Optical matrix operation device
The optical matrix calculation device addresses scalability and speed limitations by using a wavelength multiplexed and spatial light modulator system for parallel operations, achieving petascale calculations and overcoming conventional limitations in optical processing.
Patent Information
- Application Number
- PCT/JP2023/020550
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2023-06-01
- Publication Date
- 2024-12-05
AI Technical Summary
Current optical matrix calculation devices face limitations in scalability and speed due to manufacturing technology and material characteristics, restricting large-scale matrix multiplication operations and failing to leverage the broadband characteristics of light for high-speed calculations.
The proposed optical matrix calculation device employs a wavelength multiplexed optical signal generation and spatial light modulator system, enabling parallel matrix operations through a spatial optical system with a high-density light beam focus, and an optical waveguide front end to increase beam density, allowing for efficient large-scale optical matrix calculations.
This configuration supports high-speed optical modulation signals by achieving petascale calculations per unit time, surpassing conventional technologies in both spatial channel count and filter performance, thereby enabling efficient large-scale optical processing.
Smart Images

Figure JP2023020550_05122024_PF_FP_ABST
Abstract
Description
Optical matrix calculation device
[0001] The present disclosure relates to an optical matrix calculation device.
[0002] Information processing using machine learning with deep neural networks (DNNs) is gaining attention. DNNs are large-scale nonlinear networks in which numerous neurons with nonlinear responses are connected by synapses, and deep learning techniques using networks with multi-layered neurons in particular have been widely applied. This technology has been reported to demonstrate excellent performance in a wide range of fields, including image and speech recognition, robot control, and artificial data generation.
[0003] On the other hand, as the number of layers and neurons (nodes) increases, the scale of the required nonlinear matrix operations explodes, posing a major challenge: the increase in the time and power required for inference and learning. In recent years, attention has been focused on information processing technology that uses optical circuits as computing elements as a method to solve these issues.
[0004] In general machine learning based on DNN, the signal x in the Lth layer (L) The output is obtained by applying the transformation of the following equation (1) to
[0005] where f is a nonlinear function such as tanh or relu, and x (L) , x (L+1) are M-dimensional and N-dimensional input / output vectors, respectively, and ω (L) is a transformation matrix with N rows and M columns called a weight matrix. In machine learning, the weight matrix ω (L) By learning and updating, the desired final output is obtained.
[0006] In a DNN based on optical computation using optical circuits, part or all of this nonlinear matrix multiplication is performed optically. A known example of conventional optical computation devices is the optical computation device described in Non-Patent Document 1. The optical computation device described in Non-Patent Document 1 uses a multi-array optical interference system as an optical interference system array, thereby converting the intensity of the input signal x according to the phase value of the interference system. The output optical signal y in this case can be expressed as in Equation (2).
[0007] Equation (2) is the ω of the calculation equation (1) in the DNN mentioned above. (L) x (L) Regarding nonlinear conversion, in the case of Non-Patent Document 1, the received light is converted into an electrical signal and then digitally processed. However, by placing some kind of nonlinear optical element in the subsequent stage of the interference system, it is possible to completely execute equation (1) including nonlinear calculations. In equation (1), the matrix product ω (L) x (L) Since the number of operations is limited by the nonlinear operation, it is also useful to perform the nonlinear operation after the output optical signal is photoelectrically converted. In this embodiment, the weight learning in equation (1) is realized by learning and updating the set phase value of the interference system. The phase value of the interference system can be dynamically changed by changing the refractive index of the waveguide, for example, by the thermo-optic effect.
[0008] The optical arithmetic device in Non-Patent Document 1 is constructed on a plane using optical waveguides, but due to the manufacturing technology and the characteristics of optical materials, loss is large and it is difficult to realize a large-scale arithmetic element, and at present, the circuit scale is limited to the implementation of matrix multiplication operations of about 16 × 16.
[0009] A configuration using a three-dimensional spatial optical system has been proposed for the purpose of implementing larger-scale matrix multiplication calculations (Non-Patent Document 2). In a configuration such as Non-Patent Document 2, a light beam is emitted into space, and large-scale matrix multiplication calculations can be realized using a diffraction grating (DOE), a metasurface, a spatial light modulator (SLM), or the like. In such a configuration, the input signal is spatially modulated using a display element, resulting in a low modulation frequency of 100 Hz or less. This makes it difficult to perform calculations that utilize the broadband characteristics of light exceeding the terahertz band, and currently, the number of calculations per unit time is limited to approximately teracycles per second.
[0010] Y. Shen et al., “Deep learning with coherent nanophotonic circuits”, Nature photonics, Vol.11, P441-447, June 2017. <URL:https: / / www.nature.com / articles / nphoton.2017.93?TB_iframe=true&width=921.6&height=921.6>X. Lin et al., “All-optical machine learning using diffractive deep neural networks”, Science, Vol.361, Issue 6406, P1004-1008, June 26, 2018. <URL: https: / / www.science.org / doi / 10.1126 / science.aat8084>J. Feldmann et al., “Parallel convolutional processing using an integrated photonic tensor core”, Nature, VOL.589, p.52-58, January 6, 2021. <URL: https: / / www.nature.com / articles / s41586-020-03070-1>M. Nakajima et al., “Optical Arbitrary Waveform Processing of Over 100 Spatial Channels for Optical Performance Monitoring”, JOURNAL OF LIGHTWAVE TECHNOLOGY, January 15, 2019, VOL. 37, NO. 2, p.291-299. <URL:https: / / ieeexplore.ieee.org / abstract / document / <8465965>M Shoeybi et al., “Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism”, arxiv, 2019. <URL: https: / / arxiv.org / pdf / 1909.08053.pdf>M. Nakajima et al., “Multilane Photonic Spectral Processor Integrated in a Spatial and Planar Optical Circuit for a Space-Division Multiplexing Network”, JOURNAL OF LIGHTWAVE TECHNOLOGY, January 15, 2018, VOL. 36, NO. 2, p.309-317.<URL:https: / / ieeexplore.ieee.org / abstract / document / <8031404>.
[0011] The present disclosure aims to provide a large-scale optical matrix operation device that supports high-speed optical modulation signals used in optical communication and the like.
[0012] An optical matrix operation device according to an embodiment of the present disclosure includes a wavelength-division multiplexed optical signal generation unit that generates a vector or a tensor by a wavelength-division multiplexed optical signal, an optical matrix operation unit, a multi-channel optical reception unit, and an optical transmission means that transmits the wavelength-division multiplexed optical signal output from the wavelength-division multiplexed optical signal generation unit to the optical matrix operation unit. The optical matrix operation unit includes a spatial light modulator that can independently modulate the light condensed on the spatial light modulator for each predetermined region by driving a plurality of pixels arranged in a matrix on a plane by a driving unit, M (M is a natural number of 2 or more) input optical paths, M output optical paths, an input port connected to the optical transmission means, and an output port connected to each of the M input optical paths. An optical splitter, and an optical element group arranged to disperse the wavelength-division multiplexed optical signal of the input optical path for each wavelength and condense it at different positions on the spatial light modulator, and combine each signal light emitted from the different positions of the spatial light modulator into the M output optical paths. The multi-channel optical reception unit is characterized by including an optical receiver connected to each output optical path.
[0013] FIG. 1 is a diagram for explaining the principle of an optical matrix calculation device of the present disclosure. FIG. 2 is a diagram for explaining the principle of an optical matrix calculation device of the present disclosure. FIG. 3 is a diagram schematically showing an optical matrix calculation device of a first embodiment of the present disclosure. FIG. 4 is a diagram for explaining wavelength filter performance α of the optical matrix calculation device. FIG. 5 is a diagram schematically showing an optical matrix calculation device of a second embodiment of the present disclosure. FIG. 6 is a diagram schematically showing an optical waveguide front end of the second embodiment of the present disclosure. FIG. 7 is a diagram for explaining the advantage of the second embodiment in terms of calculation speed OPS. FIG. 8 is a diagram schematically showing an optical matrix calculation device of a third embodiment of the present disclosure.
[0014] Hereinafter, embodiments of the present disclosure will be described in detail with reference to the drawings. Note that the following description is an example, and some configurations may be modified without departing from the gist of the present disclosure.
[0015] The present disclosure provides a large-scale optical processing device that supports high-speed optical modulation signals currently used in optical communications, etc. The device according to an exemplary embodiment of the present disclosure includes a wavelength-multiplexed optical signal generation unit that generates an input vector or an input tensor using a wavelength-division multiplexed (WDM) optical signal, an optical transmission means that transmits the wavelength-multiplexed optical signal to an optical matrix processing unit, an optical matrix processing unit that includes a spatial light modulator (SLM), an optical branching unit that branches the input wavelength-multiplexed optical signal light, and an optical element group that includes a lens group and a diffraction grating that form images of each branched wavelength-multiplexed optical signal at different positions on the spatial wavelength modulator for each wavelength, and a multi-channel optical receiving unit that receives each branched wavelength-multiplexed optical signal collectively and performs summation of each wavelength signal in parallel.
[0016] In the device of this embodiment, the input wavelength-multiplexed optical signal is modulated individually in each region on the spatial light modulator. This allows the device of this embodiment to perform matrix operations in parallel in the optical region. In this way, this embodiment uses a wavelength-multiplexed optical signal as an input signal and performs matrix operations in a large-scale parallel manner in the spatial optical system, thereby achieving calculations that exceed those of conventional examples.
[0017] The computational efficiency of the optical matrix operation device of this embodiment is determined by the density of the light beams imaged on the spatial light modulator. Therefore, this disclosure also proposes an embodiment that makes it possible to increase the density of the light beams imaged on the spatial light modulator by using an optical waveguide front end.
[0018] Furthermore, in the optical matrix calculation device disclosed herein, the wavelength-multiplexed optical signal generation unit and the optical matrix calculation unit that performs the matrix calculation can be connected by a single optical fiber. Therefore, the present disclosure also proposes an embodiment in which the wavelength-multiplexed optical signal generation unit that generates the input signal and the optical matrix calculation unit are arranged separately. According to this embodiment, the input signal transmitted from the wavelength-multiplexed optical signal generation unit can be directly calculated by the optical matrix calculation circuit unit, thereby achieving the excellent effect of enabling calculation to be performed in the optical domain while the input signal is being communicated.
[0019] (Principle of the Optical Matrix Operation Device of the Present Disclosure) First, the basic principle of the optical matrix operation device of the present disclosure and the configuration for realizing it will be described using Fig. 1. Fig. 1 shows a configuration for realizing the calculation of the above-mentioned equation (2) in the optical domain. The optical matrix operation device 100 of Fig. 1 is composed of a wavelength-multiplexed optical signal generation unit 110, an optical matrix operation unit 120, and a multi-channel optical receiving unit 130.
[0020] The wavelength multiplexed optical signal generating unit 110 includes N laser arrays 111 to 111 that oscillate at different wavelengths. N and an input signal X input by a digital circuit 112 via a digital-to-analog converter (DAC) 113. i The wavelength λ output from each laser array is i The laser light is modulated to generate a light intensity signal X λi Optical modulator arrays 114 to 114 N and a wavelength multiplexer (wavelength MUX) 115 that wavelength-multiplexes the signals from each optical modulator to generate a wavelength-multiplexed optical signal (i=1, 2, 3, . . . , N, where N is a natural number equal to or greater than 2).
[0021] The wavelength-multiplexed optical signal generated by the wavelength-multiplexed optical signal generation unit 110 is an input signal vector (X λ1 , X λ2 ..., X λNThis wavelength-multiplexed optical signal is introduced into the optical matrix calculation unit 120 by the optical transmission unit 140. The optical matrix calculation unit 120 transmits the wavelength-multiplexed optical signal input from the optical transmission unit 140 through M optical paths 121 1 to 121 M an optical branching device 122 that branches into optical paths 1211 to 121 M wavelength weighting banks 123 1 to 123 2 that weight each wavelength of the M wavelength-multiplexed optical signals. M are arranged (M is a natural number of 2 or more).
[0022] Wavelength weighting banks 1231 to 123 M wavelength demultiplexers (wavelength DUMUX) 1241 to 124 M , each wavelength λ1 to λ N N variable optical attenuators (VOAs) 125 connected to the optical path of M1 ~125 MN , wavelengths λ1 to λ2 from each variable optical attenuator N Wavelength multiplexers (wavelength MUX) 1261 to 126 that multiplex the wavelengths of the light M It is composed of:
[0023] The M branched wavelength multiplexed signal lights are respectively transmitted to wavelength demultiplexers (wavelength DEMUX) 1241 to 1244 M The wavelengths are separated by λ i (i=1, 2, 3, . . . , N) and each optical path 121 j Each variable optical attenuator (VOA) 125 for each j=1, 2, 3, . . . , M ji At ω ji and the wavelength multiplexers 126 1 to 126 M The wavelengths are multiplexed at
[0024] The multi-channel optical receiving unit 130 comprises optical receivers 131 1 to 131 MThe optical receiver includes a multi-channel receiver array configured with the above components, an analog-to-digital converter (ADC) 132 that converts the optical intensity signals detected by each optical receiver into digital signals, and a digital circuit 133 that processes the digital signals from the ADC. As described above, the nonlinear conversion may be performed by the digital circuit 133, or may be performed by disposing some kind of nonlinear optical element. In this configuration, the nonlinear conversion is performed by the digital circuit 133. When a nonlinear optical element is used, for example, the gain saturation phenomenon of optical intensity caused by a semiconductor optical amplifier can be utilized.
[0025] As shown in FIG. 1, the wavelength multiplexers 126 1 to 126 M The M output signals from the optical receivers 131 1 to 131 M The output signals consisting of N wavelengths that have been intensity-modulated based on the weight for each wavelength from the optical matrix calculation unit are each received by one optical receiver, and the respective wavelength signals are subjected to intensity summation at this time. Then, the optical receiver 131 for the j-th optical path j Output y j is expressed by the following equation (3): As described above, j=1, 2, 3, . . . , M, and i=1, 2, 3, .
[0026] Therefore, the components of the output signal from each optical receiver correspond to the respective vector components of the vector-matrix product described in the above equation (2). The wavelength-multiplexed optical signal generation unit 110 and the multi-channel optical receiving unit 130 in Figure 1 can be wavelength-multiplexed transmitters and optical receivers that are generally used in optical communications.
[0027] The configuration in FIG. 1 performs a matrix operation on the input signal in intensity space. When values are expressed using only light intensity, negative components cannot be expressed. For this reason, some additional processing is required when performing subtraction processing. For example, subtraction processing in such cases is performed by separately calculating positive components and negative components, as in Non-Patent Document 3. This processing is performed by, for example, ω ji The positive and negative components of ω + and ω - and the weight of the positive component ω+ Calculation using only ω and negative component weight ω - In this case, the weight of the negative component, ω - The absolute value of the weight of the negative component is used. After each calculation result is obtained independently, it can be performed by performing the calculation of equation (4) on a digital circuit.
[0028] The number of calculation operations on the digital circuit for this subtraction process is O(M). Therefore, when the numbers M and N are large, this value is sufficiently small compared with the number of calculation processes O(MN) executed by the optical matrix calculation unit.
[0029] Another method for performing the subtraction process is to use the configuration shown in Fig. 2. The optical matrix calculation unit 220 shown in Fig. 2 performs the calculation of equation (4) optically by separately performing positive and negative calculation processes for each optical path j and receiving each calculation result using a differential optical receiver.
[0030] 2 is a diagram showing an outline of the configuration of the optical matrix calculation device 200. The optical matrix calculation device 200 in FIG. 2 is a wavelength weighting bank 123 arranged for each optical path j of the optical matrix calculation device 100 shown in FIG. j are stored in two wavelength weighting banks 123 for positive and negative calculation processing. j1 and 123 j2 2, the same reference numerals as in FIG. 1 denote the same components as in FIG. 1, and therefore the explanation thereof will be omitted here.
[0031] The wavelength weighting bank 123 of the optical matrix calculation unit 220 in FIG. 11 , 123 12 , ..., 123 M1 , 123 M2 1, the wavelengths λ1 to λ2 separated by the wavelength demultiplexer (wavelength DUMUX) 124, and the wavelengths λ1 to λ2 separated by the wavelength demultiplexer. N N number of variable optical attenuators (VOA) 125 connected to the optical path of N Each optical path 121 is configured by a wavelength multiplexer (wavelength MUX) 126 that multiplexes the wavelengths of the light. j1(j=1, 2, 3, . . . M) wavelength weight bank 123 j1 Now, the positive weight ω + For each optical path 121 j2 wavelength weighting bank 123 j2 Now, the negative weight ω - At this time, the weight of the negative component, ω - The absolute value of the weight of the negative component is used.
[0032] Wavelength Weighting Bank 123 j1 and 123 j2 The output of the differential optical receiver 231 j Each differential optical receiver 231 receives the light. j The output signal from is expressed by the above-mentioned equation (4). In the configuration of Figure 2, the required circuit scale increases twice as much as that of the optical matrix calculation unit of Figure 1, but an excellent function is realized, such as eliminating the need for digital processing for subtraction in the subsequent stage.
[0033] 1 and 2 illustrate an arithmetic device configured with an optical intensity modulator, an optical intensity attenuator, and an optical intensity receiver, but the optical signal may be calculated in complex space. Calculating the optical signal in complex space can be achieved by providing an optical phase shifter to a coherent optical modulator that performs optical modulation in complex space, a coherent optical receiver, and an intensity attenuator in the optical matrix calculation unit. Performing calculations in complex space provides excellent functionality, such as the ability to perform addition and subtraction operations in one go.
[0034] The optical matrix operation units 120 and 220 of the optical matrix operation devices shown in Figures 1 and 2 require a large number of multiplexing / demultiplexing elements and variable attenuator arrays. Specifically, to be able to perform the matrix operation of N rows and M columns of the above equation (1), M optical weight banks are required, and each optical weight bank requires two multiplexing / demultiplexing elements and N variable attenuators. Therefore, in the configurations described in Figures 1 and 2, it is not practical to increase M.
[0035] The present disclosure proposes an optical matrix calculation device equipped with an optical matrix calculation unit that increases MN and efficiently realizes large-scale optical matrix calculations. The present disclosure proposes an optical matrix calculation device equipped with an optical matrix calculation unit that can realize large-scale optical matrix calculations, which makes it possible to perform matrix calculations at calculation speeds that support high-speed optical modulation signals used in optical communications, etc.
[0036] First Embodiment A first embodiment of the present disclosure will be described with reference to Fig. 3. In the first embodiment, an optical matrix calculation unit of an optical matrix calculation device that realizes the calculation principle described in Fig. 1 and Fig. 2 is configured using a spatial optical system.
[0037] FIG. 3 is a diagram schematically illustrating an optical matrix calculation unit 320 used in an optical matrix calculation device according to a first embodiment of the present disclosure. The optical matrix calculation device according to the first embodiment includes the wavelength-multiplexed optical signal generation unit and multi-channel optical receiving unit shown in FIG. 1 or 2 , and the optical matrix calculation unit 320. The optical matrix calculation unit 320 in FIG. 3 includes an optical splitter 321, an optical fiber array 322, an optical element group 323, and a spatial light modulator 324. FIG. 3( a) is a perspective view, FIG. 3( b) is a view from the x-axis direction, and FIG. 3( c) is a view from the y-axis direction. In FIG. 3 , the direction in which the optical fibers of the optical fiber array 322 are arranged is indicated as the y-axis, the optical axes of the optical fibers are indicated as the z-axis, and the direction perpendicular to the y-axis and z-axis is indicated as the y-axis.
[0038] For ease of understanding, Fig. 3(b) shows only the collimator lens 333 and the Fourier lens 336, which affect the optical path in the yz plane when viewed from the x-axis direction, of the optical element group 323 shown in Fig. 3(a), and omits the other elements. Similarly, Fig. 3(c) shows only the collimator lens 333, the diffraction grating 335, and the scan lens 337, which affect the optical path in the xz plane when viewed from the y-axis direction, and omits the other elements.
[0039] The optical splitter 321 has an input port connected to the optical transmission means 140 and output ports respectively connected to M input optical fibers that form the input optical paths of the optical matrix operation unit. Similar to the optical splitters 122 and 222 in Figures 1 and 2, the optical splitter 321 splits the wavelength-multiplexed optical input signal transmitted and input from the wavelength-multiplexed optical signal generation unit 110 by the optical transmission means 140 into M optical paths.
[0040] The optical fiber array 322 includes M input optical fibers 331 1 to 331 M , and M output optical fibers 332 1 to 332 M Input optical fibers 331 to 331 M form M input optical paths j of the optical matrix calculation unit.
[0041] The optical element group 323 includes input optical fibers 331 1 to 331 M input collimator lenses 3331 to 333 that direct the light emitted from the M a diffraction grating 335 for wavelength dispersion of the incident wavelength multiplexed optical input signal; and output collimator lenses 334 1 to 334 2 for collecting the wavelength dispersed light by the diffraction grating at different positions on the spatial light modulator for each of M optical paths j and for each wavelength, and for transmitting these signals to output optical fibers corresponding to the input optical fibers. M , a Fourier lens 336 and a scan lens 337 for coupling via a Fourier lens 336 and a scan lens 337, and an output collimating lens.
[0042] 3, the light dispersed by the diffraction grating is developed as a continuous wavelength spectrum in the x-axis direction on the spatial light modulator 324. Therefore, the light on the spatial light modulator 324 is developed as a continuous wavelength spectrum in the x-axis direction on the spatial light modulator 324. N The M input signals expanded into the above are focused onto different N×M regions (note that in the embodiment of the present disclosure, N and M are natural numbers equal to or greater than 2, as in the above-described FIGS. 1 and 2 ).
[0043] The spatial light modulator 324 is an element in which a plurality of pixels are arranged in a matrix on the xy plane, and has a driver (not shown) that drives each of the plurality of pixels. By driving the plurality of pixels in the matrix by the driver, it is possible to independently control the intensity or phase of light irradiated onto the spatial light modulator for each predetermined region. As described above, in the spatial light modulator, the input signal light is irradiated along the optical path (j) and the wavelength channel (λ i Therefore, in this configuration, by displaying independent patterns in the regions on the spatial light modulator that correspond to the respective focusing positions, the spatial channels (optical paths j) and wavelength channels (λ i ) can be modulated individually and independently.
[0044] Spatial light modulators are broadly divided into phase modulation types (for example, liquid crystal on silicon, LCOS) and intensity modulation types (for example, digital mirror device, DMD), and either type can be used in this embodiment.
[0045] In the intensity modulation type and the phase modulation type, the weight ω ji For example, in the case of a spatial light modulator capable of multilevel intensity modulation, the desired attenuation intensity ω ji By simply displaying the pixels according to ji In addition, if multi-value modulation is not possible and only binarized values can be set, the ratio of pixel values (corresponding to 1) that combine with respect to the area of the region irradiated by each light beam and pixel values (corresponding to 0) that decay can be controlled to generate a multi-value ω ji It is possible to implement
[0046] In the case of a phase modulation type spatial light modulator, it is necessary to convert the phase modulation into intensity modulation. For example, by displaying a periodic phase pattern (e.g., a well-known phase control pattern such as a sine wave or a sawtooth wave) in the area on the spatial light modulator where each light beam is irradiated, and controlling the period of the phase pattern for each area, the diffraction angle of the light emitted from the spatial light modulator and the distribution ratio of the light intensity to each order of diffracted light can be controlled. This makes it possible to individually control the coupling efficiency to the output optical fiber. In this way, it is possible to convert the phase modulation by the spatial light modulator into intensity modulation.
[0047] In the present disclosure, various types of spatial light modulators and control methods can be applied, including not only liquid crystal but also MEMS (Micro Electro Mechanical System), etc. For example, when a ferroelectric liquid crystal is used as the liquid crystal, it is possible to self-maintain the weights assigned to each signal.
[0048] The optical signals modulated for each wavelength by the spatial light modulator are wavelength-multiplexed again by the diffraction grating 335, and output from the output optical fiber 332 according to the optical path j. 1~M The output optical fibers form the output optical paths of the optical matrix operation unit. The output optical fibers are connected to optical receivers (PDs) 131 that constitute the multi-channel optical receiving unit 130 described in FIG. 1. The signals received by each optical receiver are described by equation (3), as in FIG. 1. When this embodiment performs addition and subtraction by optical operations as described in FIG. 2, the differential optical receivers 231 of the multi-channel optical receiving unit 230 in FIG. 2 are each connected to two adjacent output optical fibers.
[0049] As described above, in the first embodiment of the present disclosure, the wavelength demultiplexing elements and wavelength multiplexing elements provided in each of the M optical weighting banks constituting the optical matrix calculation unit are realized by a single diffraction grating, and the N variable optical attenuators (VODs) provided in each of the M optical weighting banks are realized by a single spatial light modulator. In this way, in this embodiment, the optical matrix calculation unit that realizes the calculation principle described in Figures 1 and 2 is realized by a simple configuration of the optical element group and spatial light modulator of Figure 3.
[0050] The advantage of the optical matrix calculation method according to the present disclosure is that calculations are performed in parallel in the directions of optical path M and wavelength N. The calculation speed of the optical matrix calculation unit is determined by the product MN of the number of optical paths M and the number of wavelengths N that can be implemented. Furthermore, since the optical calculations of this method are performed for each symbol of the wavelength-multiplexed optical input signal, the calculation speed also varies depending on the Baud rate. These relationships are summarized in the following equation (5):
[0051] Here, OPS is the number of product-sum calculations per unit time, and B is the Baud rate of the optical signal. (opt) The relationship between the number of wavelengths N and the Baud rate is expressed by equation (6).
[0052] The relationship of this equation (6) is that the signal spreads in the wavelength direction (frequency direction) according to the Baud rate, so that the signal spreads in the wavelength direction (frequency direction) according to the Baud rate. (opt) This is because there is a limit to the number of wavelengths N that can be packed into the optical matrix calculation device. Here, α indicates the wavelength filter performance of the optical matrix calculation device. Ideally, ω set on a rectangular wave in the wavelength direction ji However, in reality, a rectangular filter cannot be created due to the influence of the beam diameter, etc.
[0053] As shown in FIG. (grid) The effective filter bandwidth BW within a given wavelength grid band is shown by (eff) BW is the ratio of (eff) / BW (grid) This determines the wavelength utilization efficiency on the spatial light modulator. By substituting equation (6) into equation (5), the following equation (7) is obtained.
[0054] In the case of the optical matrix calculation unit 320 of this embodiment, BW (opt) reflects the wavelength area that fits within the pixel area of the spatial light modulator, so the number of pixels (l x ) Therefore, the following equation (8) is obtained.
[0055] Here, β is a proportionality constant proportional to the dispersion power of the diffraction grating. α is a parameter that depends on the beam diameter in the wavelength direction and the pixel size. It has been shown in the design of switch optical systems for optical communications that, with appropriate spatial optical design, α = 0.8 (3 dB bandwidth 40 GHz) can be achieved when the wavelength spacing is 50 GHz (see Non-Patent Document 6).
[0056] In the configuration of this embodiment, the number of spatial channels M in the optical path direction is determined by the beam density ρ in the y-axis direction in FIG. 3 and the number of pixels (l y ) is determined by the following equation (9).
[0057] From these relationships, the upper limit of the number of calculations in this embodiment is given by the following equation (10).
[0058] In equation (10), α and β are in a trade-off relationship. (opt) If the dispersion power of the diffraction grating is reduced in order to improve , the wavelength dispersion on the spatial light modulator decreases, and the filter performance deteriorates. Therefore, it is difficult to significantly improve the value of (α / β) in equation (9) in an optimized optical system.
[0059] From the above, the total number of pixels of the spatial light modulator is l x ×l y and the beam density ρ in the y-axis direction are parameters in the above equation (10) that give the upper limit of the number of calculations. x ×l y As display standards evolve, the options for ρ have expanded, with the number of options increasing by about four times every few years, from high resolution (approximately 1000 x 2000), 4K (approximately 2000 x 4000), to 8K (approximately 4000 x 8000). ρ is a design factor that depends on the optical system configuration, and it is desirable to design it so that it approaches 1.
[0060] In this embodiment, the beam density ρ in the y-axis direction is determined by the pitch interval p of the fiber array and the beam diameter ω of the light emitted from the optical fiber. fiber These relationships are described by the following equation (11):
[0061] In the case of a general one-dimensional optical fiber array, p is the diameter of the fiber core, ω fiber is determined by the diameter of the fiber core. These values have no design freedom due to standards, mechanical strength, etc., and ρ is usually about 0.01. Therefore, the upper limit of the calculation speed in the first embodiment is restricted by ρ.
[0062] Second Embodiment A second embodiment of the present disclosure will be described with reference to FIG. 5 . The second embodiment is characterized by enabling an increase in beam density ρ. FIG. 5 is a diagram schematically illustrating an optical matrix calculation unit 520 used in an optical matrix calculation device according to the second embodiment. Similar to the first embodiment, the optical matrix calculation device according to the second embodiment includes the wavelength-multiplexed optical signal generation unit, the multi-channel optical receiving unit, and the optical matrix calculation unit 520 shown in FIG. 1 or 2 . The optical matrix calculation unit 520 in FIG. 5 includes an optical branching unit (not shown), an optical waveguide front end 522, an optical element group 323, and a spatial light modulator 324. FIG. 5( a) is a perspective view, FIG. 5( b) is a view from the x-axis direction, and FIG. 5( c) is a view from the y-axis direction. Similar to FIGS. 3( b) and 3( c) illustrating the first embodiment, FIGS. 5( b) and 5( c) omit some of the optical elements that constitute the optical element group and do not affect the optical path in the observation direction.
[0063] The optical element group 323 and spatial light modulator 324 of the second embodiment have the same functions as the optical element group 323 and spatial light modulator 324 of the first embodiment. This embodiment differs from the first embodiment in that, instead of the optical fiber array 322 that forms the input and output optical paths of the optical matrix calculation unit, the input and output optical paths are formed by an optical waveguide front end 522. In Fig. 5, the same reference numerals as in Fig. 3 are the same as those in Fig. 3, and therefore, description thereof will be omitted here.
[0064] 5, in the optical matrix calculation unit 520 of this embodiment, each multi-wavelength optical signal is emitted into space at a different angle for each optical path j from the optical phased array emission unit 524 of the optical waveguide front end 522, and each multi-wavelength optical signal is wavelength-demultiplexed by the diffraction grating 335. The wavelength-demultiplexed signals are input to the spatial light modulator 324 for optical path j and wavelength λ as in the first embodiment. i As described in the first embodiment, on the spatial light modulator 324, different spatial lights can be modulated independently by displaying independent patterns in the respective regions.
[0065] The optical signals modulated by the spatial light modulator 324 are wavelength-multiplexed again by the diffraction grating 335 and imaged at the optical phased array incident unit 526 of the optical waveguide front end at different angles according to the optical path j. The light waves incident on the optical phased array incident unit are output to the corresponding output optical path j of the output-side waveguide unit according to the incident angle for each optical path j to the optical phased array incident unit 526. As is clear from the above description, in the second embodiment as well, an arbitrary attenuation state ω can be obtained according to each wavelength λi and optical path j, as in the first embodiment. ji can be given.
[0066] Next, the optical waveguide front end 522 will be described with reference to Fig. 6. Fig. 6 is a diagram showing a schematic configuration of the optical waveguide front end. Fig. 6 shows an optical circuit board 640 and an optical waveguide front end 522 connected to the optical circuit board 640. The optical circuit board 640 includes an optical branching unit 641 that branches a wavelength-multiplexed optical input signal input from a wavelength-multiplexed optical signal generating unit via an optical transmission means 140 into M optical paths, and optical waveguide arrays 6421 to 6422 each having M waveguides connected to an output-side waveguide unit of the optical waveguide front end. M It is equipped with:
[0067] The optical waveguide front end 522 includes an input waveguide section 650 connected to the output port of the optical splitter, and an output waveguide section 660 connected to the waveguide array 642 .
[0068] The input waveguide section 650 includes a plurality of first input waveguide sections 651 1 to 651 Man input slab waveguide 652 that expands and diffuses the optical beam diameter of the signal light input from each of the first input waveguide sections, and guides the signal light; and a plurality of second input waveguide sections 653 to 653, into which the signal light diffused by the input slab waveguide 652 is incident. M’ and an optical phased array output unit 524 that converts the output angle in accordance with the wavefront of the signal light output from each of the second input waveguide units, and outputs each signal light with the converted output angle to an optical element group.
[0069] The output-side waveguide unit 660 includes an optical phased array incident unit 526 into which each signal light output from the spatial light modulator via the optical element group is incident, and a plurality of first output waveguide units 663 to 663 that respectively guide the signal light incident on the optical phased array incident unit 526. M’ and each of the first output waveguide sections 663 1 to 663 M’ and a plurality of second output waveguide portions 661 to 661 for guiding the signal lights output from the output slab waveguide 662. M It includes:
[0070] Here, the numbers M and M' (M' is a natural number of 2 or more) of the first input waveguide section, the second input waveguide section, the first output waveguide section, and the second output waveguide section can be designed independently, but it is desirable to design M'≧M in order to reduce waveguide loss and crosstalk of each output beam. WG and the pitch interval P of the output section of the second input waveguide section. WG , and M', M'·P WG ≧2ω WG It is desirable to design it so that the following relationship holds:
[0071] As shown in the figure, a plurality of second output waveguide sections 666 1 to 661 M are optical waveguide arrays 6421 to 642 M These optical waveguide arrays 642 1 to 642 Mare connected to the optical receivers of the multi-channel optical receiving unit shown in Fig. 1. When addition and subtraction are performed by optical arithmetic as described in Fig. 2 in this embodiment, the differential optical receivers 231 of the multi-channel optical receiving unit 230 shown in Fig. 2 are connected to two adjacent waveguides.
[0072] When the optical waveguide front end 522 of this embodiment is used, the beam density ρ in the y-axis direction on the spatial light modulator is M Pitch interval P d and the distance f of the input slab waveguide 652 sbt (See Non-Patent Document 4.) These relationships can be expressed as the following equation (12).
[0073] The second embodiment of the present disclosure has an excellent feature in that, by using the optical waveguide front end 522, it is possible to freely design the beam density ρ in the y-axis direction, which was difficult in the first embodiment.
[0074] 7A and 7B are diagrams for explaining the advantages of the second embodiment in terms of the operation speed OPS compared to the conventional technology. Fig. 7A shows the relationship between the number of spatial channels M realized by the optical matrix operation device and the operation speed OPS. Fig. 7B shows the bandwidth of the conventional method of Non-Patent Document 3 and its problems.
[0075] The solid line in Figure 7(a) plots the calculated relationship between the number of spatial channels M and the calculation speed OPS achieved using the optical matrix calculation device of the second embodiment. The dashed-dotted line indicates the theoretical upper limit of OPS (the Nyquist limit region where α = 1). For comparison, the reported OPS values of the prior art are listed as the values in Non-Patent Document 1 [1], Non-Patent Document 2 [2], and Non-Patent Document 3 [3].
[0076] In the trial calculation of the calculation speed OPS of this embodiment, a beam density ρ=0.032 that can be realized using an optical waveguide front end was used. This value allows for sufficient margin for crosstalk between spatial channels, controllability of the spatial light modulator, etc., and is a realistic value that allows for high ease of design when realizing the optical matrix calculation device of this embodiment. In addition, α=0.74 was used in the trial calculation of the filter performance. This value also refers to a value that can be fully realized with existing optical systems.
[0077] As shown in FIG. 7A, the number of pixels in the y-axis direction of the spatial light modulator is l y The value of the number of spatial channels (the number of optical paths j) M can be increased depending on the number of spatial channels. For example, a 4K spatial light modulator can achieve 128 spatial channels. In this case, the number of calculations per unit time is 888 TOPS, making it possible to achieve high-speed calculations approaching the petascale. In this way, this embodiment can provide a large-scale optical arithmetic device that supports high-speed optical modulation signals used in optical communications, etc.
[0078] The major difference between the optical matrix calculation device of the present disclosure and the conventional technology shown for reference is the difference in the number of realizable spatial channels M and the value of the filter performance α of the optical matrix calculation device. For example, in the method of Non-Patent Document 3, the resonance by the microring resonator is expressed as ω ji , the bandwidth is inherently narrow, resulting in a low value of α (see FIG. 7(b)). As a result, the OPS is significantly lower than the theoretical upper limit (the dashed-dotted line in FIG. 7(a), the Nyquist limit region where α = 1), as shown by [3] in the figure. Regarding the number of spatial channels M, in the method of Non-Patent Document 3, the optical input signal must pass through the microring resonator structure a number of times proportional to the number M. This causes optical loss proportional to M, which limits the scale of the system, and is currently limited to systems with M = 4 or so.
[0079] In contrast to this, in the first and second embodiments of the present disclosure, a spatial optical system is used to continuously shape the spectrum, thereby enabling a much wider bandwidth than the method of Non-Patent Document 3. Furthermore, as is clear from the above explanation, even if the number of spatial channels M increases, the number of components that pass through does not increase (they are spatially parallelized), and therefore, an increase in fundamental loss as occurs in the optical system of the method of Non-Patent Document 3 does not occur.
[0080] As described above, the embodiment according to the present disclosure can dramatically increase the number of spatial channels M and the filter performance α compared to the conventional technology, thereby providing an optical matrix operation device with excellent performance, with the number of calculations per unit time approaching the petascale.
[0081] Third Embodiment Finally, a third embodiment will be described with reference to Fig. 8. As described above, the optical matrix calculation device of the present disclosure includes a wavelength-multiplexed optical signal generation unit, an optical matrix calculation unit, and a multi-channel optical receiving unit. The wavelength-multiplexed optical signal generation unit and the optical matrix calculation unit are connected by an optical transmission means. The wavelength-multiplexed optical signal transmitted from the wavelength-multiplexed optical signal generation unit 810 to the optical matrix calculation unit 820 is wavelength-multiplexed. Therefore, the optical transmission means 840 can be formed of a single optical fiber.
[0082] The transmission of the wavelength-multiplexed optical signal from the wavelength-multiplexed optical signal generation unit 810 to the optical matrix operation unit 820 is generally the same as in a simple wavelength-multiplexed transmission system. Therefore, as shown in Figure 8, it is possible to arrange the wavelength-multiplexed optical signal generation unit and the optical matrix operation unit separately. In addition, the optical fiber in this transmission section can be extended.
[0083] Therefore, in this embodiment, the wavelength-multiplexed optical signal generation unit 810 and the optical matrix operation unit 820 are arranged in different, spaced-apart positions. Furthermore, by separately arranging the optical matrix operation unit 820 and the multi-channel optical receiving unit 830 and connecting them with M optical transmission means 850 corresponding to the number of spatial channels, the calculation result from the optical matrix operation unit 820 can be transmitted as an optical signal to the multi-channel optical receiving unit 830. Therefore, the optical matrix operation unit 820 of the optical matrix operation device can also be arranged between the wavelength-multiplexed optical signal generation unit 810 and the multi-channel optical receiving unit 830, connected by optical transmission means such as optical fiber.
[0084] As described above, the optical matrix calculation unit 820 disclosed herein makes it possible to perform large-scale matrix calculations in the optical domain at calculation speeds that support high-speed optical modulation signals used in optical communications, etc.
[0085] Therefore, in this embodiment, by arranging the wavelength multiplexing generation unit, optical matrix calculation unit, and multi-channel optical receiving unit that constitute the optical matrix calculation device of the present disclosure as shown in Figure 8, it becomes possible to perform processing such as performing some matrix calculations in the optical domain during transmission of an optical signal from an optical transmitter (wavelength multiplexing signal generation unit 810) to an optical receiver (multi-channel optical receiving unit 830), and new usage forms can be realized using this.
[0086] In addition, in this usage scenario, the optical transmission and reception system, which is the main factor in the power consumption and cost of the optical matrix calculation device, can be shared with the optical transmission device. In addition, since the matrix calculation is performed in the optical domain, there is no need to store it in memory, transfer it to the CPU / GPU, or perform calculation processing, which has the excellent effect of improving memory utilization efficiency and calculation latency.
[0087] For example, as described in Non-Patent Document 5, matrix operations within a deep learning model may be distributed to each computing unit, and the results of these operations may be integrated via optical communication. If each distributed matrix operation is performed by the optical matrix operation unit of the present disclosure, it becomes possible to perform the matrix operation while transmitting signals between each computing unit. As a result, each computing unit does not need to receive optical signals and store them in memory, which provides excellent effects such as improved performance in terms of power consumption, cost, latency, and memory efficiency.
[0088] It is possible to provide a large-scale optical arithmetic device that supports high-speed optical modulation signals used in optical communications and the like.
Claims
1. An optical matrix calculation device comprising: a wavelength-multiplexed optical signal generation unit which generates a vector or tensor using a wavelength-multiplexed optical signal; an optical matrix calculation unit; a multi-channel optical receiving unit; and an optical transmission means which transmits the wavelength-multiplexed optical signal output from the wavelength-multiplexed optical signal generation unit to the optical matrix calculation unit, wherein the optical matrix calculation unit comprises: a spatial light modulator capable of independently modulating light focused on the spatial light modulator for each predetermined region by driving a plurality of pixels arranged in a matrix on a plane using a driving unit; an optical splitter comprising M input optical paths (M is a natural number of 2 or more) and M output optical paths; an input port connected to the optical transmission means and an output port connected to each of the M input optical paths; and a group of optical elements arranged to disperse the wavelength-multiplexed optical signal of the input optical path by wavelength and focus it at different positions on the spatial light modulator, and to couple each of the signal light emitted from the different positions of the spatial light modulator to the M output optical paths, an optical matrix calculation device, wherein the multi-channel optical receiving unit includes optical receivers connected to the output optical paths, respectively.
2. The optical matrix calculation device according to claim 1, characterized in that the wavelength-multiplexed optical signal generating unit comprises: a light source that generates light of N different wavelengths (N is a natural number equal to or greater than 2); N optical modulators that individually modulate the light of the N different wavelengths from the light source; and a wavelength combiner that combines the light from each of the optical modulators and outputs it to the optical transmission means.
3. The optical matrix calculation device according to claim 2, characterized in that the group of optical elements includes a diffraction grating that converts each of the M wavelength-multiplexed optical signals from the input optical path into N wavelength-dispersed light, and a group of lenses that focus each of the M x N wavelength-dispersed light at different positions on the optical spatial modulator.
4. An optical matrix operation device according to claim 1, wherein said input optical path and said output optical path are formed by optical fibers.
5. An optical waveguide front-end including an input-side waveguide section forming the input optical path and an output-side waveguide section forming the output optical path, wherein the input-side waveguide section includes a plurality of first input waveguide sections connected to an output port of the optical splitter, an input-side slab waveguide that expands and guides the optical beam diameters of each of the signal lights input from the plurality of first input waveguide sections, a plurality of second input waveguide sections into which the signal lights diffused by the input-side slab waveguide are respectively incident, and an optical phased array output section that converts an output angle in accordance with a wavefront of the signal lights respectively output from each of the plurality of second input waveguide sections and outputs the converted output angle to the optical element group, 2. The optical matrix calculation device according to claim 1, comprising an optical waveguide front end, wherein the output-side waveguide unit includes an optical phased array input unit that inputs each signal light output from the optical spatial modulator via the optical element group, a plurality of first output waveguide units that respectively guide each signal light input to the optical phased array input unit, an output slab waveguide that guides each signal light input from each of the plurality of first output waveguide units, and a plurality of second output waveguide units that guide each signal light input from the output slab waveguide.
6. The optical matrix calculation device according to claim 1, characterized in that the wavelength-multiplexed optical signal generating unit includes a coherent optical modulator, the spatial light modulator modulates the intensity and phase of the optical signal, and the multi-channel optical receiving unit includes a coherent optical receiver.
7. An optical matrix calculation device according to claim 1, wherein the optical receivers connected to the respective output optical paths output the difference between two adjacent output optical paths.
8. An optical matrix calculation device as described in any one of claims 1 to 7, characterized in that the optical transmission means is formed of a single optical fiber, and the wavelength-multiplexed optical signal generation unit and the optical matrix calculation unit are arranged at separate locations.
Citation Information
Patent Citations
Photon Tensor Accelerator for Artificial Neural Networks
JP2022529870A
Cited By
Optical computing device
WO2026126440A1
