CN113960718B - Photoelectric hybrid incoherent complex value matrix operation processor and complex value matrix operation method - Google Patents
Photoelectric hybrid incoherent complex value matrix operation processor and complex value matrix operation method Download PDFInfo
- Publication number
- CN113960718B CN113960718B CN202111193964.9A CN202111193964A CN113960718B CN 113960718 B CN113960718 B CN 113960718B CN 202111193964 A CN202111193964 A CN 202111193964A CN 113960718 B CN113960718 B CN 113960718B
- Authority
- CN
- China
- Prior art keywords
- matrix operation
- matrix
- real
- optical
- complex value
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G02—OPTICS
- G02B—OPTICAL ELEMENTS, SYSTEMS OR APPARATUS
- G02B6/00—Light guides; Structural details of arrangements comprising light guides and other optical elements, e.g. couplings
- G02B6/24—Coupling light guides
- G02B6/26—Optical coupling means
- G02B6/28—Optical coupling means having data bus means, i.e. plural waveguides interconnected and providing an inherently bidirectional system by mixing and splitting signals
- G02B6/293—Optical coupling means having data bus means, i.e. plural waveguides interconnected and providing an inherently bidirectional system by mixing and splitting signals with wavelength selective means
- G02B6/29331—Optical coupling means having data bus means, i.e. plural waveguides interconnected and providing an inherently bidirectional system by mixing and splitting signals with wavelength selective means operating by evanescent wave coupling
- G02B6/29335—Evanescent coupling to a resonator cavity, i.e. between a waveguide mode and a resonant mode of the cavity
- G02B6/29338—Loop resonators
-
- G—PHYSICS
- G02—OPTICS
- G02B—OPTICAL ELEMENTS, SYSTEMS OR APPARATUS
- G02B6/00—Light guides; Structural details of arrangements comprising light guides and other optical elements, e.g. couplings
- G02B6/24—Coupling light guides
- G02B6/26—Optical coupling means
- G02B6/28—Optical coupling means having data bus means, i.e. plural waveguides interconnected and providing an inherently bidirectional system by mixing and splitting signals
- G02B6/293—Optical coupling means having data bus means, i.e. plural waveguides interconnected and providing an inherently bidirectional system by mixing and splitting signals with wavelength selective means
- G02B6/29346—Optical coupling means having data bus means, i.e. plural waveguides interconnected and providing an inherently bidirectional system by mixing and splitting signals with wavelength selective means operating by wave or beam interference
- G02B6/29347—Loop interferometers, e.g. Sagnac, loop mirror
-
- G—PHYSICS
- G02—OPTICS
- G02B—OPTICAL ELEMENTS, SYSTEMS OR APPARATUS
- G02B6/00—Light guides; Structural details of arrangements comprising light guides and other optical elements, e.g. couplings
- G02B6/24—Coupling light guides
- G02B6/26—Optical coupling means
- G02B6/28—Optical coupling means having data bus means, i.e. plural waveguides interconnected and providing an inherently bidirectional system by mixing and splitting signals
- G02B6/293—Optical coupling means having data bus means, i.e. plural waveguides interconnected and providing an inherently bidirectional system by mixing and splitting signals with wavelength selective means
- G02B6/29379—Optical coupling means having data bus means, i.e. plural waveguides interconnected and providing an inherently bidirectional system by mixing and splitting signals with wavelength selective means characterised by the function or use of the complete device
- G02B6/2938—Optical coupling means having data bus means, i.e. plural waveguides interconnected and providing an inherently bidirectional system by mixing and splitting signals with wavelength selective means characterised by the function or use of the complete device for multiplexing or demultiplexing, i.e. combining or separating wavelengths, e.g. 1xN, NxM
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/14—Fourier, Walsh or analogous domain transformations, e.g. Laplace, Hilbert, Karhunen-Loeve, transforms
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
Abstract
The invention discloses a photoelectric hybrid incoherent complex value matrix operation processor and a complex value matrix operation method, which belong to the field of integrated light calculation and comprise the following steps: the device comprises an electronic control unit, a wavelength division multiplexing unit, an optical complex value matrix operation unit and a data acquisition unit; the optical complex value matrix operation unit comprises a tunable array, wherein a transmission matrix is coded, and a mixed optical signal coded with an input vector is divided into multiple paths of signals with equal power and respectively coupled to each row of resonance devices to complete the complex value matrix operation of a non-negative real number domain; splitting the operation into two matrix operations according to the positive and negative elements and then subtracting the two matrix operations to realize the matrix operation of the full real number domain; the matrix operation is divided into four matrix operations and two times of electrical addition and subtraction through the separation of a real part and an imaginary part, and the complex value matrix operation is realized; and large-scale complex value matrix operation is realized through matrix partitioning, and signal transformation is realized. The invention can construct a processor which is photoelectric mixed, is based on an incoherent framework and can realize large-scale complex value matrix operation.
Description
Technical Field
The invention belongs to the field of integrated optical computing, and particularly relates to a photoelectric hybrid incoherent complex value matrix operation processor and a complex value matrix operation method.
Background
With the rapid development of artificial intelligence technology, the increasing demand for high performance computing has driven the development of custom hardware to speed up certain classes of computing. However, as the exponential scale of electronic transistors reaches the physical limits revealed by moore's law, electronic hardware-based processors have encountered bottlenecks where performance cannot continue to grow. Optical processors use photons instead of electrons for computation, so optical computation can significantly speed up computation by overcoming the limitations inherent in electrons. Unlike integrated circuits, integrated optical circuits have excellent characteristics of ultra-wide bandwidth, high frequency, low energy consumption, and the like, making optical computing methods a viable and competitive candidate for artificial intelligence accelerators.
The current integrated optical computing architecture mainly includes two types, namely a cascaded Mach-Zehnder interferometer (Mach-Zehnder Interferometers) network and a Micro Ring Resonator (MRR) array. Both of these architectures are electrical for control and optical for computational acceleration. The MZI network is a coherent network that retains optical phase information and therefore can perform complex operations. However, the transmission matrix thereof needs to be configured by means of an iterative algorithm, is not suitable for matrix calculation of real-time response, and the power consumption of the MZI device is larger compared with the resonant device. The MRR is a resonant device with compact structure, has low power consumption, and is suitable for large-scale integrated arrays. However, the MRR array has a drawback in that it is an incoherent network, and in optical calculation, only intensity information and no phase information are available, so that complex operations cannot be performed, and the application field is limited. Therefore, it is necessary to construct a processor which is based on an opto-electronic hybrid and non-coherent architecture and can realize large-scale complex matrix operation.
Disclosure of Invention
Aiming at the defects and the improvement requirements of the prior art, the invention provides an optoelectronic hybrid incoherent complex value matrix operation processor and a complex value matrix operation method, and aims to construct an optoelectronic hybrid processor based on an incoherent architecture and capable of realizing large-scale complex value matrix operation.
To achieve the above object, according to an aspect of the present invention, there is provided an optoelectric hybrid incoherent complex value matrix operation processor, including:
an electronic control unit for generating a vector I carrying N dimensions 0 The microwave signal to be detected; n is a positive integer;
a data input unit connected with the electronic control unit for receiving N paths of optical signals with different wavelengths and outputting a vector I 0 Respectively coding the N-dimensional information into the intensity of the N paths of optical signals to obtain N paths of optical carrier microwave signals;
the wavelength division multiplexing unit is used for converging the N paths of optical carrier microwave signals into a path of mixed optical signal;
an optical complex value matrix arithmetic unit connected with the electronic control unit and comprising a tunable array formed by M rows and N columns of resonance devices and used for dividing the mixed optical signal into M paths of signals with equal power and respectively coupling the signals to M rowsIn the resonant device; the resonance wavelengths of the N columns of resonance devices are respectively aligned with the wavelengths of the N optical signals, and the electronic control unit is further used for aligning the matrix X of M multiplied by N 0 Encoding into transmission coefficients of respective resonant devices, thereby obtaining O 0 ＝X 0 I 0 Performing optical complex value matrix operation; m is a positive integer;
the data acquisition unit is connected with the electronic control unit and used for acquiring the result of the optical complex value matrix operation, converting the result into an electric signal and transmitting the electric signal to the electronic control unit; the electronic control unit is also used for storing the matrix operation result and performing addition and subtraction operation on the stored matrix operation result.
The invention provides a photoelectric hybrid incoherent complex value matrix operation processor, wherein a tunable array for realizing optical complex value matrix operation is composed of resonance devices and is an incoherent framework; the electronic control unit is used for storing and adding and subtracting the operation result of the optical complex value matrix, and large-scale complex value matrix operation can be realized through decomposition of matrix operation. Therefore, the invention provides a processor which is photoelectric mixed and based on an incoherent framework and can realize large-scale complex value matrix operation, and the processor can effectively solve the problem that the conventional photoelectric mixed incoherent integrated optical calculation scheme cannot carry out large-scale complex value matrix operation.
In some alternative embodiments, the resonant devices in the tunable array are microring resonators.
In some alternative embodiments, the resonating devices in the tunable array are tunable MZI-assisted microring resonators.
The micro-ring resonator (MRR) and the tunable MZI-assisted micro-ring resonator (MZI-MRR) are compact resonance devices, low in power consumption and suitable for large-scale integrated arrays; the invention utilizes MRR or MZI-MRR to form a tunable array, and has low power consumption, high real-time performance and strong universality during operation; particularly, when the MZI-MRR is used, the MZI-MRR is a novel micro-ring structure based on a combined MZI arm, the extinction ratio and the resonance peak of the micro-ring can be respectively adjusted, and the reconfigurability of the micro-ring is greatly improved.
Furthermore, the data acquisition unit comprises M balanced photodetectors corresponding to the M rows of resonance devices; each balance photoelectric detector is connected with the electronic control unit;
the balance photoelectric detector is used for detecting the output optical power of the straight-through end and the downloading end of the corresponding row of the resonant devices, and converting the output optical power into an electric signal after differential operation.
Furthermore, all devices in the optical complex value matrix arithmetic unit are integrated on a silicon-based chip.
The invention integrates all devices in the optical complex value matrix arithmetic unit on the silicon-based chip, is developed by the mature silicon-based process platform at present, is compatible with the CMOS process, has strong universality and has the potential of large-scale application.
The invention provides a complex value matrix operation method based on a photoelectric hybrid incoherent complex value matrix operation processor, which comprises the following steps:
a non-negative real number domain matrix operation step: carrying an N-dimensional vector I by an electronic control unit 0 The microwave signal to be measured is input into the data input unit, and the matrix X with the scale of M multiplied by N is input into the data input unit 0 Encoding the obtained data into transmission coefficients of the resonant devices to output matrix operation result O by the data acquisition unit 0 ＝X 0 I 0 ；I 0 And X 0 All belong to the non-negative real number domain.
Through the steps, the invention can realize the matrix operation of the non-negative real number domain.
Further, the complex-valued matrix operation method provided by the invention further comprises the following steps:
and (3) performing full real number domain matrix operation: splitting an N-dimensional vector I into N-dimensional vectors I containing all positive elements + And an N-dimensional vector I containing the absolute values of all negative elements - Respectively calculating P = XI by using non-negative number domain matrix operation steps + And Q = XI - Calculating a matrix operation result O = XI = P-Q through differential operation; i belongs to the real number domain, X belongs to the non-negative real number domain, and X has a size of M N.
Through the steps, the matrix calculation is divided into two times by the invention, and the two times are respectively composed of the vector I containing all positive elements + And contains all negative elementsVector of absolute values I - And finally, the transmission matrix and the output vector are expanded to a full-real number domain from a non-negative domain.
Further, the complex-valued matrix operation method provided by the invention further comprises the following steps:
a full complex field matrix operation step: the N-dimensional vector I P Splitting into real parts real (I) P ) And imaginary part imag (I) P ) Dividing the matrix X of M × N P Split into real parts real (X) P ) And imaginary imag (X) P ) Separately calculating real (X) through non-negative real number field matrix operation steps P )real(I P )、imag(X P )imag(I P )、imag(X P )real(I P ) And real (X) P )imag(I P ) And respectively obtaining matrix operation results O by addition and subtraction P ＝X P I P Real part of (O) P )＝real(X P )real(I P )-imag(X P )imag(I P ) And imaginary component imag (O) P )＝real(X P )imag(I P )+imag(X P )real(I P ) Thereby obtaining a matrix operation result of O P ＝real(O P )+i*imag(O P )；I P And X P All belong to the full real number domain.
Through the steps, the invention divides the matrix operation into four optical matrix vector product operations, namely real (X) P )real(I P )、imag(X P )imag(I P )、imag(X P )real(I P ) And real (X) P )imag(I P ) And two times of electrical addition and subtraction operations, therefore, the invention can realize complex matrix operation only by respectively calculating the four matrix vector products in the optical domain and performing addition and subtraction operations in the electrical domain, and the matrix operation is expanded to a full complex number domain.
Further, the complex-valued matrix operation method provided by the invention further comprises the following steps:
large-scale matrix operation steps: will vector I L Splitting the block with the dimension of N to obtain
Will matrix X L Splitting the blocks with the size of M multiplied by N to obtain
Separately computing X through a full complex field matrix operation step kj I j Then, a matrix operation result O is obtained by addition operation L ＝X L I L The k-th block in (a) is O L (k)＝X k1 I 1 +X k2 I 2 +…+X kn I n Thereby obtaining a matrix operation result of
I L And X L All belong to a full real number domain, k is more than or equal to 1 and less than or equal to m, and j is more than or equal to 1 and less than or equal to n.
Through the steps, the invention simplifies the dimension expansion problem of the matrix into multiple times of low-dimension optical matrix vector product multiplication and multiple times of electrical addition by using the principle of matrix partitioning, and finally realizes the operation of a large-scale complex value matrix.
Furthermore, the complex value matrix operation method provided by the invention is used for realizing discrete Walsh-Hadamard transform, discrete cosine transform or discrete Fourier transform.
Because the invention constructs a processor which is photoelectric mixed, based on a non-coherent framework and capable of realizing large-scale complex value matrix operation, based on the processor, the invention can correctly realize three typical signal transformations of Discrete Walsh-Hadamard transform (WHT), discrete Cosine Transform (DCT) and Discrete Fourier Transform (DFT).
Generally, by the above technical solution conceived by the present invention, the following beneficial effects can be obtained:
(1) The invention provides a photoelectric hybrid incoherent complex value matrix operation processor, wherein a tunable array for realizing optical complex value matrix operation is composed of resonance devices and is an incoherent framework; the electronic control unit is used for storing and adding and subtracting the operation result of the optical complex value matrix, and large-scale complex value matrix operation can be realized through decomposition of matrix operation. Therefore, the invention provides a processor which is photoelectric mixed and based on an incoherent framework and can realize large-scale complex value matrix operation, and the processor can effectively solve the problem that the conventional photoelectric mixed incoherent integrated optical calculation scheme cannot carry out large-scale complex value matrix operation.
(2) The photoelectric hybrid incoherent large-scale complex value matrix operation processor provided by the invention divides an input vector and a transmission matrix into a real part and an imaginary part by adopting a complex value matrix decomposition mode, calculates a matrix vector product in an optical domain, and performs addition and subtraction operation in an electric domain to realize complex value matrix operation, thereby expanding the operation domain of the incoherent matrix operation processor.
(3) The photoelectric hybrid incoherent large-scale complex value matrix operation processor provided by the invention adopts a matrix blocking mode, the dimension expansion problem of the matrix is equivalent to the combination of optical matrix vector product multiplication and electrical addition for a plurality of times, and the matrix operation is expanded to a higher dimension under the condition of hardware with limited scale.
(4) The photoelectric hybrid incoherent large-scale complex value matrix operation processor provided by the invention can realize multi-wavelength parallel operation by introducing a wavelength division multiplexing technology, can realize a transmission matrix of any scale in principle, and has hardware expansion capability.
(5) The photoelectric hybrid incoherent large-scale complex value matrix operation processor provided by the invention adopts a novel micro-ring structure based on a combined MZI arm, the extinction ratio and the resonance peak of a micro-ring can be respectively adjusted, and the reconfigurability of the micro-ring is greatly improved.
(6) All on-chip integrated devices adopted by the photoelectric hybrid incoherent large-scale complex value matrix operation processor provided by the invention are developed by the mature silicon-based process platform at present, are compatible with a CMOS (complementary metal oxide semiconductor) process, have strong universality and have large-scale application potential.
Drawings
Fig. 1 is a schematic structural diagram of an optoelectronic hybrid incoherent complex value matrix arithmetic processor according to an embodiment of the present invention;
fig. 2 is a diagram of an experimental apparatus of an optoelectronic hybrid incoherent complex value matrix arithmetic processor according to an embodiment of the present invention;
FIG. 3 is a schematic diagram of an MRR device in an optical complex-valued matrix arithmetic unit according to an embodiment of the present invention; wherein, (a) is the schematic diagram of the MRR device structure, and (b) is the variation of the micro-ring resonance wavelength along with the micro-ring electrode power;
fig. 4 is a schematic diagram illustrating a principle of extending a matrix operation of an optoelectronic hybrid incoherent complex value matrix operation processor to a full complex field according to an embodiment of the present invention;
fig. 5 is a schematic diagram illustrating a principle that a matrix operation of the optoelectronic hybrid incoherent complex value matrix operation processor according to an embodiment of the present invention is expanded to a higher dimension;
fig. 6 is a result of implementing complex matrix operation by the optoelectronic hybrid incoherent complex matrix operation processor according to the embodiment of the present invention; wherein, (a) is a transmission matrix value obtained by measurement, (b) is a corresponding theoretical transmission matrix value, (c) is an ith row vector product output value and an error value obtained by an experiment, and (d) is the distribution condition of the absolute value of the error;
fig. 7 is a schematic diagram of an optoelectronic hybrid incoherent complex value matrix processor according to an embodiment of the present invention implementing three typical signal transformations; wherein, (a) is original signal of WHT transform, (b) is original signal of even symmetric DCT transform, (c) is original signal of DCT transform of half of the aforesaid sequence, (d) is original signal of DFT transform, (e) is result and theoretical value of WHT transform, (f) is result and theoretical value of even symmetric DCT transform, (g) is result and theoretical value of DCT transform of half of the aforesaid sequence, (h) is result and theoretical value of DFT transform;
the same reference numbers will be used throughout the drawings to refer to the same or like elements or structures, wherein:
1-data input unit:
11-a first electro-optical intensity modulator; 12-a second electro-optical intensity modulator; 13-a third electro-optical intensity modulator; 14-a fourth electro-optical intensity modulator;
2-wavelength division multiplexing unit:
3-optical complex value matrix arithmetic unit:
31-an in-coupling grating; 32. 33, 34, 35, 36, 37, 38, 39-out-coupling gratings; 6. 7, 8-MMI 3dB optical splitter;
4-data acquisition unit:
41. 42, 43, 44-balanced photodetectors;
5-an electronic control unit.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the invention. In addition, the technical features involved in the embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.
In the present application, the terms "first," "second," and the like (if any) in the description and the drawings are used for distinguishing between similar elements and not necessarily for describing a particular sequential or chronological order.
In order to construct a processor which is based on an opto-electric hybrid and incoherent architecture and can realize large-scale complex value matrix operation, the opto-electric hybrid incoherent complex value matrix operation processor provided by the invention, as shown in fig. 1, comprises:
an electronic control unit 5 for generating a vector I carrying N dimensions 0 The microwave signal to be detected; n is a positive integer;
a data input unit 1 connected with the electronic control unit 5 for receiving N paths of optical signals with different wavelengths and outputting a vector I 0 Respectively coding the N-dimensional information into the intensity of the N paths of optical signals to obtain N paths of optical carrier microwave signals;
the wavelength division multiplexing unit 2 is used for merging the N paths of optical carrier microwave signals into a path of mixed optical signal;
an optical complex matrix operation unit 3 connected with the electronic control unit 5 and including a tunable array composed of M rows and N columns of resonance devices for dividing the mixed optical signal into M equal in powerThe path signals are respectively coupled to the M rows of resonance devices; the resonance wavelengths of the N columns of resonator devices are aligned with the wavelengths of the N optical signals, respectively, and the electronic control unit 5 is further configured to align the M X N matrix X 0 Encoding into transmission coefficients of respective resonant devices, thereby obtaining O 0 ＝X 0 I 0 Performing optical complex value matrix operation; m is a positive integer;
the data acquisition unit 4 is connected with the electronic control unit 5 and used for acquiring matrix operation results, converting the matrix operation results into electric signals and transmitting the electric signals to the electronic control unit 5; the electronic control unit 5 is further configured to store a matrix operation result and perform addition and subtraction operations on the stored matrix operation result;
the photoelectric hybrid incoherent complex value matrix operation processor is an incoherent framework, and a tunable array for realizing optical complex value matrix operation consists of resonance devices; the electronic control unit is used for storing and adding and subtracting the operation result of the optical complex value matrix, and the optical matrix operation which can only carry out multiplication and addition can be expanded to carry out multiplication, addition and subtraction through the decomposition of the matrix operation, thereby finally realizing large-scale complex value matrix operation. Therefore, the processor is a processor which is photoelectric mixed and based on an incoherent framework and can realize large-scale complex value matrix operation, and can effectively solve the problem that the existing photoelectric mixed incoherent integrated optical calculation scheme cannot perform large-scale complex value matrix operation;
based on the structure, the invention can realize the vector I with the dimension of N 0 And a matrix X of size MxN 0 The matrix-vector product operation of (1) finally obtains an output vector O with the dimension of N 0 (ii) a Wherein, M and N can be any positive integer, and can be equal or unequal; for convenience of description, in the following embodiments, M =4 and n =4 are used as examples for explanation. The following are examples.
Example 1:
an optoelectronic hybrid incoherent complex value matrix arithmetic processor, as shown in fig. 1 and 2, comprising: the device comprises a data input unit 1, a wavelength division multiplexing unit 2, an optical complex value matrix arithmetic unit 3, a data acquisition unit 4 and an electronic control unit 5;
continuous optical signals with different wavelengths are all generated externally and input into the processor, as shown in fig. 2, alternatively, in this embodiment, the optical signals are all generated by a multi-channel light source; due to the input vector I 0 Is 4, and accordingly, optical signals of 4 different wavelengths are inputted from the outside, the 4 wavelengths being λ 1 、λ 2 、λ 3 、λ 4 The four optical signals are equal in intensity; the data input unit 1 is used for loading input data onto an optical carrier to obtain an intensity-coded optical carrier microwave signal, the wavelength division multiplexing unit 2 combines a plurality of paths of optical carrier microwave signals with different wavelengths into one path, the electronic control unit 5 controls the loading of the microwave signals and the voltage of a metal electrode of a resonance device in the optical complex value matrix operation unit 3, the optical complex value matrix operation unit 3 is used for executing incoherent large-scale complex value matrix operation in an optical domain, and an output operation result optical signal is acquired by the data acquisition unit 4 and is sent to the electronic control unit 5 to be processed to obtain an operation result.
As shown in FIG. 1, in the present embodiment, to input a vector I 0 The data input unit 1 specifically includes four electro-optical intensity modulators, that is, a first electro-optical intensity modulator 11, a second electro-optical intensity modulator 12, a third electro-optical intensity modulator 13, and a fourth electro-optical intensity modulator 14, the electronic control unit 5 generates a microwave signal to be measured, and loads a corresponding microwave signal to be measured on each electro-optical intensity modulator in the data input unit, so that the input data information is encoded on the intensity of the optical carrier microwave signal, and four paths of optical carrier microwave signals with different wavelengths and encoded input data information can be obtained. The input is a 4 x 1 vector I 0 ＝[i 1 ,i 2 ,i 3 ,i 4 ] T 4 electro-optical intensity modulators for modulating the wavelength lambda 1 、λ 2 、λ 3 、λ 4 Of the optical signal of which the input vector I is 0 Element i of 1 ，i 2 ，i 3 ，i 4 Are respectively intensity-encoded at a wavelength λ 1 、λ 2 、λ 3 、λ 4 Thereby obtaining 4 optical carrier signals.
As shown in fig. 1, in this embodiment, the wavelength division multiplexing unit 2 includes a wavelength division multiplexer, and combines the multiple optical carrier signals with different wavelengths output by the data input unit 1 by using a wavelength division multiplexing technology, and couples the multiple optical carrier signals to the same optical fiber for transmission, where an output includes four wavelengths λ 1 、λ 2 、λ 3 、λ 4 The mixed optical signal encoding the input vector I 0 ＝[i 1 ,i 2 ,i 3 ,i 4 ] T The information of (2) is sent to the optical complex matrix arithmetic unit 3.
As shown in fig. 1, in the present embodiment, an input vector I is encoded and outputted via a wavelength division multiplexing unit 2 0 ＝[i 1 ,i 2 ,i 3 ,i 4 ] T The mixed optical signal of the information is coupled into the optical complex value matrix arithmetic unit 3 through the input coupling grating 31; the optical complex value matrix arithmetic unit 3 comprises an MRR array with a 4 × 4 scale, in order to divide the mixed optical signal into 4 paths of signals with equal power to be respectively coupled to 4 rows of MRRs, the optical complex value matrix arithmetic unit 3 further comprises a light splitting module composed of three MMI 3dB light splitters 6, 7 and 8 between the input coupling grating 31 and the MRR array, the mixed optical signal output by the input coupling grating 31 is divided into two paths by the MMI 3dB light splitter 6, one of the paths is divided into two paths by the MMI 3dB light splitter 7, and the two paths are respectively output from the optical complex value matrix arithmetic unit 3 from the output coupling gratings 32, 33 and 34 and 35 after passing through the 1 st row and the 2 nd row of the MRR array; the other path is divided into two paths by the MMI 3dB coupler 8, and after passing through the 3 rd line and the 4 th line of the MRR array respectively, the two paths are output from the optical complex value matrix calculation unit 3 through output coupling gratings 36, 37 and 38, 39.
The optical complex-valued matrix arithmetic unit 3 is a silicon-based chip. On the chip, an input vector I is encoded 0 ＝[i 1 ,i 2 ,i 3 ,i 4 ] T Data comprising four wavelengths lambda 1 、λ 2 、λ 3 、λ 4 Is divided into 4 parts of equal power, thenAnd then separately input into each row of a 4 x 4 scale tunable MRR array. There are 4 rows of MRRs on the chip, each row has 4 MRRs, and the resonant wavelength of each MRR located in the same row is precisely aligned with the 4 optical wavelengths λ 1, λ 2, λ 3, λ 4 of the input mixed optical signal, respectively, to form a one-to-one corresponding operational relationship.
The basic principle of the matrix operation implemented by the complex optical value matrix operation unit 3 will be briefly described below with reference to the specific structure of MRR. Fig. 3 (a) is a schematic diagram of the MRRs, and the resonance wavelength of each MRR is aligned with only a specific wavelength of the input mixed optical signal, and therefore does not affect the input of other wavelengths. As shown in fig. 3 (a), MRR is a micro-ring structure based on a combination of straight waveguides, and is composed of two straight waveguides and a ring waveguide, and a metal electrode is disposed on the ring waveguide. The temperature of the metal electrode of the annular waveguide can be changed by changing the voltage or current applied to the electrode, so that the refractive index of the waveguide is changed, and further the circular phase shift of the annular waveguide is changed, so as to change the position of the resonance peak of the MRR, and fig. 3 (b) reflects the process of realizing the position control of the resonance wavelength of the MRR, wherein the bandwidth of the resonance peak of a single MRR is about 0.09nm, and the Free Spectral Range (FSR) is about 11nm; then applying different voltages on the annular waveguide electrode to determine the change of the micro-ring spectral response; the resistance of this electrode was 0.9k omega. Fig. 3 (b) shows the amount of shift in the resonance wavelength of MRR as a function of the phase shifter power. It can be seen that when the voltage applied to the metal electrode was changed from 1.1V to 3.2V, the resonance wavelength was shifted toward the long wavelength by 5.6nm.
Therefore, by establishing a mapping relationship between the transmission spectrum of the MRR and the matrix element values, each MRR can be used to control the transmission coefficient of the optical signal with a specific wavelength in a certain row, and the tunable MRR array with 4 × 4 scale can be used to simulate a transmission matrix with 4 × 4 scale.
Since there is a one-to-one correspondence between MRR and input light in each row, a vector multiplication can be constructed. Assuming that the MRR is lossless, the pass-through and download ends of the MRR should be complementary without loss. Whereby the i-th row and j-th column of micro-rings are at their corresponding wavelengthsλ j Where the transmission coefficient of the down loading end to the incident light is a ij And the transmission coefficient of the straight end to the incident light is 1-a ij . The optical power detected at the downloading end of the ith row can be written in the form of a cross product, i.e. the optical power detected at the downloading end of the ith row
The concept of the vector product of each line is popularized to all the lines and written into a matrix form, and the optical power detected by the downloading end can be obtained
Similarly, the optical power detected at all the through ports can be written as
Due to a ij The ideal value range of [0,1]Therefore, the transmission coefficient x of MRR ij ＝1-2a ij The value range is [ -1,1](ii) a Finally, the operation result of the matrix-vector product can be expressed as:
in practical application, the transmission coefficient of the MRR can be adjusted to a specified value by adjusting the voltage or current value on the metal electrode of the annular waveguide of the MRR; in this embodiment, this adjustment is achieved by the electronic control unit 5.
As shown in fig. 1, in order to collect the matrix operation result output by the optical complex value matrix operation unit 3, in this embodiment, the data collection unit 4 is specifically a balanced photodetector array, which includes 4 balanced photodetectors 41, 42, 43, and 44, and corresponds to 8 output ports of the optical complex value matrix operation unit, and is used for collecting the result of the optical complex value matrix operation, converting the result into an electrical signal, and then sending the electrical signal to the electronic control unit 5, and the electrical control unit 5 stores the electrical signal, and performs subsequent addition and subtraction operations. In the data acquisition unit 4, each flat photodetector can simultaneously detect the optical power output by the through end and the downloading end of one row of MRR array, perform differential operation, and then convert the differential operation result into an electrical signal.
In this embodiment, the electronic control unit 5 is an electronic computer or a field programmable gate array, and is configured to provide an electrical signal carrying input data, configure parameters of the optical complex value matrix operation unit, store operation results, and provide real-time feedback based on an artificial intelligence algorithm according to the working state of each unit of the processor, so that each unit is in an ideal working state.
In this embodiment, both the electro-optical intensity modulator and the optical complex value matrix operation unit chip in the data input unit 1 are polarization-sensitive, so as shown in fig. 2, in this embodiment, the front and the back of the data input unit are also respectively connected with a polarization controller.
It should be noted that, under the architectures shown in fig. 1 and fig. 2, some devices in this embodiment may also be replaced by other devices that achieve the same function, for example, in some other embodiments of the present invention, the input coupling grating 31 and the output coupling gratings 32 to 39 may also be replaced by silicon-based horizontal couplers for coupling with external optical fibers; for another example, in some other embodiments of the present invention, the data acquisition unit may also use two photodetectors to respectively detect optical powers output by the through terminal and the download terminal of the MRR device in the same row; for another example, in some other embodiments of the present invention, the MRR device may be replaced with other incoherent resonator devices such as a tunable MZI-assisted micro-ring resonator (MZI-MRR), where the MZI-MRR is a novel micro-ring structure based on a combination with MZI arms, and the extinction ratio and the resonance peak of the micro-ring can be respectively adjusted, so as to greatly improve the reconfigurability thereof.
Example 2:
a complex value matrix operation method based on the photoelectric hybrid incoherent complex value matrix operation processor provided in embodiment 1 includes:
non-negative real number domain matrix operation step: carrying an N-dimensional vector I by an electronic control unit 0 The microwave signal to be measured is input into the data input unit, and the matrix X with the scale of M multiplied by N is input into the data input unit 0 Encoding the obtained data into transmission coefficients of the resonant devices to output matrix operation result O by the data acquisition unit 0 ＝X 0 I 0 ；I 0 And X 0 All belong to the non-negative real number domain.
Through the above steps, the present embodiment can implement the matrix operation in the non-negative real number domain.
The data of the input vector is intensity-coded by the electro-optical intensity modulator and the intensity is non-negative, so the input vector is still non-negative, and in order to extend the transmission matrix and the output vector from the non-negative real domain to the full real domain, the input vector I is divided into a vector I + containing all positive elements and a vector I-containing all negative elements absolute values, the relation between them and I can be obtained:
I＝I + -I -
two non-negative vectors I + and I-can thus be obtained, both of which can be taken as values representing the light intensity; according to this principle, the matrix calculation can be split into two operations, I + and I-as inputs, respectively, resulting in P and Q. The two output vectors are sent to an electrical calculation part for subtraction to obtain the final real output O. Thus, the matrix-vector product operation of a real number domain can be divided into two non-negative domain optical matrix-vector product operations and one electrical subtraction operation. This process can be formulated as:
P＝XI + ,Q＝XI -
O＝P-Q＝XI
based on the above analysis, the present embodiment further includes:
and (3) performing full-real number domain matrix operation: splitting an N-dimensional vector I into N-dimensional vectors I containing all positive elements + And an N-dimensional vector I containing the absolute values of all negative elements - Respectively calculating P = XI by using non-negative number domain matrix operation steps + And Q = XI - Calculating a matrix operation result O = XI = P-Q through differential operation; i belongs to the real number domain, X belongs to the non-negative real number domain, and X has a size of M N.
For illustrative purposes, a set of real experimental data obtained based on this principle is presented here. The input vector is [1, -0.25,0.5, -1] T According to the method described above, the input vector is divided into I containing only non-negative elements + Has a value of [1,0,0.5,0] T And I containing only the absolute value of the non-positive element - Has a value of [0,0.25,0,1]And T. The transfer matrix X of the chip can be represented as
When two groups of input are respectively input, two groups of output P and Q can be obtained, and the two results are subtracted to obtain an output vector; theoretical values of P and Q are [1,0, -0.5,0.5 respectively] T And [0,0,1, -1] T The theoretical value of the output vector is [1,0.25, -1.5,1.5] T (ii) a The actual values of P and Q calculated by the processor are [1.00, -0.01, -0.49,0.47 respectively] T And [ -0.01,0.21,1.00, -0.99] T The actual output vector value obtained by the subtraction of the computer is [1.01,0.22, -1.49,1.47] T . Cosine similarity is generally used to measure the similarity between two vectors in vector space, and the closer the cosine value is to 1, the closer the included angle is to 0 degree, i.e. the more similar the two vectors are, for vector X = (X) 1 ,x 2 ,…,x n ) And Y = (Y) 1 ,y 2 ,…,y n ) The other chord similarity formulas are:
the cosine similarity between the theoretical value and the actual value is calculated to be 0.99876, which is very close to 1, indicating that the actual value and the theoretical value are very close, that is, based on the full-real number domain matrix operation step of the embodiment, the embodiment can accurately realize the matrix operation of the full-real number domain.
This embodiment may also extend the matrix operation to the full complex field. Fig. 4 is a schematic illustration of the principle of matrix operation extending to the full complex field. The complex number contains a real part and an imaginary part, so that the input vector and the transmission matrix can be split into sums of the real part and the imaginary part, and the operations of the real part and the imaginary part are separately performed
O P ＝X P I P ＝(real(X P )+i*imag(X P ))(real(I P )+i*imag(I P ))
The output is calculated by splitting into real and imaginary parts
real(O P )＝real(X P )real(I P )-imag(X P )imag(I P )
imag(O P )＝real(X P )imag(I P )+imag(X P )real(I P )
According to the above formula, the matrix operation can be split into four optical matrix-vector product operations, real (X) P )real(I P )、imag(X P )imag(I P )、imag(X P )real(I P ) And real (X) P )imag(I P ) And two times of electric addition and subtraction operations, so that complex operations can be realized only by respectively calculating the four matrix vector products in the optical domain and performing addition and subtraction operations in the electric domain.
Based on the above analysis, the present embodiment further includes:
a full complex field matrix operation step: vector I of N dimension P Splitting into real parts real (I) P ) And imaginary part imag (I) P ) Dividing the matrix X of M × N P Split into real parts real (X) P ) And imaginary imag (X) P ) By non-negative real number field matrix operation stepsCalculating real (X) by step P )real(I P )、imag(X P )imag(I P )、imag(X P )real(I P ) And real (X) P )imag(I P ) And respectively obtaining matrix operation results O by addition and subtraction P ＝X P I P Real part of (O) P )＝real(X P )real(I P )-imag(X P )imag(I P ) And imaginary part imag (O) P )＝real(X P )imag(I P )+imag(X P )real(I P ) Thereby obtaining a matrix operation result of O P ＝real(O P )+i*imag(O P )；I P And X P All belong to the full real number domain.
Likewise, for illustrative purposes, a set of real experimental data obtained based on this principle is presented herein. The value of the input vector is I P ＝[1+0.25i,0.5i,0.5+1i,0] T Can be split into real parts [1,0,0.5,0] T And imaginary parts [0.25i,0.5i,1i,0] T . Transmission matrix X P Is also split into real and imaginary parts. The real part and the imaginary part of the optical fiber are respectively
As shown in fig. 4, it can be seen that the multiplication is divided into four groups, arranged from top to bottom, and the results are P, Q, R, and S, respectively. The subtraction of P and Q yields the real part of the output matrix, and the addition of R and S yields the imaginary part of the output matrix, thus the output matrix O P = (P-Q) + i (R + S). Theoretical values of output vector [ -0.66-0.23i,1.23-0.54i, -2.21-0.28i,0.15-0.93i] T . The experimental values of the output vectors are plotted in two-dimensional coordinates in the figure, and the final results are [ -0.57-0.36i,1.23-0.58i, -2.14-0.32i,0.18-0.95i [ -0.57-0.36i ], (ii)] T 。
In order to further enlarge the scale of the processing matrix, the embodiment uses the principle of matrix partitioning to reduce the problem of matrix dimension expansion into multiple times of multiplication of vector products of low-dimensional optical matrices and multiple times of electrical addition, and specifically, the embodiment further includes: large-scale matrix operation steps: will vector I L Splitting the block with the dimension of N to obtain
Will matrix X L Splitting the blocks with the size of M multiplied by N to obtain
Separately computing X by a full complex field matrix operation kj I j Then, a matrix operation result O is obtained by addition operation L ＝X L I L The k-th block in (a) is O L (k)＝X k1 I 1 +X k2 I 2 +…+X kn I n Thereby obtaining a matrix operation result of
I L And X L All belong to a full real number domain, k is more than or equal to 1 and less than or equal to m, and j is more than or equal to 1 and less than or equal to n;
FIG. 5 illustrates the process of extending a 4 × 4 order operator to an 8 × 8 order operator; the input vector and the output vector are both 8 multiplied by 1 order vectors, and the transmission matrix is an 8 multiplied by 8 order matrix; to split the operation into 4 × 4 order, to match the hardware of the experiment, the input and output vectors are respectively blocked into 2 4 × 1 vectors; meanwhile, the transmission matrix is divided into 4 × 4 matrices. From the principle of matrix blocking, it is clear that the following formula can be obtained
The problem of dimension expansion of the matrix is simplified into 4 times of multiplication of optical matrix vector products and 2 times of electric addition, and theoretical results and experimental results of the multiplication are drawn in a three-dimensional bar graph as shown in figure 5. In fig. 5, a histogram is used to show a real set of experimental data. After four operations are decomposed according to a matrix method, four output vector experimental values P = X are respectively obtained by measurement 11 I 1 ，Q＝X 21 I 1 ，R＝X 12 I 2 ，S＝X 22 I 2 (ii) a Collate the four measurements intoThe final output, i.e.
Then there is O L ＝[-0.43,-1.87,0.13,-0.98,-2.65,-0.66,-0.15,-2.4] T The order number of the matrix is expanded from 4 multiplied by 4 to 8 multiplied by 8, so that the expansion of the matrix operation scale is realized.
To analyze the reliability of the system in mass data calculations, 576 sets of input data and transfer matrices were loaded to obtain statistical results, as shown in FIG. 6; fig. 6 (a) is a transmission matrix value obtained by measurement; fig. 6 (b) shows the corresponding theoretical transmission matrix value; FIG. 6 (c) shows the output value and error value of the ith row vector product obtained by the experiment, where the abscissa of the data point in the graph is the data point number and the ordinate represents the vector product of the input vector and the ith row of the transfer matrix
Fig. 6 (d) shows the distribution of absolute values of errors. From the results shown in FIG. 6, it can be seen that the absolute value of the error is more than half in the range of 0 to 0.1, and falls mostly in the range of 0 to 0.2. Therefore, the processor provided by the embodiment still has higher calculation precision in a large number of operations.
Example 3:
a signal transformation method based on the complex matrix operation method provided in embodiment 2 is one of the following three typical signal transformation methods: discrete Walsh-Hadamard transform (WHT), discrete Cosine Transform (DCT), and Discrete fourier transform.
The results of three exemplary signal transformation methods are shown in fig. 7, (a) is the original signal of WHT transformation, and (b) is the original signal of even symmetric DCT transformation; (c) DCT transformed original signal being half of the preceding sequence; (d) is the DFT-transformed original signal; (e) results and theoretical values of WHT transform; (f) is the result and theoretical value of even symmetric DCT transform; (g) The result and theoretical values of the DCT transform for half of the preceding sequence; (h) is the result of the DFT transform and the theoretical value.
As can be seen from the results shown in fig. 7, the present embodiment can correctly realize the three typical signal conversions.
Generally speaking, the invention provides a photoelectric hybrid incoherent large-scale matrix operation processor, which divides an input vector and a transmission matrix into a real part and an imaginary part in a complex value matrix decomposition mode, calculates a matrix vector product in an optical domain, and performs addition and subtraction operation in an electric domain to realize complex value matrix operation, thereby expanding the operation domain of the incoherent matrix operation processor. And the dimension expansion problem of the matrix is equivalent to the combination of a plurality of times of optical matrix vector product multiplication and electricity addition in a matrix blocking mode, and the matrix operation is expanded to a larger scale under the condition of hardware with limited scale. By introducing the wavelength division multiplexing technology, the multi-wavelength parallel operation can be realized, the transmission matrix of any scale can be realized in principle, and the expansion capability on hardware is realized.
It will be understood by those skilled in the art that the foregoing is only a preferred embodiment of the present invention, and is not intended to limit the invention, and that any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the scope of the present invention.
Claims (10)
1. An optoelectronic hybrid incoherent complex valued matrix arithmetic processor, comprising:
an electronic control unit (5) for generating a vector I carrying N dimensions 0 The microwave signal to be detected; n is a positive integer;
a data input unit (1) connected with the electronic control unit (5) and used for receiving N paths of optical signals with different wavelengths and converting the vector I 0 Respectively coding the N-dimensional information into the intensity of the N paths of optical signals to obtain N paths of optical carrier microwave signals;
the wavelength division multiplexing unit (2) is used for converging the N paths of optical carrier microwave signals into a path of mixed optical signal;
an optical complex matrix arithmetic unit (3) connected to the electronic control unit (5) and comprising an adjustable resonant device comprising M rows and N columnsThe harmonic array is used for dividing the mixed optical signal into M paths of signals with equal power and respectively coupling the signals into M rows of resonant devices; the resonance wavelengths of the N columns of resonance devices are respectively aligned with the wavelengths of the N optical signals, and the electronic control unit (5) is also used for aligning the matrix X of M multiplied by N 0 Encoding into the transmission coefficient of each resonant device, thereby obtaining the final O 0 ＝X 0 I 0 Performing matrix operation; m is a positive integer;
the data acquisition unit (4) is connected with the electronic control unit (5) and is used for acquiring the optical power output by the through end and the downloading end of each row of resonance devices, carrying out differential operation, converting the differential operation result into an electric signal as a matrix operation result and transmitting the electric signal to the electronic control unit (5); the electronic control unit (5) is also used for storing the matrix operation result and performing addition and subtraction operation on the stored matrix operation result.
2. The optoelectronic hybrid incoherent complex valued matrix arithmetic processor of claim 1, wherein the resonant devices in the tunable array are microring resonators.
3. The optoelectronic hybrid incoherent complex valued matrix arithmetic processor of claim 1, wherein the resonant devices in the tunable array are tunable MZI assisted microring resonators.
4. The optoelectronic hybrid incoherent complex valued matrix arithmetic processor of claim 1, characterized in that the data acquisition unit (4) comprises M balanced photodetectors, corresponding to M rows of resonant devices; each balance photoelectric detector is connected with the electronic control unit (5);
the balance photoelectric detector is used for detecting the output optical power of the straight-through end and the downloading end of the corresponding line of the resonant device, and converting the optical power into an electric signal after differential operation.
5. The optoelectronic hybrid incoherent complex valued matrix arithmetic processor of any of claims 1 to 4, characterized in that each device in the optical complex valued matrix arithmetic unit (3) is integrated on a silicon-based chip.
6. A complex-valued matrix operation method based on the optoelectric hybrid incoherent complex-valued matrix operation processor of any one of claims 1 to 5, comprising:
non-negative real number domain matrix operation step: will carry the vector I of N dimension by means of the electronic control unit (5) 0 Is input to the data input unit (1) and a matrix X of size mxn is applied 0 Encoding the data into transmission coefficients of the resonant devices to output a matrix operation result O by the data acquisition unit 0 ＝X 0 I 0 ；I 0 And X 0 All belong to the non-negative real number domain.
7. The complex-valued matrix operation method of claim 6, further comprising:
and (3) performing full real number domain matrix operation: splitting an N-dimensional vector I into N-dimensional vectors I containing all positive elements + And an N-dimensional vector I containing the absolute values of all negative elements - Respectively calculating P = XI by using the non-negative number domain matrix operation step + And Q = XI - Calculating a matrix operation result O = XI = P-Q through differential operation; i belongs to the real number domain, X belongs to the non-negative real number domain, and the scale of X is M N.
8. The complex-valued matrix operation method of claim 7, further comprising:
a full complex field matrix operation step: vector I of N dimension P Splitting into real parts real (I) P ) And imaginary part imag (I) P ) Dividing the matrix X of M × N P Split into real parts real (X) P ) And imaginary imag (X) P ) Respectively calculating real (X) through the non-negative real number domain matrix operation step P )real(I P )、imag(X P )imag(I P )、imag(X P )real(I P ) And real (X) P )imag(I P ) And respectively obtaining matrix operation results O by addition and subtraction P ＝X P I P Real part of (O) P )＝real(X P )real(I P )-imag(X P )imag(I P ) And imaginary part imag (O) P )＝real(X P )imag(I P )+imag(X P )real(I P ) Thereby obtaining a matrix operation result of O P ＝real(O P )+i*imag(O P )；I P And X P All belong to the full real number domain.
9. The method of complex valued matrix operation of claim 8 further comprising:
large-scale matrix operation steps: will vector I L Splitting the block with the dimension of N to obtain
Will matrix X L Splitting the blocks with the size of M multiplied by N to obtain
Respectively calculating X through the full complex field matrix operation step kj I j Then, a matrix operation result O is obtained by addition operation L ＝X L I L The k-th block in (a) is O L (k)＝X k1 I 1 +X k2 I 2 +…+X kn I n Thereby obtaining a matrix operation result of
I L And X L All belong to a full real number domain, k is more than or equal to 1 and less than or equal to m, and j is more than or equal to 1 and less than or equal to n.
10. The complex-valued matrix operation method of claim 9, used to implement a discrete walsh-hadamard transform, a discrete cosine transform, or a discrete fourier transform.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111193964.9A CN113960718B (en) | 2021-10-13 | 2021-10-13 | Photoelectric hybrid incoherent complex value matrix operation processor and complex value matrix operation method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111193964.9A CN113960718B (en) | 2021-10-13 | 2021-10-13 | Photoelectric hybrid incoherent complex value matrix operation processor and complex value matrix operation method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113960718A CN113960718A (en) | 2022-01-21 |
| CN113960718B true CN113960718B (en) | 2022-10-14 |
Family
ID=79464447
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111193964.9A Active CN113960718B (en) | 2021-10-13 | 2021-10-13 | Photoelectric hybrid incoherent complex value matrix operation processor and complex value matrix operation method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113960718B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2023162857A1 (en) * | 2022-02-24 | 2023-08-31 | 国立大学法人東京大学 | Optical integrated circuit, information processing device, and computation method |
| CN114815959B (en) * | 2022-06-27 | 2022-11-01 | 之江实验室 | Photon tensor calculation acceleration method and device based on wavelength division multiplexing |
| CN117787369A (en) * | 2024-02-27 | 2024-03-29 | 苏州元脑智能科技有限公司 | Optical computing system, complex value detection method and data processing method |
Family Cites Families (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103678258B (en) * | 2013-12-25 | 2017-01-25 | 中国科学院半导体研究所 | Method for improving data resolution ratio of silica-based optical matrix processor |
| US10644808B2 (en) * | 2017-08-23 | 2020-05-05 | Seagate Technology Llc | Silicon photonics based optical network |
| CN109639359B (en) * | 2019-01-07 | 2021-10-08 | 上海交通大学 | Photonic neural network convolution layer chip based on micro-ring resonator |
| CN110501779B (en) * | 2019-08-26 | 2020-04-28 | 南京航空航天大学 | Micro-ring delay matrix and microwave photon integrated multi-beam phased array chip and system |
| CN113267911B (en) * | 2021-05-12 | 2022-05-27 | 华中科技大学 | Large-capacity photoelectric hybrid programmable logic operation chip |
-
2021
- 2021-10-13 CN CN202111193964.9A patent/CN113960718B/en active Active
Also Published As
| Publication number | Publication date |
|---|---|
| CN113960718A (en) | 2022-01-21 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113960718B (en) | Photoelectric hybrid incoherent complex value matrix operation processor and complex value matrix operation method | |
| TWI735886B (en) | Computing system | |
| CN109639359B (en) | Photonic neural network convolution layer chip based on micro-ring resonator | |
| Bandyopadhyay et al. | Single chip photonic deep neural network with accelerated training | |
| CN112912900A (en) | Photoelectric computing system | |
| KR20220039775A (en) | Systems and Methods for Analog Computation Using a Linear Photonic Processor | |
| TWI806042B (en) | Optoelectronic processing apparatus, system and method | |
| CN113496281A (en) | Photoelectric computing system | |
| Peserico et al. | Integrated photonic tensor processing unit for a matrix multiply: a review | |
| Ma et al. | High-density integrated photonic tensor processing unit with a matrix multiply compiler | |
| CN113961035A (en) | Deep learning-oriented high-efficiency reconfigurable all-optical neural network computing chip architecture | |
| WO2023051550A1 (en) | Programmable high-dimensional quantum computing chip structure based on integrated optics | |
| CN113570051A (en) | Photoelectric processing system | |
| CN110908428A (en) | Parallel optical computing system for efficiently realizing large-scale matrix operation | |
| CN111561953A (en) | On-chip integrated optical matrix vector multiplier based on wavelength division multiplexing and balanced detection | |
| Zhu et al. | Dota: A dynamically-operated photonic tensor core for energy-efficient transformer accelerator | |
| CN115905792A (en) | Mach-Zehnder interferometer network for optical real number matrix calculation | |
| US20220337333A1 (en) | Scalable, Ultra-Low-Latency Photonic Tensor Processor | |
| CN114488650A (en) | Silicon-based photonic integrated chip | |
| WO2022060908A1 (en) | Frequency multiplexed photonic neural networks | |
| Li et al. | A wavelength-multiplexed photonic tensor processor based on Mach-Zehnder modulator | |
| TWI831932B (en) | Hybrid analog-digital processors and method performed therein | |
| Zhang et al. | Time-stretch optical neural network with time-division multiplexing | |
| Tian et al. | Photonic Neural Networks with Kramers–Kronig Activation | |
| CN115641494A (en) | Neural network image processing system based on micro-ring modulator |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |