CN115374828A - Tensor convolution kernel acceleration chip and method - Google Patents
Tensor convolution kernel acceleration chip and method Download PDFInfo
- Publication number
- CN115374828A CN115374828A CN202211091673.3A CN202211091673A CN115374828A CN 115374828 A CN115374828 A CN 115374828A CN 202211091673 A CN202211091673 A CN 202211091673A CN 115374828 A CN115374828 A CN 115374828A
- Authority
- CN
- China
- Prior art keywords
- optical
- sub
- signals
- wavelength
- convolution kernel
- 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.)
- Pending
Links
Images
Landscapes
- Optical Communication System (AREA)
Abstract
The invention discloses a tensor convolution kernel acceleration chip which is suitable for all deep learning networks containing convolution operation. The invention integrates the wavelength division demultiplexer, the modulator, the delay weighting unit, the coupling array and the detector which finish the tensor convolution kernel accelerated operation through the photon integration technology. The method comprises the steps of loading signals to be processed on a plurality of optical carriers respectively based on a wavelength division multiplexing technology, realizing convolution kernel coefficient weighting and time interleaving of different carrier signals through a micro-ring and a delay line, and realizing summation operation after weighting through a coupling array. The invention takes light as an information carrier, and can realize tensor calculation acceleration of signals based on a plurality of delay weighting units and coupling arrays, thereby greatly improving the calculation rate and the energy efficiency ratio of a neural network.
Description
Technical Field
The invention relates to the technical field of photonic integration, in particular to a tensor convolution kernel acceleration chip and a tensor convolution kernel acceleration method.
Background
Superimposing multidimensional data into tensors provides us with the opportunity to discover intrinsic structural features hidden in the data that are not visible in bi-directional (matrix) data analysis. For example, a multi-path representation of electroencephalographic (EEG) data is an efficient way of neuroscience data processing, while tensors stacked across time, space, and spectrum are advantageous for detecting features in electromagnetic waveforms. Because the tensor is matched with the high-dimensional nature of the world, the concept of multipath analysis generates a wide range of signal processing methods in the fields of life sciences, radars, data mining, machine learning and the like. In the elementary operation of the tensor, the convolution can effectively extract structural features from the data, and when the convolution kernel traverses the tensor, the target features are filtered out. Convolutional neural networks, a miniature of modern Artificial Intelligence (AI), are designed under the concept of multi-channel tensor processing. Given that tensor processing, particularly in the field of artificial intelligence, is consuming more and more computing resources, there is a pressing need for high throughput and power efficient processors. Photonics has recently proven to be a promising candidate for constructing high performance matrix processors. By designing the photonic circuit as a linear transformation function, matrix multiplication can be done as the light passes through the circuit. The broadband spectrum of photonic circuits increases the clock frequency to tens of gigahertz. Thus, photonic circuits have proven to be excellent GeMM processors with High throughput and energy efficiency (see [ Xu, shaofu, et al. "" High-order transducer flow processing integrated photonic circuits. "" arXiv preprint xiv:2112.12322 (2021) ]). Indeed, another advantage of photonics over electronics is the abundance of available degrees of freedom for light. For example, linear transformations with wavelength, guided mode, time and space have been successfully investigated.
Disclosure of Invention
The technical problem to be solved by the invention is as follows: the method overcomes the defects of the prior art, realizes tensor calculation of tensor signals in an optical domain by utilizing a delay weighting unit comprising a delay waveguide and a coupling array, solves the problems of limited calculation power and large power consumption caused by calculation separation and data dimension conversion in the traditional electronic technology, can flexibly expand tensor core, and is suitable for multidimensional data tensor calculation. Except for the light source, all the photonic components of the whole acceleration chip are integrated on one chip, the system is compact and simple, small in size and low in cost, and the convolution kernel matrix can be flexibly expanded.
In order to achieve the purpose, the invention provides the following technical scheme:
the application discloses a tensor convolution kernel acceleration chip which is integrated by a wavelength division demultiplexer, a modulator, a delay weighting unit, a coupling array and a detector;
the optical input end of the wavelength division demultiplexer is the optical input end of the chip and is used for receiving multi-wavelength optical signals; the output end of the wavelength division multiplexer is connected with the modulator and used for outputting the sub optical signals with N wavelengths;
the optical input end of the modulator is connected with the optical output end of the wavelength division demultiplexer and used for receiving sub optical signals containing N wavelengths; the electrical input end of the modulator is used for receiving a signal to be convolved; the optical output end of the modulator is connected with the delay weighting unit and is used for outputting sub-modulation optical signals containing N wavelengths;
the optical input end of the delay weighting unit is connected with the optical output end of the modulator and is used for receiving sub-modulation optical signals containing N wavelengths; the electrical input end of the delay weighting unit is used for receiving a convolution kernel matrix control signal; the optical output end of the delay weighting unit is connected with the coupling array and is used for outputting an amplitude weighting sub-modulation optical signal;
the optical input end of the coupling array is connected with the optical output end of the delay weighting unit and is used for receiving the amplitude weighted sub-modulation optical signal; the electrical input end of the coupling array is used for receiving a coupling coefficient control signal; the output end of the coupling array is connected with the detector and used for outputting a secondary amplitude weighting sub-modulation optical signal containing N wavelengths;
the optical input end of the detector is connected with the optical output end of the coupling array and is used for receiving a secondary amplitude weighted sub-modulation optical signal containing N wavelengths; and the optical output end of the detector is used for outputting a characteristic signal obtained after the convolution operation of the signal to be convolved is completed.
A tensor convolution kernel acceleration chip as recited in claim 1, wherein: the delay weighting unit consists of a straight-through waveguide, a coupling waveguide and N micro-ring resonators; and the coupling waveguide output end of the first micro-ring resonator is the optical output end of the delay weighting unit.
Preferably, the coupling array is composed of a plurality of groups of sub-arrays, and each sub-array is composed of a plurality of couplers and a wavelength division multiplexer; the optical input end of the coupler of the first group of sub-arrays is connected with the optical output end of the delay weighting unit to be used as the optical input end of the coupling array; the optical output end of the coupler of each group of subarrays is respectively connected with the optical input end of the corresponding wavelength division multiplexer and the optical input end of the coupler of the next group of subarrays; and the optical output end of the wavelength division multiplexer is the optical output end of the coupling array.
Preferably, the N micro-ring resonators in the delay weighting unit are provided with delay waveguides between the through waveguide ends.
Preferably, the chip is integrated based on a III-V material integration process or a silicon-based integration process.
Preferably, the radii of the N micro-ring resonators in the delay weighting unit are sequentially increased, and the free spectral range corresponding to the micro-ring resonator with the largest radius is larger than the spectral range occupied by the multi-wavelength optical signal.
The invention also discloses a tensor convolution kernel acceleration method, which comprises the following steps:
s1, sending the multi-wavelength optical signal to a wavelength division multiplexer, wherein the wavelength division multiplexer divides the multi-wavelength optical signal into M sub-optical signals containing N wavelengths and sends the sub-optical signals to M modulators;
s2, the modulator receives M signals to be convolved, and M sub-optical signals each containing N wavelengths are modulated to obtain M sub-modulation optical signals each containing N wavelengths; respectively sending the signals to M delay weighting units;
s3, the delay weighting unit receives the convolution kernel matrix control signal, controls the coupling coefficients of the N micro-ring resonators of the delay weighting unit according to the convolution kernel matrix control signal, and sequentially couples the M sub-modulation optical signals each containing N wavelengths into the coupling waveguide according to different coupling coefficients to obtain M amplitude-weighted sub-modulation optical signals; and sent into the coupling array;
s4, the coupling arrays share O groups of sub-arrays; the coupling array receives the coupling coefficient control signal, controls the M amplitude weighted sub-modulation optical signals through the coupling coefficient control signal to realize secondary amplitude weighting, obtains OM secondary amplitude weighted sub-modulation optical signals containing N wavelengths, and respectively sends the OM secondary amplitude weighted sub-modulation optical signals containing N wavelengths to the corresponding O wavelength division multiplexers; the wavelength division multiplexer combines OM second-level amplitude weighted sub-modulation optical signals containing N wavelengths into O weighted modulation multi-wavelength optical signals containing MN wavelengths; respectively sending O weighted modulation multi-wavelength optical signals containing MN wavelengths to O corresponding detectors;
and S5, performing photoelectric conversion on the O weighted modulation multi-wavelength optical signals containing MN wavelengths by the detector to obtain electric output signals, namely characteristic signals obtained after tensor convolution operation is completed on the signals to be convolved.
Preferably, the signals to be convolved in step S2 are obtained by tensor decomposition, and the tensor is obtained by inputting the dimensionality of the signals to be processed and the number of the signals to be processed; the signal to be processed is a one-dimensional signal obtained by flattening a one-dimensional signal or an actual two-dimensional signal.
Preferably, the convolution kernel matrix control signal in step S3 implements the weighting of the convolution kernel matrix coefficients of the MN wavelength modulation signals by controlling the coupling coefficients of the N micro-ring resonators in each delay weighting unit, specifically: and determining the coupling coefficient of the micro-ring resonator according to the size of the corresponding convolution kernel matrix coefficient and the initial signal intensity of each wavelength in the multi-wavelength optical signal, changing the coupling coefficient of the micro-ring resonator through a thermo-optical effect or an electro-optical effect, wherein N micro-ring resonators in each delay weighting unit correspond to one sub-optical signal.
The invention has the beneficial effects that:
1) The invention realizes tensor calculation of tensor data by combining multiple dimensions of wavelength, time and space in an optical domain based on the characteristic that photons can be obtained in parallel, and can effectively avoid the problems of calculation complexity increase and power consumption caused by electric domain calculation separation and multi-dimensional data conversion.
2) Compared with the general discrete photoelectric device, the monolithic photonic integration of all functional components does not need additional photoelectric functional devices, thereby simplifying the system, improving the stability of the system and expanding the scale of the chip in a large range.
3) The invention realizes the convolution kernel convolution accelerated calculation of data based on a plurality of delay weighting units embedded into the cascade delay waveguide, has simple and efficient scheme, is compact in system and strong in anti-interference compared with the system based on the optical fiber dispersion technology, and does not need dispersion calibration compensation.
4) The method realizes the control of the convolution kernel matrix coefficient based on a plurality of delay weighting units, can realize the quick update of the convolution kernel matrix coefficient in tensor calculation, and is suitable for real-time data processing application.
The features and advantages of the present invention will be described in detail by embodiments in conjunction with the accompanying drawings.
Drawings
FIG. 1 is a diagram of a tensor convolution kernel acceleration chip according to the present invention.
Fig. 2 is a schematic structural diagram of an embodiment of a tensor convolution kernel acceleration chip according to the present invention.
FIG. 3 is a diagram illustrating a spectrum distribution of a multi-wavelength optical signal according to an embodiment;
FIG. 4 is a spectrum distribution diagram of an output signal of a first delay weighting unit in the embodiment;
FIG. 5 is a spectrum distribution diagram of an output signal of a second delay weighting unit in the embodiment;
FIG. 6 is a diagram showing a spectrum distribution of an output signal of the Mth delay weighted micro-ring unit in the embodiment;
FIG. 7 is a diagram of the time series of the output optical signals of the first wavelength division multiplexer versus the wavelength in the embodiment;
FIG. 8 is a diagram showing a time series of output optical signals of the second wavelength division multiplexer in relation to wavelengths in the embodiment;
fig. 9 is a time series and wavelength relation diagram of the output optical signal of the O-th wavelength division multiplexer in the embodiment.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood, however, that the description herein of specific embodiments is only intended to illustrate the invention and not to limit the scope of the invention. Moreover, in the following description, descriptions of well-known structures and techniques are omitted so as to not unnecessarily obscure the concepts of the present invention.
As shown in fig. 1, the tensor convolution kernel acceleration chip of the present invention includes: a wavelength division multiplexer (DWDM), a modulator, a delay weighting unit, a coupling array and a detector (PD); all the photonic components are connected through optical waveguides; the wavelength division multiplexer is provided with 1 optical input end and M optical output ends, the optical input end is the optical input end of the whole chip and is used for receiving external multi-wavelength optical signals, and the M optical output ends are connected with the optical input ends of the corresponding modulators;
the number of the modulators is M, each modulator is provided with 1 electrical input end, 1 optical input end and 1 optical output end, the optical output ends are connected with the optical input ends of the corresponding delay weighting units, and the electrical input ends are used for receiving external signals to be convolved; the number of the delay weighting units is M, each delay weighting unit is composed of 1 through waveguide, 1 coupling waveguide and N micro-ring resonators, and the coupling waveguide output end of the first micro-ring resonator of each delay weighting unit is the optical output end of each delay weighting unit; the coupling array is provided with O groups of couplers and O wavelength division multiplexers, each coupler is provided with 1 optical input end and 2 optical output ends, wherein the optical input ends of the first group of couplers are connected with the optical output ends of the delay weighting units and are the optical input ends of the coupling array; 2 optical output ends of each group of couplers are respectively connected with the optical input ends of the next group of couplers and the optical input ends of the corresponding wavelength division multiplexers, and each wavelength division multiplexer is provided with M optical input ends and 1 optical output end; the number of the detectors is O, each detector is provided with 1 optical input end which is connected with the optical output end of the corresponding wavelength division multiplexer;
the working process of the chip is as follows: firstly, a multi-wavelength optical signal is sent to a wavelength division multiplexer, the wavelength division multiplexer divides the multi-wavelength optical signal into M sub-optical signals each containing N wavelengths and sends the sub-optical signals to M modulators, and M signals to be convolved are loaded on the corresponding sub-optical signals through the corresponding modulators respectively to obtain M sub-modulated optical signals; m sub-modulation optical signals are respectively sent into corresponding M delay weighting units, and the control signals control the coupling coefficients of MN micro-rings in the M delay weighting units based on M groups of convolution kernel matrix coefficients to realize the amplitude weighting of the MN wavelength signals; the M sub-modulation optical signals after amplitude weighting are sent to a coupling array, and the M sub-modulation optical signals after amplitude weighting are controlled by a coupling coefficient control signal to realize secondary amplitude weighting, so that OM secondary amplitude weighting sub-modulation optical signals containing N wavelengths are obtained; respectively sending OM secondary amplitude weighted sub-modulation optical signals to corresponding O wavelength division multiplexers, and combining the OM secondary amplitude weighted sub-modulation optical signals into O weighted modulation multi-wavelength optical signals containing MN wavelengths by the wavelength division multiplexers; and respectively sending the O weighted modulation multi-wavelength optical signals to O corresponding detectors to carry out photoelectric conversion to obtain O electric signals, wherein the O electric signals are characteristic signals obtained after the signals to be convolved complete tensor convolution operation.
The convolution kernel matrix control signal is MN convolution kernel matrix control signals generated based on M groups of convolution kernel matrix coefficients, the convolution kernel matrix control signals change corresponding micro-ring resonator coupling coefficients through a thermo-optical effect or an electro-optical effect, each micro-ring resonator coupling coefficient is determined according to the size of the convolution kernel matrix coefficient and the initial signal intensity of each wavelength in the multi-wavelength optical signal, the MN wavelength signals in the multi-wavelength optical signal are equal in amplitude or unequal, M and N are positive integers, the M and N are the number of signals to be processed and the number of convolution kernel matrix coefficients in each group, and the maximum number of the signals to be processed and the maximum number of the convolution kernel matrix coefficients can be supported are preferred. Preferably, the radiuses of MN micro-ring resonators in the time delay weighting micro-ring array are sequentially increased and respectively correspond to one resonant wavelength, and the free spectral range Δ f corresponding to the micro-ring resonator with the largest radius FSR The spectrum range MN Δ f occupied by the multi-wavelength optical signal is larger than.
The coupling coefficient control signal is used for controlling the coupling coefficients of the OM couplers, respectively realizing the amplitude regulation of N sub-modulation optical signals in the M paths of weighted modulation multi-wavelength optical signals and realizing the second-stage amplitude weighting.
The channel interval of the wavelength demultiplexer is NΔ f, wherein Δ f is the frequency interval between two wavelengths in the multi-wavelength optical signal, so that the multi-wavelength optical signal is correspondingly divided into M sub-optical signals each containing N wavelengths.
The photonic components such as the wavelength division multiplexer, the modulator, the delay weighting unit, the coupling array and the detector and the optical waveguide can be integrated by three-five material or silicon preparation, and the chip can be integrated based on mature processes such as three-five material integration process or silicon-based integration process.
To facilitate understanding of the public, the technical solution of the present invention is further described in detail by a specific embodiment.
The multi-wavelength light source is generated by a multi-wavelength laser, a laser array, a mode-locked laser, a femtosecond laser, an optical frequency comb generator, or an optical soliton optical frequency comb generator, and in this embodiment, the mode-locked laser is preferred and amplified by an optical amplifier. The convolution kernel matrix control signal and the coupling coefficient control signal are sent by a tensor kernel control signal. The characteristic signals are collected and processed by the signal collecting and processing unit. In addition, the amplitudes of the MN wavelength signals in the output signal of the multi-wavelength light source may be equal or unequal, and the present embodiment is preferably equal.
As shown in fig. 2, the tensor convolution kernel acceleration chip of the present embodiment includes: the system comprises 1 mode-locked laser, 1 optical amplifier, 1 wavelength division demultiplexer, M modulators, 1 signal to be convolved, M delay weighting units, 1 tensor kernel matrix control single signal, 1 coupling array (consisting of OM couplers and O wavelength division multiplexers), O detectors, 1 signal acquisition and processing unit and the like.
First, the mode-locked laser outputs multi-wavelength optical signals with equal wavelength intensities and amplifies the signals by an optical amplifier, and the wavelength intensities of the amplified multi-wavelength optical signals can be represented by a = [ a, …, a ] in a matrix] T MN The spectral distribution is shown in FIG. 3, where M and N are positive integers and are the maximum supportable positions respectivelyThe number of the physical signals and the number of the maximum supportable convolution kernel matrix coefficients of each group, A is the single-wavelength signal intensity. The amplified constant-amplitude multi-wavelength optical signal is sent to a wavelength division multiplexer to be divided into M sub-multi-wavelength optical signals each containing N wavelengths, and the sub-multi-wavelength optical signals can be represented as A m =[A,A,A,…,A] T N ，A m Representing the mth sub multi-wavelength optical signal, M =1,2, …, M. M modulators are in one-to-one correspondence with M output ports of the wavelength division demultiplexer, M sub multi-wavelength optical signals are respectively sent into the M modulators, M signals to be processed output by a signal source to be processed are respectively subjected to intensity modulation on the sub multi-wavelength optical signals through the corresponding modulators, and the M signals to be processed are loaded on the corresponding sub multi-wavelength optical signals to obtain M sub multi-wavelength modulated optical signals. The signal sequence to be processed can be denoted x m (n)=[x m (1), x m (2), x m (3),…, x m (P)]Where n denotes the discretized time index, x m And (n) represents an mth signal sequence to be processed, P is the length of the signal to be processed, the signal to be processed is a one-dimensional signal obtained by flattening a one-dimensional signal or an actual two-dimensional signal, and the flattening operation is to convert the two-dimensional matrix into the one-dimensional matrix. The dimensionalities of M signals to be processed output by the signal source to be processed can be expressed as [ D ] by tensor data , Sin ]In which D is data For inputting the dimension of the signal to be processed, for a one-dimensional signal, D data Indicating the number of signal data; for two-dimensional signals, D data =[W,H]W and H represent the number of data of width and height of the two-dimensional signal, S in For the number of signals to be processed, here S in And (c) = M. Sub-multi-wavelength modulated optical signal S Mod_m The matrix can be expressed as:
m sub multi-wavelength modulation optical signals are sequentially coupled to M delay weighting units by an optical fiber waveguide coupling technology, the structural schematic diagram of the delay weighting micro-ring unit is shown in figure 3, and each delay weighting micro-ring unit consists of 1 straight-through waveguide and 1 straight-through waveguideA coupling waveguide and N micro-ring resonators having a length of Δ l = c Δ t/N between the ends of the through waveguide w In which n is w Is the effective refractive index of the waveguide delay line, Δ t =1/S M For a single symbol duration of the signal to be processed, SM is the symbol rate of the signal to be processed. The resonance characteristics of the N micro-rings in each delay weighting unit correspond to one wavelength in sequence. The convolution kernel matrix control signal output by the tensor kernel control signal firstly controls the resonance characteristic of the first micro-ring resonator, so that the corresponding wavelength sub-intensity modulated optical signal transmitted in the through waveguide is coupled into the coupling waveguide according to a specific coupling coefficient, and the coupling coefficient is set according to the size of the convolution kernel matrix coefficient, thereby realizing the weighting of the convolution kernel matrix coefficient. And the sub multi-wavelength modulation optical signals in the through waveguide enter the delay waveguide of the through waveguide after passing through the first micro-ring resonator to realize the Δ t delay. And the delayed multi-wavelength intensity modulation optical signals realize coefficient weighting on corresponding wavelength signals through a second micro-ring resonator, and all wavelength signal weighting is completed in sequence after delay. And obtaining M sub multi-wavelength weighted modulation optical signals at the output end of the coupling waveguide.
Let the mth set of convolution kernel matrix coefficients be W mN =[w m1 ,w m2, w m3, …,w mN ] T Let D be the dimension of the convolution kernel k Denotes that D when the convolution kernel is one-dimensional k Is the number N of coefficients of a one-dimensional convolution kernel, and when the convolution kernel is two-dimensional, D k =[C,L]C and L are the number of coefficients of the two-dimensional convolution kernel row and column, respectively, and C · L = N. Sub-multi-wavelength weighted modulation optical signal S output by delay weighted micro-ring unit coupling waveguide Modcon_m Can be expressed as:
the sub multi-wavelength weighted modulation optical signal spectrograms are shown in fig. 4, fig. 5 and fig. 6, and it should be noted that, for the convenience of understanding of the public, fig. 4, fig. 5 and fig. 6 correspond to the sub multi-wavelength weighted modulation optical signals respectively output by the 1 st, 2 nd, m delay weighted micro-ring units. Based on the waveguide fiber coupling technology, the M sub-weighted intensity modulated optical signals are coupled into the fiber and then input into the coupling array. The coupling array comprises OM couplers and O wavelength division multiplexers, coupling coefficients of the couplers are controlled based on coupling coefficient control signals output by tensor kernel matrix signals, amplitude adjustment of M sub-modulation optical signals in each path of weighted modulation multi-wavelength optical signals is achieved respectively, and secondary amplitude weighting is achieved. Coupling coefficient M of OM couplers in coupling array con Can be represented by a matrix as:
v is the coupling coefficient corresponding to each coupler, and the OM secondary amplitude weighted sub-modulated optical signals are combined into O weighted modulated multi-wavelength optical signals containing MN wavelengths by O wavelength division multiplexers. Dimension D for setting dimension of coupled array m =[S Min , S Mout ]Indicates the number of valid input ports for the coupled array and the number of output ports of the wavelength division multiplexer, here S Min = M, here S Mout And (c) = O. The time series of the weighted modulated multi-wavelength optical signal versus wavelength is shown in fig. 7, 8, and 9. It should be noted that fig. 7 corresponds to a time series and wavelength relationship of the weighted modulated multi-wavelength optical signal output by the first wavelength division multiplexer, fig. 8 corresponds to a time series and wavelength relationship of the weighted modulated multi-wavelength optical signal output by the second wavelength division multiplexer, and fig. 9 corresponds to a time series and wavelength relationship of the weighted modulated multi-wavelength optical signal output by the O-th wavelength division multiplexer. And respectively sending the O-path weighted and modulated multi-wavelength optical signals to O detectors to complete photoelectric conversion to obtain electric output signals. The signal within the valid timing sequence of the electrical output signal may be represented as:
wherein S is ca_o(r) The result of the r-th tensor calculation for the o-th detector. After the acquisition processing unit acquires the O electrical output signals, the acquisition processing unit is effectiveAnd processing the sequence signals to obtain O tensor calculation results. That is, the dimension of data after tensor acceleration operation can be expressed as [ D ] by tensor data , Sout ]In which D is data For the output signal dimension, the same as the input signal dimension to be processed, S out For the number of signals to be processed, here S out And (c) = O. The dimensionality of the overall tensor convolution kernel can be expressed as [ D k , ,D m , S in , S out ]。
The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents or improvements made within the spirit and principle of the present invention should be included in the scope of the present invention.
Claims (9)
1. A tensor convolution kernel acceleration chip is characterized in that: the chip is integrated by a wavelength division demultiplexer, a modulator, a delay weighting unit, a coupling array and a detector;
the optical input end of the wavelength division demultiplexer is the optical input end of the chip and is used for receiving multi-wavelength optical signals; the output end of the wavelength division demultiplexer is connected with the modulator and used for outputting sub optical signals containing N wavelengths;
the optical input end of the modulator is connected with the optical output end of the wavelength division demultiplexer and is used for receiving sub optical signals containing N wavelengths; the electrical input end of the modulator is used for receiving a signal to be convolved; the optical output end of the modulator is connected with the delay weighting unit and is used for outputting sub-modulation optical signals containing N wavelengths;
the optical input end of the delay weighting unit is connected with the optical output end of the modulator and is used for receiving sub-modulation optical signals containing N wavelengths; the electrical input end of the delay weighting unit is used for receiving a convolution kernel matrix control signal; the optical output end of the delay weighting unit is connected with the coupling array and is used for outputting an amplitude weighting sub-modulation optical signal;
the optical input end of the coupling array is connected with the optical output end of the delay weighting unit and is used for receiving the amplitude weighting sub-modulation optical signal; the electrical input end of the coupling array is used for receiving a coupling coefficient control signal; the output end of the coupling array is connected with the detector and used for outputting a secondary amplitude weighted sub-modulation optical signal containing N wavelengths;
the optical input end of the detector is connected with the optical output end of the coupling array and is used for receiving a secondary amplitude weighted sub-modulation optical signal containing N wavelengths; and the optical output end of the detector is used for outputting a characteristic signal obtained after the convolution operation of the signal to be convolved is completed.
2. A tensor convolution kernel acceleration chip as recited in claim 1, wherein: the delay weighting unit consists of a straight-through waveguide, a coupling waveguide and N micro-ring resonators; and the coupling waveguide output end of the first micro-ring resonator is the optical output end of the delay weighting unit.
3. A tensor convolution kernel acceleration chip as recited in claim 1, wherein: the coupling array consists of a plurality of groups of sub-arrays, and each sub-array consists of a plurality of couplers and a wavelength division multiplexer; the optical input end of the coupler of the first group of sub-arrays is connected with the optical output end of the delay weighting unit to be used as the optical input end of the coupling array; the optical output end of the coupler of each group of subarrays is respectively connected with the optical input end of the corresponding wavelength division multiplexer and the optical input end of the coupler of the next group of subarrays; and the optical output end of the wavelength division multiplexer is the optical output end of the coupling array.
4. A tensor convolution kernel acceleration chip as recited in claim 2, wherein: and delay waveguides are arranged between the N micro-ring resonators in the delay weighting unit at the ends of the through waveguides.
5. A tensor convolution kernel acceleration chip as recited in claim 1, wherein: the chip is integrated based on a III-V material integration process or a silicon-based integration process.
6. A tensor convolution kernel acceleration chip as recited in claim 2, wherein: the radiuses of the N micro-ring resonators in the delay weighting unit are sequentially increased, and the free spectral range corresponding to the micro-ring resonator with the largest radius is larger than the spectral range occupied by the multi-wavelength optical signal.
7. A tensor convolution kernel acceleration method is characterized by comprising the following steps:
s1, sending the multi-wavelength optical signal to a wavelength division multiplexer, wherein the wavelength division multiplexer divides the multi-wavelength optical signal into M sub-optical signals containing N wavelengths and sends the sub-optical signals to M modulators;
s2, the modulator receives M signals to be convolved, and M sub-optical signals each containing N wavelengths are modulated to obtain M sub-modulation optical signals each containing N wavelengths; respectively sending the signals to M delay weighting units;
s3, the delay weighting unit receives the convolution kernel matrix control signal, controls the coupling coefficients of the N micro-ring resonators of the delay weighting unit according to the convolution kernel matrix control signal, and sequentially couples the M sub-modulation optical signals each containing N wavelengths into the coupling waveguide according to different coupling coefficients to obtain M amplitude-weighted sub-modulation optical signals; and sent into the coupling array;
s4, the coupling arrays share O groups of sub-arrays; the coupling array receives the coupling coefficient control signal, controls the M amplitude weighted sub-modulation optical signals to realize secondary amplitude weighting through the coupling coefficient control signal, obtains OM secondary amplitude weighted sub-modulation optical signals containing N wavelengths, and respectively sends the OM secondary amplitude weighted sub-modulation optical signals containing N wavelengths to the corresponding O wavelength division multiplexers; the wavelength division multiplexer combines OM second-level amplitude weighted sub-modulation optical signals containing N wavelengths into O weighted modulation multi-wavelength optical signals containing MN wavelengths; respectively sending O weighted modulation multi-wavelength optical signals containing MN wavelengths to O corresponding detectors;
and S5, performing photoelectric conversion on the O weighted modulation multi-wavelength optical signals containing MN wavelengths by the detector to obtain electric output signals, namely characteristic signals obtained after tensor convolution operation is completed on the signals to be convolved.
8. A tensor convolution kernel acceleration method as recited in claim 7, wherein: in the step S2, the signals to be convolved are respectively obtained by tensor decomposition, and the tensor is obtained by inputting the dimensionality of the signals to be processed and calculating the number of the signals to be processed; the signal to be processed is a one-dimensional signal obtained by flattening a one-dimensional signal or an actual two-dimensional signal.
9. A tensor convolution kernel acceleration method as recited in claim 7, wherein: in the step S3, the convolution kernel matrix control signal realizes the weighting of the convolution kernel matrix coefficients of the MN wavelength modulation signals by controlling the coupling coefficients of the N micro-ring resonators in each delay weighting unit, which specifically includes: and determining the coupling coefficient of the micro-ring resonator according to the size of the corresponding convolution kernel matrix coefficient and the initial signal intensity of each wavelength in the multi-wavelength optical signal, changing the coupling coefficient of the micro-ring resonator through a thermo-optical effect or an electro-optical effect, wherein N micro-ring resonators in each delay weighting unit correspond to one sub-optical signal.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211091673.3A CN115374828A (en) | 2022-09-07 | 2022-09-07 | Tensor convolution kernel acceleration chip and method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211091673.3A CN115374828A (en) | 2022-09-07 | 2022-09-07 | Tensor convolution kernel acceleration chip and method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115374828A true CN115374828A (en) | 2022-11-22 |
Family
ID=84071738
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211091673.3A Pending CN115374828A (en) | 2022-09-07 | 2022-09-07 | Tensor convolution kernel acceleration chip and method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115374828A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117784313A (en) * | 2024-02-28 | 2024-03-29 | 之江实验室 | Two-dimensional photon convolution operation chip and system based on cyclic array waveguide grating |
-
2022
- 2022-09-07 CN CN202211091673.3A patent/CN115374828A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117784313A (en) * | 2024-02-28 | 2024-03-29 | 之江实验室 | Two-dimensional photon convolution operation chip and system based on cyclic array waveguide grating |
| CN117784313B (en) * | 2024-02-28 | 2024-06-07 | 之江实验室 | Two-dimensional photon convolution operation chip and system based on cyclic array waveguide grating |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN114358271B (en) | Time-wavelength interweaving photon neural network convolution acceleration chip | |
| CN114815959B (en) | Photon tensor calculation acceleration method and device based on wavelength division multiplexing | |
| CN114819132B (en) | Photon two-dimensional convolution acceleration method and system based on time-wavelength interleaving | |
| CN109459816B (en) | Silicon-based optical arbitrary waveform generation chip | |
| CN115169542B (en) | Two-dimensional photon convolution acceleration system and device for convolution neural network | |
| CN115167610B (en) | Two-dimensional photon convolution acceleration chip and application system and method | |
| CN117077750B (en) | Parallel photon convolution operation chip and system | |
| CN115374828A (en) | Tensor convolution kernel acceleration chip and method | |
| CN116432726B (en) | Photoelectric hybrid deep neural network operation device and operation method | |
| CN114157391A (en) | Beam forming device and beam forming method thereof | |
| CN117784313B (en) | Two-dimensional photon convolution operation chip and system based on cyclic array waveguide grating | |
| CN115567118A (en) | Silicon-based photon terahertz wireless communication transmitting front-end chip, system and method | |
| US11934943B1 (en) | Two-dimensional photonic neural network convolutional acceleration chip based on series connection structure | |
| CN116484931B (en) | Photon matrix multiplication operation device and operation method for neural network | |
| CN116029350B (en) | Two-dimensional photon coherent convolution acceleration chip based on time interleaving and application system thereof | |
| CN116540353A (en) | Array waveguide grating system and demodulation method for demodulating fiber gratings | |
| You et al. | All-optical photonic signal processors with negative coefficients | |
| Xu et al. | Single photonic perceptron based on a soliton crystal Kerr microcomb for high-speed, scalable, optical neural networks | |
| CN117709423B (en) | Deep neural network photon acceleration chip and operation system thereof | |
| Xu et al. | High-speed optical neural networks based on microcombs | |
| CN113541809B (en) | Multi-band adjustable photo-generated millimeter wave generation system | |
| Moss | Laser and Photonics Reviews photonic single perceptron based on time-wavelength multiplexing with a microcomb for optical neural networks | |
| CN116107037B (en) | Optical computing network structure of optical device on chip of dense waveguide array | |
| CN116739065A (en) | Photon tensor convolution calculation method and system for multichannel data processing | |
| CN115906977A (en) | Full gloss convolver based on two combs |
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 |