A compact and efficient photonic convolutional neural network accelerator based on a dual-function microring resonator

Abstract

The invention provides a compact and efficient photon convolutional neural network accelerator based on a dual-function microring resonator, comprising: a plurality of tiles, each tile communicating with an optical router; the tile comprising: an optical matrix vector multiplication module and a nonlinear module; the optical matrix vector multiplication module comprising: an input signal module, an optical matrix vector operation module and an output signal module; the input signal module comprising: a laser array, a wavelength division multiplexer and an optical splitter; the laser array generates n continuous optical signals of different wavelengths; the wavelength division multiplexer multiplexes the n continuous optical signals of different wavelengths into the same waveguide to obtain a multiplexed optical signal; the optical splitter performs optical splitting on the multiplexed optical signal to obtain k-way optical splitting signals; the optical matrix vector operation module modulates the input optical splitting signal through an embedded PN junction, and converts the k-way modulated optical signal into an electrical signal through an added GST analog weight; the nonlinear module performs nonlinear operation on the k electrical signals to obtain a calculation result.

Description

Technical Field

[0001] The present invention belongs to the field of optical signal processing, and in particular relates to a compact and efficient photonic convolutional neural network accelerator based on a dual-function microring resonator. Background Art

[0002] As a key technology, convolutional neural network (CNN) has been widely used in many fields such as image recognition and speech processing. CNN is a computationally intensive network. As the scale of CNN models continues to increase, traditional electronic CNN accelerators will face more and more challenges due to the limitations of clock frequency, latency, power consumption and other factors. In contrast, photonic CNN accelerators have shown great potential in terms of computing rate, parallelism, ultra-low latency and high energy efficiency. Researchers have proposed some different optical CNN accelerators, which are mainly based on two optical devices, namely Mach–Zehnder Interferometer (MZI) and Microring Resonator (MR). Among them, the structure based on MZI occupies a large area, making it difficult to achieve large-scale integration and limiting the scale of operation. MR has the advantages of compact footprint and support for wavelength division multiplexing (WDM), showing superior computing performance.

[0003] In MR-based optical accelerators, most of them use electrically controlled MR to simulate weights, and a continuous bias voltage needs to be applied to MR to maintain the weight state. However, in the process of CNN reasoning, the weights of the convolution kernel do not need to be updated frequently after training. Frequent weight data reading and weight state maintenance will cause a lot of energy consumption. Therefore, the non-volatile phase change material Ge2Sb2Te5 (GST) can be used to form a storage array to realize parallel photon memory computing, simplifying the calculation to measure the light transmission of reconfigurable and non-resonant passive components, which can significantly reduce the data transmission overhead. On the other hand, the current MR-based optical CNN accelerator usually uses an independent set of MR as a modulator array in the input part to simulate the image pixel value, and then uses another set of MR arrays based on electrical modulation or phase change materials to simulate the weight. The total number of MRs required for this structure is the sum of the weight MR array and the input modulation MR array, and the area cost is relatively large. At the same time, MR has a thermo-optical effect and is easily affected by the ambient temperature and causes the resonant wavelength to shift. The increase in the number of MRs also increases the possibility of calculation errors in the system. Summary of the invention

[0004] In order to solve the problems existing in the background technology, the present invention proposes a compact and efficient photonic convolutional neural network accelerator based on a dual-function microring resonator based on considerations such as area overhead, energy consumption and computing rate. The accelerator integrates the non-volatile phase change material GST on a microring resonator embedded in a PN junction to form a dual-function photonic computing engine that can simultaneously perform input wavelength power amplitude modulation and weight simulation. In addition, a matrix multiplier supporting wavelength division multiplexing is built based on the dual-function photonic computing engine. By simultaneously inputting data of different wavelengths as independent signal carriers into the matrix for multiplication and addition operations, the scale of parallel computing is increased. The present invention is expected to realize a large-scale, low-power, and low-hardware area overhead convolutional neural network accelerator to meet the future large-scale convolution operations The demand for high-performance computing hardware.

[0005] In order to achieve the above technical objectives, the present invention provides the following technical solutions: a compact and efficient photonic convolutional neural network accelerator based on a dual-function microring resonator, comprising: a plurality of tiles, each tile communicating with an optical router;

[0006] The tile includes: an optical matrix vector multiplication module and a nonlinear module;

[0007] The optical matrix vector multiplication module includes: an input signal module, an optical matrix vector operation module and an output signal module; wherein the input signal module includes: a laser array, a wavelength division multiplexer and a splitter;

[0008] The laser array is used to generate n continuous optical signals of different wavelengths;

[0009] The wavelength division multiplexer is used to multiplex n continuous optical signals of different wavelengths into the same waveguide to obtain a multiplexed optical signal;

[0010] The optical splitter is used to split the multiplexed optical signal to obtain k-way split optical signals;

[0011] The optical matrix vector operation module includes: k dual-function photon computing engines, each dual-function photon computing engine includes: n dual-function micro-ring resonators with added GST and embedded PN junctions;

[0012] The optical matrix vector operation module modulates the input k-way split light signals respectively through k dual-function photon computing engines to obtain k-way modulated optical signals; wherein the input split light signals are modulated through the embedded PN junctions, and the input split light signals are multiplied with the weights while being modulated through the added GST analog weights;

[0013] The output signal module comprises: k photoelectric detectors, the k photoelectric detectors are respectively used to convert k modulated optical signals into electrical signals;

[0014] The nonlinear module is used to perform nonlinear operations on k electrical signals to obtain calculation results.

[0015] Preferably, the laser array comprises: n continuous wave lasers; different continuous wave lasers generate continuous light signals of different wavelengths, and the n continuous wave lasers generate n continuous light signals of different wavelengths.

[0016] Preferably, the dual-function microring resonator comprises: two straight waveguides and a ring waveguide arranged on a substrate; the two straight waveguides are parallel to each other and correspond to each other front and back; the ring waveguide is arranged between the two front and rear straight waveguides, and has coupling regions with the two straight waveguides respectively; a P-type section is arranged in the circumferential inner region of the ring waveguide; an N-type section is arranged on the left or right side of the ring waveguide along the circumference of the ring waveguide; a GST layer is arranged on the upper surface of the ring waveguide along the circumference of the ring waveguide; the two straight waveguides of the n dual-function microring resonators of the dual-function photonic computing engine are cascaded in the horizontal direction respectively.

[0017] Preferably, the width of the GST layer is the same as the width of the annular waveguide; the length of the GST layer is 0.01 to 0.1 of the circumference of the annular waveguide.

[0018] Preferably, the two straight waveguides of the dual-function microring resonator are used to transmit split light signals having multiple wavelengths; the resonance conditions of the n dual-function microring resonators correspond to the light signals of n wavelengths in the split light signals respectively; when a light signal of a certain wavelength in the split light signals meets the resonance condition, the light signal of this wavelength is coupled to the dual-function microring resonator in the coupling region for modulation.

[0019] Preferably, different voltages are applied to the PN junctions of the n dual-function micro-ring resonators according to the input information of the convolutional neural network so that the dual-function micro-ring resonators perform corresponding power amplitude modulation on the input optical signal.

[0020] Preferably, the crystallinity state of the GST is changed by applying an external light pulse to the GST layer to achieve reading, writing, storing or erasing operations on the weights.

[0021] Preferably, the nonlinear module uses a nonlinear function to perform nonlinear operation on the electrical signal output by the output signal module to obtain a calculation result.

[0022] Preferably, the compact and efficient photonic convolutional neural network accelerator also includes: an analog-to-digital converter, a digital-to-analog converter, an input buffer, an output buffer and a dynamic random access memory; the analog-to-digital converter is used to convert the calculation result obtained by the nonlinear operation into a digital signal, the output buffer is used to cache the digital signal converted by the analog-to-digital converter, and the dynamic random access memory is used to store the signal cached by the output buffer; the input buffer is used to read the input signal from the dynamic random access memory; the digital-to-analog converter is used to convert the input signal read by the input buffer into an analog signal, and apply a voltage to the PN junction of the dual-function microring resonator.

[0023] The present invention has at least the following beneficial effects

[0024] The present invention applies photonic technology to the design of computing architecture, uses optical computing instead of electrical computing, and combines wavelength division multiplexing technology to effectively increase the number of channels of the optical link, realize parallel processing of optical signals, and provide faster computing speed and higher energy efficiency than traditional electrical neural network accelerators; integrates non-volatile phase change material GST on a micro-ring resonator embedded with a PN junction to form a dual-function photonic computing engine, which can simultaneously perform input wavelength power amplitude modulation and weight simulation, realizing "in-memory computing" of input and weight; applying different voltages to the PN junction on each dual-function micro-ring resonator can realize the input optical signal Power amplitude modulation, without the need to add an additional set of micro-ring resonator arrays to modulate the input wavelength power amplitude, reduces the number of micro-rings required under the same computing scale, and greatly reduces the hardware area overhead; using the non-volatility of GST and the high contrast between the amorphous state and the crystalline state, the weights are mapped to each GST to achieve the simulation and storage of the weights, avoiding the power loss caused by the continuous external power supply; in addition, the introduction of wavelength division multiplexing technology, based on the dual-function photon computing engine to build a matrix multiplier, can simultaneously input data of different wavelengths into the matrix for multiplication and addition operations, thereby increasing the scale of parallel computing. The accelerator architecture proposed in the present invention is committed to achieving low-energy, high-density photon computing to meet the needs of future large-scale image recognition tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 The photonic convolutional neural network accelerator architecture of the present invention;

[0026] Figure 2 Schematic diagram of the structure of a dual-function microring resonator with GST and PN junction added;

[0027] Figure 3 Schematic diagram of the optical matrix-vector multiplication module structure. DETAILED DESCRIPTION

[0028] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0029] Among them, the drawings are only used for illustrative explanations, and they only represent schematic diagrams rather than actual pictures, and should not be understood as limitations on the present invention. In order to better illustrate the embodiments of the present invention, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0030] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right", "front", "rear", etc. indicate the orientation or position relationship, they are based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation. Therefore, the terms describing the position relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0031] See also Figure 1 ,The present invention provides a compact and efficient photonic convolutional neural network accelerator based on a dual-function micro-ring resonator, comprising: a plurality of tiles, each tile communicating with an optical router;

[0032] The tile includes: an optical matrix vector multiplication module and a nonlinear module;

[0033] The optical matrix vector multiplication module includes: an input signal module, an optical matrix vector operation module and an output signal module; wherein the input signal module includes: a laser array, a wavelength division multiplexer and a splitter;

[0034] The laser array is used to generate n continuous optical signals of different wavelengths;

[0035] The wavelength division multiplexer is used to multiplex n continuous optical signals of different wavelengths into the same waveguide to obtain a multiplexed optical signal;

[0036] The optical splitter is used to split the multiplexed optical signal to obtain k-way split optical signals;

[0037] The optical matrix vector operation module includes: k dual-function photon computing engines, each dual-function photon computing engine includes: n dual-function micro-ring resonators with added GST and embedded PN junctions;

[0038] The optical matrix vector operation module modulates the input k-way split light signals respectively through k dual-function photon computing engines to obtain k-way modulated optical signals; wherein the input split light signals are modulated through the embedded PN junctions, and the input split light signals are multiplied with the weights while being modulated through the added GST analog weights;

[0039] The output signal module comprises: k photoelectric detectors, the k photoelectric detectors are respectively used to convert k modulated optical signals into electrical signals;

[0040] The nonlinear module is used to perform nonlinear operations on k electrical signals to obtain calculation results.

[0041] In this embodiment, Figure 1 This is a schematic diagram of the structure of the convolutional neural network accelerator proposed in the present invention, in which any tile communicates with the optical router; each tile can be used to calculate part of the computing tasks of a single CNN neuron, and can also be used to calculate the computing tasks of multiple CNN neurons, and the calculation results between multiple tiles are transmitted through the optical router; the nonlinear module is used for nonlinear operations, and the optical matrix-vector multiplication module is used to perform convolution operations on optical signals to simulate the computing principle of the convolution layer; photonic technology is applied to the design of computing architecture, optical computing is used instead of electrical computing, and wavelength division multiplexing technology is combined to effectively increase the number of channels of the optical link, realize parallel processing of optical signals, and provide faster computing speed and higher energy efficiency than traditional electrical neural network accelerators.

[0042] The optical splitter is a 1-to-k optical splitter, where k is an integer greater than or equal to 3, and n is an integer greater than or equal to 3.

[0043] Preferably, the laser array comprises: n continuous wave lasers; different continuous wave lasers generate continuous light signals of different wavelengths, and the n continuous wave lasers generate n continuous light signals of different wavelengths.

[0044] In this embodiment, the laser array includes a plurality of continuous wave lasers, and different lasers generate continuous optical signals of different wavelengths, from λ1 to λ n For example, multiple optical signals of different wavelengths generated by a laser array are multiplexed into the same waveguide through a wavelength division multiplexer, and the multiplexed signal of the same waveguide is passed through an optical splitter to generate a multi-path split signal.

[0045] See also Figure 2 Preferably, the dual-function microring resonator comprises: two straight waveguides and a ring waveguide arranged on a substrate; the two straight waveguides are parallel to each other and correspond to each other front and back; the ring waveguide is arranged between the two front and rear straight waveguides, and has coupling regions with the two straight waveguides respectively; a P-type section is arranged in the inner circumferential region of the ring waveguide; an N-type section is arranged on the left or right side of the ring waveguide along the circumference of the ring waveguide; a GST layer is arranged on the upper surface of the ring waveguide along the circumference of the ring waveguide; the two straight waveguides of the n dual-function microring resonators of the dual-function photonic computing engine are cascaded in the horizontal direction respectively.

[0046] The present invention provides an implementation manner, wherein an N-type section is arranged on the left side of the annular waveguide along the circumference of the annular waveguide; and a GST layer is arranged on the upper surface of the right end of the annular waveguide along the circumference of the annular waveguide.

[0047] Figure 2 The schematic diagram of the dual-function microring resonator. Figure 2 (a) is an up-and-down path type microring resonator used to realize a dual-function computing engine; it is composed of two straight waveguides and one ring waveguide, and the microring resonator contains a PN junction and a section of non-volatile phase change material GST; when the input optical signal meets the resonance condition of the microring resonator, the ring waveguide resonant cavity is in a resonant state, and the input optical signal is completely output from the "download" port, that is, the transmittance of the "download" end is 1, and the transmittance of the "through" end is 0; when the ring waveguide resonant cavity is detuned, the output power value of the "download" end will decrease, and the output power of the "through" end will increase; the present invention is provided with a P-type node on the inner side of the ring waveguide; an N-type node is provided on the left or right side of the ring waveguide; a PN node is formed; applying different voltages to the PN junction can realize power amplitude modulation of the input optical signal, and different weight simulations are realized by changing the crystallinity of GST.

[0048] Preferably, the width of the GST layer is the same as the width of the ring waveguide; the length of the GST layer is 0.01 to 0.1 of the circumference of the ring waveguide. Only one embodiment of the present invention is given here. For those skilled in the art, it is possible to adjust the above parameters in a targeted manner through a limited number of experiments on the premise of understanding this solution, which all fall within the protection scope of the present invention.

[0049] Preferably, the thickness of the GST layer is 5 to 20 nm. In this embodiment, the substrate includes a Si (silicon) bottom layer and a SiO2 buried oxide layer stacked in sequence; wherein, a ring waveguide and a straight waveguide are arranged on the SiO2 buried oxide layer, and both the straight waveguide and the ring waveguide include: a Si waveguide layer and a ridge arranged on the Si waveguide layer; the width of the ridge (made of Si (silicon)) is equal to the width of the GST layer; the thickness of the SiO2 buried oxide layer is 3.3 μm; the thickness of the Si waveguide layer is 70 nm; the thickness of the ridge is 150 nm; the width of the GST layer is 450 nm; the thickness of the GST layer is 10 nm. Only one embodiment of the present invention is given here. For those skilled in the art, it is possible to adjust the above parameters in a targeted manner through limited experiments on the premise of understanding this scheme, which all belong to the protection scope of the present invention.

[0050] Preferably, the two straight waveguides of the dual-function microring resonator are used to transmit split light signals having multiple wavelengths; the resonance conditions of the n dual-function microring resonators correspond to the light signals of n wavelengths in the split light signals respectively; when a light signal of a certain wavelength in the split light signals meets the resonance condition, the light signal of this wavelength is coupled to the dual-function microring resonator in the coupling region for modulation.

[0051] Among them, the resonance condition of the microring resonator is:

[0052] 2πRn eff =mλ0

[0053] Where R is the radius of the microring resonator ring waveguide, n eff is the effective refractive index of the waveguide, m is the resonance order, and λ0 is the resonant wavelength of the microring resonator.

[0054] As mentioned above, applying different voltages to the PN junction can achieve power amplitude modulation of the input optical signal. The specific principle is that the main material Si of the microring resonator has a plasma dispersion effect, and the refractive index and absorption coefficient of Si change with the carrier concentration to meet the following relationship:

[0055]

[0056] Where Δn is the change in refractive index, Δα is the change in absorption coefficient, e is the electron charge, λ is the input light wavelength, c is the propagation speed of light in vacuum, ε0 is the free space dielectric constant, n is the refractive index of Si material, ΔN e and ΔN h Respectively represent the changes in free electron concentration and hole concentration, μ e and μ h are the mobilities of electrons and holes, respectively. and Represent the effective mass of electrons and holes respectively; Existing research shows that the empirical formula for the plasma dispersion effect of Si material at 1550nm is:

[0057] Δn=Δn e +ΔN h =-8.8×10 -22 ΔN e -8.5×10 -18 ΔN h 0.8

[0058] Δα=Δα e +Δα h =8.5×10 -18 ΔN e +6×10 -18 ΔN h

[0059] Therefore, by applying different voltages to the PN junction, the free carrier concentration in the silicon material can be changed, thereby changing the effective refractive index of Si, and realizing power amplitude modulation of the input optical signal;

[0060] Preferably, different voltages are applied to the PN junctions of the n dual-function micro-ring resonators according to the input information of the convolutional neural network so that the dual-function micro-ring resonators perform corresponding power amplitude modulation on the input optical signal.

[0061] Preferably, the crystallinity state of the GST is changed by applying an external light pulse to the GST layer to achieve reading, writing, storing or erasing operations on the weights.

[0062] As mentioned above, the specific principle of using non-volatile phase change material GST for weight simulation is:

[0063] GST is a non-volatile phase change material with high contrast between crystalline and amorphous states. When GST is in the crystalline state, it has a strong absorption effect, and the light in the waveguide is completely absorbed by GST. At this time, the transmittance of the waveguide is 0. When GST is in the amorphous state, it almost does not absorb light signals, and almost all the light in the waveguide is output from the waveguide. At this time, the transmittance of the waveguide is 1. When GST is between the crystalline and amorphous states, the waveguide transmittance is between 0 and 1, so GST can arbitrarily realize weight simulation between 0 and 1. Multiple crystallinity levels can be set for GST according to the crystallinity formula. The specific formula is as follows:

[0064]

[0065] Where p is the crystallinity of GST, ε eff (p) is the effective refractive index of GST when the crystallinity is p, ε c and ε aThey represent the dielectric constants of GST in the crystalline and amorphous states respectively; the real and imaginary parts of the effective refractive index of GST with different degrees of crystallization are different, and the corresponding phase and attenuation coefficients are also different. The calculation formula is as follows:

[0066]

[0067] In the formula, θ and η are the attenuation factor and phase factor respectively, n ef,wg and k eff,wg are the real and imaginary parts of the effective refractive index of the waveguide, L GST is the length of GST, n eff,GST and k eff,GST are the real and imaginary parts of the effective refractive index of GST respectively; R is the radius of the ring waveguide.

[0068] Based on the above principle, assuming that the input power in the waveguide is i after being modulated by the PN junction, and the weight represented by the GST is w, then the output power of the optical signal after passing through the PN junction and the GST is o=i×w.

[0069] Figure 2 (b) shows a schematic diagram of the cross-section of a waveguide integrated with GST. It can be seen that GST is integrated on the ring waveguide and is the same width as the ring waveguide, so no additional area is added. As a non-volatile phase change material, after the shape of GST is fixed, the information stored inside it can be preserved for decades. In addition, GST is easy to read, write, and erase. Only an external light pulse needs to be applied to modify or erase the weight value stored inside. In this case, the specific implementation plan is:

[0070] Write operation: A high-power input signal is injected into the input end, and the evanescent wave coupling between the light in the waveguide and the GST is used to absorb the input signal into the GST, thereby increasing the temperature of the GST surface. When the temperature is greater than the crystal threshold, the state of the GST begins to change (i.e., the amorphization process), affecting the transmittance of the waveguide, thereby realizing the writing of the transmittance (weight) value.

[0071] Storage operation: After the GST transmittance value is successfully written, the GST is quickly cooled to room temperature, so that the atomic state structure of the GST can be fixed and the weight can be stored. At room temperature, this state can remain unchanged for decades, so GST is non-volatile.

[0072] Read operation: A low-power input signal of magnitude A is input at the input end. Its energy is lower than the crystal threshold of GST and will not change the crystal state of GST. At this time, the signal is transmitted to the output end through GST. Assuming that the weight stored in GST is B, the signal power C received at the output end is the product of the input signal power A and the GST weight B, so the weight can be read.

[0073] Erase operation: Inputting a high-power input pulse signal can make GST transition from amorphous state to crystalline state, thereby erasing the weight data.

[0074] Figure 3 It is a schematic diagram of the structure of an optical matrix-vector multiplication module. The schematic diagram of the architecture includes an input signal module, an optical matrix-vector calculation module and an output signal module. The input signal module is composed of a laser array, a wavelength division multiplexer and a splitter. Figure 3 The laser array in the n ) input optical signals, which are multiplexed by the wavelength division multiplexer and transmitted to the optical splitter along the waveguide. The optical splitter splits the optical signals multiplexed by the wavelength division multiplexer, and the split optical signals are transmitted to the optical matrix vector operation module; the optical matrix vector operation module includes: k dual-function photon computing engines, each dual-function photon computing engine includes: n dual-function micro-ring resonators with added GST and embedded PN junctions; the PN junction embedded in the dual-function micro-ring resonator can modulate the input optical signal to change its power amplitude, and the GST can simulate and store weights, so that the input optical signal can be multiplied with the weight while performing power amplitude modulation. The optical signal after the multiplication operation is output through the download end of the micro-ring resonator and coupled to the same waveguide, that is, an addition operation is performed.

[0075] The output signal module is composed of several photodetectors, which are photodiodes. The photodiodes are used to convert the optical signal after the multiplication and addition operation of the optical matrix vector operation module into an electrical signal. The signal is then transmitted to the nonlinear module, and the output result is generated through the nonlinear action of activation, pooling and fully connected layers.

[0076] Preferably, the nonlinear module uses a nonlinear function to perform nonlinear operation on the electrical signal output by the output signal module to obtain a calculation result.

[0077] Preferably, the compact and efficient photonic convolutional neural network accelerator also includes: an analog-to-digital converter, a digital-to-analog converter, an input buffer, an output buffer and a dynamic random access memory; the analog-to-digital converter is used to convert the calculation result obtained by the nonlinear operation into a digital signal, the output buffer is used to cache the digital signal converted by the analog-to-digital converter, and the dynamic random access memory is used to store the signal cached by the output buffer; the input buffer is used to read the input signal from the dynamic random access memory; the digital-to-analog converter is used to convert the input signal read by the input buffer into an analog signal, and apply a voltage to the PN junction of the dual-function microring resonator.

[0078] The above embodiment illustrates the working process of the photonic convolutional neural network based on a dual-function microring resonator. CNN is a computationally intensive network, in which convolution operations occupy 86% to 94% of the execution time of CNN. Therefore, it is necessary to design a CNN accelerator to improve computing efficiency, reduce energy consumption and optimize the hardware architecture. The photonic technology of the present invention is applied to the design of computing architecture, using optical computing instead of electrical computing, and combining wavelength division multiplexing technology to effectively increase the number of channels of the optical link, realize parallel processing of optical signals, and provide faster computing speed and higher energy efficiency than traditional electrical neural network accelerators. Secondly, the non-volatile phase change material GST is integrated on the microring resonator embedded in the PN junction to form a dual-function photonic computing engine, which can simultaneously perform input wavelength power amplitude modulation and weight simulation, and realize the "in-memory computing" of input and weight. Such a design reduces the number of microrings required under the same computing scale, greatly reduces the hardware area overhead, and at the same time, the non-volatility of GST can avoid the power loss caused by continuous external power supply. The accelerator architecture proposed in the present invention is committed to achieving low-energy, high-density photon computing, and is expected to be applied in unmanned driving, aerospace, multi-bit image processing, biomedicine and other fields.

[0079] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

[0080] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A compact and efficient photonic convolutional neural network accelerator based on a dual-function microring resonator, characterized in that: include: Multiple tiles, each tile communicates with an optical router; The tile includes: an optical matrix vector multiplication module and a nonlinear module; The optical matrix vector multiplication module includes: an input signal module, an optical matrix vector operation module and an output signal module; wherein the input signal module includes: a laser array, a wavelength division multiplexer and a splitter; The laser array is used to generate n continuous optical signals of different wavelengths; The wavelength division multiplexer is used to multiplex n continuous optical signals of different wavelengths into the same waveguide to obtain a multiplexed optical signal; The optical splitter is used to split the multiplexed optical signal to obtain k-way split optical signals; The optical matrix vector operation module includes: k dual-function photon computing engines, each dual-function photon computing engine includes: n dual-function micro-ring resonators with added GST and embedded PN junctions; The optical matrix vector operation module modulates the input k-way split light signals respectively through k dual-function photon computing engines to obtain k-way modulated optical signals; wherein the input split light signals are modulated through the embedded PN junctions, and the input split light signals are multiplied with the weights while being modulated through the added GST analog weights; The dual-function micro-ring resonator comprises: two straight waveguides and a ring waveguide arranged on a substrate; the two straight waveguides are parallel to each other and correspond to each other front and back; the ring waveguide is arranged between the front and rear straight waveguides, and has coupling regions with the two straight waveguides respectively; a P-type section is arranged in the inner circumference area of ​​the ring waveguide; an N-type section is arranged on the left or right side of the ring waveguide along the circumference of the ring waveguide; a GST layer is arranged on the upper surface of the ring waveguide along the circumference of the ring waveguide; the two straight waveguides of the n dual-function micro-ring resonators of the dual-function photon computing engine are cascaded in the horizontal direction respectively; According to the input information of the convolutional neural network, different voltages are applied to the PN junctions of n dual-function micro-ring resonators respectively so that the dual-function micro-ring resonators perform corresponding power amplitude modulation on the input optical signal; The output signal module comprises: k photoelectric detectors, the k photoelectric detectors are respectively used to convert k modulated optical signals into electrical signals; The nonlinear module is used to perform nonlinear operations on k electrical signals to obtain calculation results.

2. A compact and efficient photonic convolutional neural network accelerator based on a dual-function microring resonator according to claim 1, characterized in that: The laser array comprises: n continuous wave lasers; different continuous wave lasers generate continuous light signals of different wavelengths, and n continuous wave lasers generate n continuous light signals of different wavelengths.

3. A compact and efficient photonic convolutional neural network accelerator based on a dual-function microring resonator according to claim 1, characterized in that: The width of the GST layer is the same as that of the annular waveguide; the length of the GST layer is 0.01 to 0.1 of the circumference of the annular waveguide.

4. A compact and efficient photonic convolutional neural network accelerator based on a dual-function microring resonator according to claim 1, characterized in that: The two straight waveguides of the dual-function microring resonator are used to transmit split light signals with multiple wavelengths; the resonance conditions of the n dual-function microring resonators correspond to the light signals of n wavelengths in the split light signals respectively; When an optical signal of a certain wavelength in the split optical signal meets the resonance condition, the optical signal of this wavelength is coupled to the dual-function microring resonator in the coupling region for modulation.

5. A compact and efficient photonic convolutional neural network accelerator based on a dual-function microring resonator according to claim 1, characterized in that: By applying external light pulses to the GST layer, the crystallinity state of GST is changed to achieve reading, writing, storage or erasing of the weights.

6. A compact and efficient photonic convolutional neural network accelerator based on a dual-function microring resonator according to claim 1, characterized in that: The nonlinear module uses a nonlinear function to perform nonlinear operations on the electrical signal output by the output signal module to obtain a calculation result.

7. A compact and efficient photonic convolutional neural network accelerator based on a dual-function microring resonator according to claim 1, characterized in that: The compact and efficient photonic convolutional neural network accelerator also includes: an analog-to-digital converter, a digital-to-analog converter, an input buffer, an output buffer and a dynamic random access memory; the analog-to-digital converter is used to convert the calculation result obtained by the nonlinear operation into a digital signal, the output buffer is used to cache the digital signal converted by the analog-to-digital converter, and the dynamic random access memory is used to store the signal cached by the output buffer; the input buffer is used to read the input signal from the dynamic random access memory; the digital-to-analog converter is used to convert the input signal read by the input buffer into an analog signal, and apply a voltage to the PN junction of the dual-function microring resonator.

Images