interferometer (MMI). 【0033】On the other hand, if the optical signals from the wavelength division multiplexer array 112 are input from different fibers 521 to 52M for each wavelength, it is possible to use an M:N optical splitter 520 structure as shown in Figure 5(b). This configuration can be achieved, for example, by connecting multiple M:N star couplers, 2x2 directional couplers, or MMIs in stages. Since the subsequent wavelength-batch optical modulator array 114 modulates the signal collectively across all wavelengths, it is desirable to configure the system so that wavelength dependence does not occur in any case, and it is desirable to design it so that the optical path difference between each spatial channel, which is the main cause, does not occur. 【0034】 (Second Optical Modulator Array) An embodiment of the wavelength-coordinated optical modulator array 114 (second optical modulator array) will be described below with reference to Figure 6. This configuration can be realized, for example, by arranging a plurality of optical modulators 621 to 624 in parallel, as shown in Figure 6. The wavelength-coordinated optical modulator array 114 can use, for example, a Mach-Zehnder modulator (MZM) array, an electric field absorption (EA) type modulator, etc. Since the signal is modulated in a wavelength-coordinated manner, it is desirable that the array be composed of optical modulators that are independent of wavelength, and for example, it can be composed of an MZM with no (or small) optical path difference between arms. 【0035】 Furthermore, it is desirable that the materials constituting each optical modulator 621 to 624 have low wavelength dependence. For example, when configured in the C-band for optical communication, it is desirable that the optical modulators be composed of Si, SiN, LiNbO3, etc. For the integration of different materials, techniques such as thin-film growth on a main wafer made of Si, wafer-level bonding, or local transfer of dissimilar materials by transfer printing can be used. In addition, the wavelength-coordinated optical modulator array 114 may be made with different optical chips and connected by optical fibers or waveguides. The modulated signal may also be represented in the complex domain by using an IQ modulator realized by connecting multiple MZMs in stages, etc. 【0036】(Wavelength Splitting Means and Optical Detector Array) Hereinafter, an embodiment of the wavelength splitting means 115 and the optical detector array 116 will be described with reference to Figure 7. In this configuration, for example, as shown in Figure 7(a), the wavelength splitting means 115 is composed of WDM-DEMUX 731 and 732 integrated using optical waveguide technology, and the optical detector array 116 may be composed of a plurality of photodiodes (PDs). The WDM-DEMUX 731 and 732 are realized by optical filter structures such as AWG and MRR. The PDs constituting the optical detector array are photodiodes using semiconductors such as Ge. Although not shown, it is desirable to provide analog electronic circuits such as a transimpedance amplifier (TIA) for amplifying the received signal after the optical detector array 116. 【0037】 Furthermore, as in another embodiment, as shown in Figure 7(b), the optical signal output from the wavelength-batch optical modulator array 114 may be emitted into space, a diffraction grating 710 may be used as the wavelength splitting means 115, and a two-dimensional photodetector array 720 may be used as the optical detector array 116. In this case, the optical detectors can be arranged two-dimensionally, resulting in excellent scalability, and other advantages such as the ability to use commercially available technologies such as infrared cameras and charge-coupled device (CCD) arrays as optical detector arrays. As a specific configuration, for example, the configuration in Patent Document 1 can be used. 【0038】(Time Integration Means) One embodiment of the time integration means 121 will be described. The time integration means 121 can be realized, for example, by configuring an integration circuit made of analog electronic circuits after the optical detector array 116. For example, it can be configured using a semiconductor operational amplifier and a feedback capacitor, as described in Non-Patent Literature 4. This configuration has the advantage of low error in matrix operations because linear integration is possible. In addition, a low-pass filter can be used as an integrator. The low-pass filter may be implemented on an analog electronic circuit or on an optical circuit. On an analog electronic circuit, for example, an RC integration type first-order low-pass filter circuit formed by a resistor and a capacitor can be used. On an optical circuit, for example, it can be configured as an impulse response filter composed of a delay line interference system. Furthermore, the photodetector and low-pass filter functions may be integrated and configured as a narrow-bandwidth optical detector. The optical computing device 100 in this disclosure has the advantage of being highly producible and economical because, although the integral values ​​are not necessarily added linearly, errors are likely to occur in the results of matrix operations, the required receiver bandwidth and ADC bandwidth can be narrowed. Another implementation is a method in which the signal waveforms after ADC reception are summed using digital processing. This method incurs the cost of an additional O(NK) digital processing steps for summation, but has the advantage of not requiring analog elements for summation. 【0039】(Loss Compensation Means) An embodiment of the loss compensation means 122 will be described. The loss compensation means 122 is provided as a means to compensate for losses due to principle losses associated with branching in the optical circuit of the optical computing device, as well as losses due to absorption and scattering of materials. The loss compensation means 122 may be mounted on the optical circuit, analog electronic circuit, or digital circuit. On the optical circuit, it can be realized by inserting, for example, an Er-doped optical fiber amplifier (EDFA) or a semiconductor optical amplifier (SOA) into the transmission line. The former is excellent for insertion into the fiber transmission line, and the latter is excellent for integration on the optical waveguide circuit. Amplification in the optical domain has the advantage of requiring fewer devices because it can amplify multiple wavelength signals at once. On the analog electronic circuit, it can be realized by using a broadband amplifier such as a transimpedance amplifier (TIA). On the digital circuit, it can be realized by processing that multiplies the obtained output signal by a constant. The loss compensation means 122 has drawbacks such as increased digital processing load and inability to recover a degraded signal-to-noise ratio, but it has the excellent feature of not requiring additional compensation devices. Signal compensation may be performed by combining the above-mentioned loss compensation methods. 【0040】 (Example of Overall Configuration) The optical computing device 100 has many possible implementations depending on the combination of the above-described components. As an example, a specific example of the overall configuration is shown in Figure 8. The optical computing device 100 uses a configuration in which a frequency comb light source is used as the light source 111 and an MRR type optical modulator is used in the wavelength division multiplexing optical modulator array 112. This configuration exhibits the excellent effect of integrating the optical computing device onto a single element. As an example of the configuration, a form using a frequency comb light source and an MRR type optical modulator is shown, but similar effects can be obtained by using other component or device embodiments as described above. For example, a configuration using a direct modulation laser array and WDM-MUX, which integrates the light source 111 and the wavelength division multiplexing optical modulator array 112, can also exhibit similar miniaturization and integration effects. The input matrix X, which is connected to the input signal DAC 118 and stored in the memory means 117, is introduced into the wavelength division multiplexing optical modulator array 112. 【0041】The optical signals modulated by the wavelength division multiplexed optical modulator array 112 are introduced into an optical splitter used as the optical branching means 113. By the optical branching means 113, after the optical signals are spatially separated into a plurality of signals, they are introduced into the wavelength block optical modulator array 114. The wavelength block optical modulator array 114 can be realized by arranging MZM optical modulators in parallel in each waveguide. Also, a weight signal DAC 119 is connected to the wavelength block optical modulator array 114, and the weight matrix A stored in the memory means 117 is introduced. 【0042】 Wavelength branching means 115 composed of an AWG are respectively arranged at the subsequent stage of the wavelength block optical modulator array 114 to demultiplex the optical signals by wavelength. Each of the demultiplexed optical signals is inserted into a photodetector array 116 composed of PDs and detected. Also, the detected signals are converted from analog to digital signals by an output signal ADC 120. Thereafter, the output digital signals can be integrated and loss-compensated by the time integration means 121 and the loss compensation means 122. Note that the integration result can be stored in the memory means 117 as necessary. 【0043】(Example of a combination of multiple chips) As described above, when all components of the optical computing device 100 are integrated onto a single chip, the degree of integration is limited by issues such as chip size, manufacturing uniformity, or yield. The optical computing device 100 of this disclosure can also be configured by optical wiring between multiple chips, for example, as shown in Figure 9. For example, the wavelength division multiplexer optical modulator array 112 that modulates the optical signal from the light source 111 can be integrated onto one chip (for example, a first optical integrated circuit 911 to 913), and the components from the optical branching means 113 to the optical detector array 116 can be integrated onto another chip (for example, a second optical integrated circuit 921 to 923), and then the chips can be connected using wavelength multiplexing means (for example, WDM-MUX) 930 or another optical branching means (for example, an optical splitter) 940. Furthermore, there may be overlap in the wavelength range generated between the input chips and the wavelength range received between the output chips. Therefore, by combining multiple small chips, it is possible to achieve the excellent function of improving scalability. As will be described later, in this configuration, as the size of the optical chip increases, it is possible to reduce the OPS (operations per unit time), which is an indicator of the number of calculations per unit time, and the power consumption per unit calculation. Thus, this configuration can scalably improve the OPS and power efficiency of the optical computing circuit. 【0044】(Effects of the Optical Computing Device in This Disclosure) The optical computing device in this disclosure has significant advantages over the prior art in terms of computing speed, power consumption, and the scale of matrix operations that can be implemented. For example, in Non-Patent Document 1 or 2, the prior art requires the integration of a number of active elements corresponding to the number of elements in matrix A, which limits the size of the matrix that can be calculated at one time to a matrix-vector product of about N × N (or M × N). Here, M and N are the number of wavelength channels and the number of spatial channels, respectively, and N = 16 has been reported. Therefore, in order to perform the large-scale matrix multiplication used in current AI models (for example, the Transformer model requires a large-scale matrix multiplication of about 10⁴ × 10⁴), it is necessary to decompose it into N × N submatrices and calculate the output by parallelizing it a large number of times (or by updating the optical circuit many times). Figure 10 shows the relationship between the number of parallelisms (number of updates) and the size of the matrix. As can be seen from Figure 10, for N=4, approximately 10 million parallel iterations (updates) are required, and even if N is increased to 64, more than 10,000 parallel iterations (updates) are still necessary for large-scale matrix operations, making the technical barrier extremely high. 【0045】 On the other hand, the optical computing device in this disclosure is capable of performing matrix multiplication between a matrix X with dimension M × K and a matrix A with size N × K, where K is the time-division multiplexing number. Here, the values ​​of M and N on a single optical circuit are similar to those of the optical device described above (e.g., 16), but the value of K can be set arbitrarily. Therefore, even a large matrix of about 104 can be decomposed into N × K submatrices, and by setting K = 104, the number of parallel processes (updates) can be significantly reduced. As shown in Figure 10, even in the case of M = N = 8, the number of parallel processes (updates) can be reduced by more than an order of magnitude. Furthermore, the proposed configuration can achieve integration beyond the number of M and N that can be realized on a single chip by connecting multiple chips. For example, by connecting 16 optical chips with M = 16 and N = 16 in parallel, it is possible to configure a computing device of a scale such as M = N = 256. This configuration can be realized with approximately 100 parallel processes (updates), and is scalable even for large-scale matrix operations. 【0046】Furthermore, the optical computing device in the present disclosure can perform calculations of up to 2(M×N×K) times per clock in a time TK, which is the product of the time clock T = 1 / B, the reciprocal of the baud rate B, and the time division multiplexing number K. Therefore, the number of operations per unit time, OPS, is estimated to be up to 2(M×N×B). In this configuration, since the numbers of M and N can be scalable improved through chip connection, OPS can be significantly improved compared to the prior art. 【0047】 On the other hand, the number of times of use of the DAC and ADC that limit the power consumption is only up to M×K times per unit time (DAC for the input matrix X), up to N×K times (DAC for the weight matrix A), and M×N times (ADC for the output signal). Therefore, the power efficiency improves by N times, M times, and K times respectively as the scale increases. Therefore, like OPS, there is a scaling merit in terms of power efficiency. 【0048】 Fig. 11 shows the estimated values of OPS and power efficiency. Assuming B = 30 Gbaud and K = 10000, the values of M and N are changed according to possible implementation forms. As can be seen from Fig. 11, it can be understood that the calculation speed and operation efficiency can be improved up to the regions that were impossible with conventional electronic circuits and optical computing circuits, due to the configuration of the optical computing device in the present disclosure. 【0049】 The optical computing device in the present disclosure provides a configuration that can perform large-scale matrix calculations even with a small chip by utilizing the spatial, wavelength, and time parallelism of light. 【0050】100 Optical computing device 111 Light source 112 Wavelength division multiplexer array (first optical modulator array) 113 Optical branching means 114 Wavelength batch optical modulator array (second optical modulator array) 115 Wavelength branching means 116 Optical detector array 117 Memory means 118 Input signal DAC array 119 Weight signal DAC array 120 Output signal ADC array 121 Time integration means 122 Loss compensation mechanism 311-31M Laser (LD) 321-32M Waveguide (or optical fiber) 360, 450, 731, 732 Wavelength demultiplexer (WDM-DEMUX) 341-34M Waveguide (or optical fiber) 350 Frequency comb light source 411-414 Waveguide (or optical fiber) 421-424 Optical modulators 430, 930 Wavelength multiplexers (WDM-MUX) 441-444 Direct modulation LDs 461-46M MRR modulators 510, 520, 940 Optical splitters 521-52M Waveguides (or optical fibers) 611-614 Waveguides (or optical fibers) 621-624 Optical modulators 710 Diffraction gratings 720 Two-dimensional photodetector arrays 911-913 First optical integrated circuit 921-923 Second optical integrated circuit

Claims

1. An optical computing device comprising: a light source that emits optical signals of multiple wavelengths; a first optical modulator array that modulates the optical signals of multiple wavelengths according to an input matrix, wherein the input matrix consists of rows in the wavelength domain and columns in the time domain; an optical branching means for spatially distributing the modulated optical signals of multiple wavelengths; a second optical modulator array that modulates the spatially parallelized optical signals of multiple wavelengths into a wavelength-collectively optical signal according to a weight matrix, wherein the weight matrix consists of rows in the time domain and columns in the spatial domain; a wavelength branching means for separating the optical signals output from the second optical modulator array into their respective wavelengths; an optical detector array for detecting each of the separated optical signals of the multiple wavelengths; and a time integration means for integrating the optical signals of multiple wavelengths detected in the time domain for each of the wavelength domain and the spatial domain to generate a predetermined matrix and matrix product signal.

2. The optical computing device according to claim 1, further comprising: a memory means configured to store the input matrix and the weight matrix; a DAC array that converts the input matrix and the weight matrix from digital signals to analog signals; and an ADC array that converts the plurality of wavelength optical signals detected by the optical detector array from analog signals to digital signals.

3. The optical computing apparatus according to claim 1, further comprising a loss compensation mechanism configured to compensate for the loss of the matrix and matrix product signals output from the time integration means.

4. The optical computing device according to claim 1, wherein the light source is a frequency comb light source, and a wavelength demultiplexer is provided downstream of the light source.

5. The optical computing device according to claim 1, wherein a direct modulation laser array is used as the light source and the first optical modulator array, which is a combination of the light source and the first optical modulator array.

6. The optical computing apparatus according to claim 1, wherein the first optical modulator array is one of a Mach-Zehnder modulator array, an electric field absorption modulator, or a microring resonator array.

7. The optical computing apparatus according to claim 1, wherein the second optical modulator array is one of a Mach-Zehnder modulator array and an electric field absorption modulator.

8. The optical computing device according to claim 1, wherein the wavelength splitting means is a diffraction grating and the optical detector array is a two-dimensional photodetector array.

9. The optical computing device according to claim 1, wherein the time integration means is one of an RC integral type first-order low-pass filter circuit or an impulse response filter.

10. The optical computing device according to claim 1, comprising one or more first optical integrated circuits integrating the light source and the first optical modulator array, and a connection of the optical branching means, the second optical modulator array, the wavelength branching means, and one or more second optical integrated circuits integrating the optical detector array.

11. The optical computing device according to claim 10, wherein when at least one of the one or more first optical integrated circuits and the one or more second optical integrated circuits is parallelized with two or more optical integrated circuits, each of the optical integrated circuits is coupled using a wavelength multiplexer and an optical branching means.

Images