High-flux low-energy-consumption optical convolution arithmetic unit based on on-chip multi-dimensional multiplexing and phase change material

Abstract

The invention discloses a high-throughput low-energy-consumption optical convolution arithmetic unit based on on-chip multi-dimensional multiplexing and a phase change material, and aims to solve the limitation of an existing optical neural network accelerator in the aspects of energy consumption, throughput and calculation efficiency. Conventional optical computing architectures generally employ single-dimensional multiplexing, which causes the scale of optical computing to be limited. Through the innovative wavelength and mode hybrid multiplexing technology, parallel processing of multiple computing channels is achieved in the same chip, the parallel computing scale is effectively increased, and the computing power is improved. Besides, the optical convolution arithmetic unit performs weight storage by using a non-volatile phase change material GSST, and can break through the limitation of von Neumann architecture to realize integration of optical storage and calculation, so that the power consumption is further reduced, and the calculation energy efficiency is improved.

Description

Technical Field

[0001] The present invention belongs to the field of optical signal processing, and in particular relates to a high-throughput, low-energy consumption optical convolution operator based on on-chip multi-dimensional multiplexing and phase change materials. Background Art

[0002] Artificial neural networks (ANNs) have become the core driving force in many rapidly developing technology fields, including computer vision, decision support, medical analysis, virtual reality, and autonomous driving. In the calculation process of ANNs, convolution operations account for more than 80% of the computing time and resources, and the core of its operation is multiplication and accumulation operations.

[0003] The traditional electrical convolution operation process relies on the von Neumann architecture design, which strictly separates the processor from the memory. This requires that a large amount of matrix and vector data be transferred from the memory to the processor for calculation and then the results are transferred back to the memory during the convolution operation. Frequent data movement causes significant delays and consumes a lot of power, which limits the overall computing speed and reduces the energy efficiency of the system. With the continuous growth in demand for high-performance neural networks, optical artificial neural networks (PANNs) are gradually becoming an effective supplement to traditional electronic ANNs with their significant advantages in high parallel processing capabilities, high speeds, and low power consumption.

[0004] Optical convolution computing devices, with their highly parallel characteristics, can not only significantly reduce the energy consumption of ANNs, but also greatly improve computing density and processing speed. By using non-volatile phase change materials (PCM) to store weights, optical convolution operators can achieve storage and computing integration to break through the von Neumann bottleneck. When the optical convolution operator calculates, PCM directly stores the weight value and participates in the calculation. The input optical signal passes through the PCM and interacts with the stored weight (modulates the light intensity or phase) to complete the convolution of the weight and the input without the need for additional electrical processing steps. In addition, the non-volatile characteristics of PCM allow weights to be stored for a long time without continuous power supply, and the storage itself consumes almost no energy, further reducing the energy consumption of the optical convolution operator.

[0005] At present, most of the schemes for designing on-chip optical convolution operators are based on wavelength division multiplexing (WDM) technology to achieve parallel computing. WDM technology uses optical signals of different wavelengths to transmit independent data, thereby performing parallel multiplication and addition (MAC) operations, effectively improving computing efficiency. However, WDM technology is limited by the available wavelength range and wavelength stability, and it is difficult to meet the expansion requirements of larger matrix operation scales. Therefore, the development of new parallel design schemes is crucial for optical convolution operators.

[0006] In this context, mode division multiplexing (MDM) technology has shown its outstanding potential. MDM technology significantly improves channel capacity by using different orders of optical modes for orthogonal signal transmission in a single bus waveguide, greatly enhances the parallel processing capability of optical convolution operators, and provides an innovative solution to overcome the scalability bottleneck of existing technologies.

[0007] Combining MDM and WDM technologies, by utilizing different optical modes and different light wavelengths, the optical convolution operator can process more data streams at the same time to achieve highly parallel matrix operations. This composite technology greatly improves the data processing capability and computing density of the optical convolution operator, and provides a technical foundation for large-scale matrix operations. Large-scale matrix operations are the core of the efficient operation of neural networks. It can not only significantly accelerate the training process of the model, but also perform reasoning tasks faster and more accurately. Therefore, the combination of MDM and WDM provides key support for fully realizing the computing potential of neural networks, enabling complex deep learning models to have higher performance and efficiency in practical applications.

[0008] CN115085854A discloses an optical matrix multiplier that supports storage and computing integration and wavelength-mode hybrid multiplexing. In this invention, non-volatile phase change materials GST and MR are used to form an optical dot product engine. Among them, MR is used to guide optical signals of different wavelengths, and the weights are stored in the GST integrated on the optical waveguide. In addition, a matrix multiplier for wavelength and mode hybrid multiplexing is built, and optical signals of different wavelengths and modes are simultaneously input into the matrix for calculation. However, this design is limited by the robustness of the high-order mode converter, the number of multiplexed modes is small, and the problem of excessive insertion loss caused by GST is not considered. Summary of the invention

[0009] The present invention aims to solve the problem of weight matrix operation scale in the current optical neural network accelerator based on wavelength division multiplexing (WDM) technology, and proposes a high-throughput, low-energy optical convolution operator structure based on on-chip multi-dimensional multiplexing and phase change materials. This optical convolution operator uses PCM to realize on-chip storage and computing integration, designs a highly robust mode converter based on sub-wavelength grating waveguide (SWG), and combines the hybrid multiplexing of mode and wavelength, breaking through the current scalability limitations of optical convolution operators in large-scale matrix processing, and further improving the parallel computing capabilities of optical convolution operators.

[0010] In order to achieve the above object, the technical solution of the present invention is as follows:

[0011] A high-throughput, low-energy optical convolution operator based on on-chip multi-dimensional multiplexing and phase-change materials, comprising an input module, a computing module and an output module, wherein the input module comprises a laser array, an N×1 wavelength division multiplexer, a 1×M first optical splitter, M×N modulators, an M×1 mode multiplexer, and a 1×P second optical splitter; N different wavelengths of light generated by the laser array enter the N×1 wavelength division multiplexer for multiplexing and integration, and then are input into the 1×M first optical splitter; the 1×M first optical splitter converts the N wavelengths of light into a single wavelength. The multiplexed optical power is divided into M equal parts to form M multi-wavelength multiplexing outputs; each of the M multi-wavelength multiplexing outputs passes through N optical modulators to realize the loading of the input signal. There are a total of M×N optical modulators, and finally an M×N input matrix is ​​formed; then the M optical signals loaded by the modulators are input to the M×1 mode multiplexer to be multiplexed into M different modes transmitted on the same waveguide; finally, the light after wavelength and mode multiplexing passes through the 1×P second optical splitter to enter the calculation module.

[0012] The computing module includes P layers of computing units, each layer of computing units includes a mode demultiplexer, M groups of ring waveguide arrays and a mode multiplexer; the mode demultiplexer is composed of a group of cascaded mode converters, and the group of cascaded mode converters converts TE M-1 The mode light to TE1 mode light is converted into TE0 mode light in order from high to low mode orders; each group of annular waveguide arrays is composed of N pairs of annular waveguides and weighted waveguides with PCM coupled between the annular waveguides, so as to realize the point multiplication function of the calculation unit; the mode multiplexer is composed of a group of cascaded mode converters, and the group of cascaded mode converters converts the TE0 mode light into TE1 to TE0 mode light in order from low to high mode orders in cascade order. M-1 mode light and outputs the wavelength mode multiplexed signal.

[0013] The output module outputs the wavelength and mode multiplexing output signals of the P layer computing units, and the P layer output signals are respectively O1, O2, O3, ..., O N express.

[0014] The laser array in the input module includes N lasers generating different wavelengths. The different wavelengths of light generated by these N lasers enter the N×1 wavelength division multiplexer for multiplexing. The mixed multi-wavelength optical signal after multiplexing is evenly divided into M equal parts by the 1×M first optical splitter and output independently, wherein each equal part includes N wavelengths. Subsequently, each of the M optical signals is independently modulated by N modulators for each wavelength, thereby completing the loading of the N×M input matrix.

[0015] The M optical signals loaded by the modulator in the input module are respectively connected to M-1 mode converters to realize the conversion from fundamental mode to TE M-1The mode conversion of the mode light to the TE1 mode light, one of which does not undergo mode conversion processing and is connected to the subsequent device as TE0 mode light. Specifically, the modulated M optical signals are connected to TE from left to right. M-1 In the TE1 mode converter, the mode order of the waveguide is converted from high to low from left to right. The waveguide at the far right does not undergo mode conversion and is connected to the subsequent device in TE0 mode.

[0016] Furthermore, the TE M-1 The TE0 mode optical signal is connected to the M×1 mode multiplexer, and M optical signals of different modes are multiplexed into the same waveguide. The multiplexed optical signal enters the 1×P second optical splitter so that the computing units of the P layer have the same optical power input.

[0017] Furthermore, the mode demultiplexer and the mode converter in the mode multiplexer are cascaded through a wedge waveguide; the mode demultiplexer forms M independent fundamental mode optical outputs, each output containing N wavelengths; the mode multiplexer receives the optical signal output after weighting by M groups of computing arrays, and realizes the weighted wavelength mode hybrid multiplexing output.

[0018] Furthermore, the mode demultiplexer in the computing module is formed by cascading M-1 mode converters through wedge-shaped waveguides, and the wavelength mode mixed multiplexed optical signal is converted into mode TE M-1 The TE0 mode light is demultiplexed from high order to low order in sequence to TE1, and the TE0 mode light is coupled to the ring waveguide array by a directional coupler after passing through the last mode converter. In this way, the mode demultiplexer forms M independent fundamental mode light outputs, and each output contains N wavelengths. Subsequently, the M optical signals containing N wavelengths are independently output to M groups of ring waveguide arrays cascaded by N pairs of ring waveguides and weighted waveguides to achieve the selection and weighting of specific wavelengths. Finally, the weighted multi-wavelength single-mode optical signal is multiplexed by the mode multiplexer to the bus waveguide for output, and the output power is the sum of the optical powers of the multi-wavelength and multi-mode signals. The mode multiplexer is composed of M-1 mode converters cascaded through wedge waveguides, and is used to receive the optical signal output after weighting by M groups of calculation arrays. Among them, the optical signal output of the first ring waveguide array is directly connected to the front end of the cascaded mode converter by a directional coupler, and then each weighted signal is converted into TE1 to TEM-1 mode light and multiplexed into the bus waveguide, thereby realizing the weighted wavelength mode mixed multiplexing output.

[0019] Furthermore, the mode converter is composed of an access waveguide and a bus waveguide, and the access waveguide and the bus waveguide form a coupling region, and when the light of different modes meets the phase matching condition in the coupling region, complete energy conversion occurs. That is, when the effective refractive index of the high-order mode light on the bus waveguide is equal to that of the fundamental mode / Bloch mode light in the access waveguide, after passing through the coupling region, the high-order mode light in the bus waveguide will be converted into the fundamental mode / Bloch mode light in the access waveguide, thereby realizing the mode multiplexing / demultiplexing function.

[0020] Furthermore, the design of the access waveguides of the mode converters in the mode demultiplexer and the mode multiplexer is different. The access waveguides used by the TE1 to TE3 mode converters are conventional silicon waveguides with the same physical parameters; M-1 The access waveguide used by the mode converter is a sub-wavelength grating waveguide with consistent physical parameters.

[0021] Furthermore, the sub-wavelength grating waveguide is an optical waveguide structure, which is composed of periodically arranged sub-wavelength scale (smaller than the wavelength of light) grating elements, and the mode in which light is guided and transmitted is the Bloch mode.

[0022] Furthermore, a Bloch mode to fundamental mode mode converter is cascaded at the output end of the sub-wavelength grating waveguide. The mode converter is composed of a series of discrete blocks whose widths gradually increase to the access waveguide width. The discrete blocks are embedded in discrete gaps of the sub-wavelength grating waveguide to form a conical mode conversion section to achieve mode conversion from Bloch mode to fundamental mode.

[0023] Furthermore, the ring waveguide array in the computing module is composed of two parallel input and output waveguides, N weighted waveguides comprising GSST material arranged between and perpendicular to the two input and output waveguides, and N pairs of two ring waveguides with completely identical physical parameters arranged left and right at the input and output ends of the weighted waveguide; one section of the ring waveguide is placed between the array input waveguide and the weighted waveguide to form a coupling region to constitute a first microring resonator (MR), and one section of the ring waveguide is placed between the array output waveguide and the weighted waveguide to form a coupling region to constitute a second ring resonator (MR).

[0024] Furthermore, when a specific wavelength in the input waveguide is consistent with the resonant wavelength of the MR, the wavelength will be demultiplexed by the first MR to the weighted waveguide and weighted with the GSST, and then re-multiplexed to the output waveguide through the second MR. The remaining wavelengths will continue to be guided by the input waveguide and weighted by the subsequent MR resonance to achieve the selection of a specific wavelength.

[0025] Furthermore, the GSST used in the weighted waveguide is a non-volatile phase change material (PCM) composed of four elements: germanium (Ge), antimony (Sb), selenium (Se) and tellurium (Te). During optical convolution calculations, GSST directly stores weight values ​​and participates in the calculation. The input optical signal passes through the GSST and interacts with the stored weights (modulating the light intensity or phase) to complete the multiplication of the weights and the input without the need for additional electrical processing steps, thus achieving optical storage and computing integration. In addition, the non-volatile nature of PCM allows weights to be stored for a long time without continuous power supply, and the storage itself consumes almost no energy.

[0026] The advantages and beneficial effects of the present invention are as follows:

[0027] The present invention proposes a high-throughput, low-energy optical convolution operator structure based on on-chip multi-dimensional multiplexing and phase change materials. Through the mixed multiplexing of the two dimensions of mode and wavelength, the scalability limitations of the current optical convolution operator in large-scale matrix processing are broken through, and the parallel computing capability of the optical convolution operator is further improved. In addition, the computing unit in the present invention greatly reduces energy consumption by storing weights with the help of non-volatile phase change materials (PCM), and enables the optical convolution operator to break through the limitations of the von Neumann architecture to achieve storage and computing integration, further improving computing performance. The present invention aims to achieve large-scale, low-power parallel convolution operations, and is expected to provide strong support for the widespread application of optical neural networks in artificial intelligence, big data analysis, and high-performance computing. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 This is a schematic diagram of the structure of a high-throughput, low-energy optical convolution operator based on on-chip multi-dimensional multiplexing and phase change materials according to an embodiment of the present invention;

[0029] Figure 2 is a schematic diagram of the structure of a mode converter and a schematic diagram of the structure of a cascaded mode converter according to an embodiment of the present invention, Figure 2 (a) shows a mode converter using a conventional access waveguide, Figure 2 (b) shows a mode converter using a sub-wavelength grating to access the waveguide. Figure 2 (c) represents a mode demultiplexer cascaded by a mode converter;

[0030] Figure 3 The curve showing the relationship between the effective refractive index and the waveguide width of the grating waveguide with different duty cycle wavelengths;

[0031] Figure 4 Schematic diagram of paired microring resonators and weighted waveguides used in the calculation module in the embodiment of the present invention and a schematic diagram of transmission curves of the Drop port and Through port of the microring resonator, Figure 4 (a) shows a paired microring resonator and a weighted waveguide. Figure 4(b) Transmission curves of the Drop port and Through port of the microring resonator;

[0032] Figure 5 Schematic diagram of the transmittance curve of the optical waveguide coupled with GSST of different lengths and crystallinity in an embodiment of the present invention. DETAILED DESCRIPTION

[0033] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0034] Figure 1 This is a structural diagram of the high-throughput, low-energy optical convolution operator based on on-chip multidimensional multiplexing and phase change materials described in the present invention, which is divided into three parts: input module, calculation module and output module.

[0035] The input module is composed of an off-chip laser array, an N×1 wavelength division multiplexer, a 1×M first optical splitter, M×N modulators, an M×1 mode multiplexer, and a 1×P second optical splitter. The N different wavelengths (λ1, λ2, λ3, ..., λ N ) enters the N×1 wavelength division multiplexer for multiplexing and integration and then output to the 1×M first optical splitter; the 1×M first optical splitter divides the multiplexed optical power of N wavelengths into M equal parts to form M-way multi-wavelength multiplexing output; each of the M-way multi-wavelength multiplexing output passes through N optical modulators to realize the loading of the input signal. There are a total of M×N such optical modulators, and finally an M×N input matrix is ​​formed; then the M optical signals loaded by the modulator are input to the M×1 mode multiplexer to be multiplexed into M different modes transmitted on the same waveguide; finally, the light after wavelength and mode multiplexing passes through the 1×P second optical splitter and enters the P-layer computing unit of the computing module.

[0036] The computing module is composed of P layers of computing units, each layer of computing units is composed of a mode demultiplexer, M groups of ring waveguide arrays, and a mode multiplexer. The mode demultiplexer is composed of a group of cascaded mode converters, which respectively convert TE M-1The mode light to TE1 mode light is converted into TE0 mode light in order from high to low mode orders; each ring waveguide array is composed of N pairs of ring waveguides and weighted waveguides with PCM coupled between the ring waveguides to realize the point multiplication function of the calculation unit; the mode multiplexer is composed of a group of cascaded mode converters, which convert the TE0 mode light into TE1 to TE0 mode light in order from low to high mode orders in cascade order. M-1 mode light and outputs the wavelength mode multiplexed signal.

[0037] The output module is composed of wavelengths of the P-layer computing units, mode multiplexing output signals and subsequent devices. The P-layer output signals are respectively O1, O2, O3, ..., O N express.

[0038] Figure 2 (a) and Figure 2 (b) is a schematic diagram of the mode converter. The mode converter used in the present invention is composed of an access waveguide and a bus waveguide. The access waveguide and the bus waveguide form a coupling region. When light of different modes meets the phase matching condition in the coupling region, complete energy conversion occurs. In the mode converter, the optical signal is transmitted along the dotted line in the figure. When the TE0 mode light is input from the lower access waveguide and reaches the coupling region, it will be completely converted into the TE0 mode light in the upper bus waveguide. M Mode light, similarly, when TE M When the mode light is input from the right end of the upper bus waveguide, it will be completely converted into the TE0 mode light in the lower access waveguide after passing through the coupling region, where TE M Refers to any higher-order mode light other than the TE0 mode.

[0039] Figure 2 The mode converter shown in (b) is a sub-wavelength grating waveguide mode converter. Unlike the traditional mode converter, the access waveguide of the sub-wavelength grating waveguide mode converter is a sub-wavelength grating structure in the coupling region. The sub-wavelength grating waveguide is an optical waveguide structure composed of periodically arranged sub-wavelength scale (smaller than the wavelength of light) grating elements, and the mode in which light is guided and transmitted is the Bloch mode. According to the effective medium theory, the effective refractive index of the grating waveguide can be defined as:

[0040]

[0041] Among them, n si and n clad are the refractive indices of silicon and silicon dioxide, respectively, and η is the duty cycle of the periodic grating structure, defined as:

[0042] η=a / Λ

[0043] Where a is the width of the grating element and Λ is the period length of the periodic structure.

[0044] The output end of the sub-wavelength grating waveguide is cascaded with a mode converter from Bloch mode to fundamental mode. The mode converter is composed of a series of discrete blocks whose widths gradually increase to the width of the access waveguide. The discrete blocks are embedded in the discrete gaps of the sub-wavelength grating waveguide to form a cone-like mode conversion section (such as Figure 2 (b) The transition cone realizes the mode conversion from Bloch mode to fundamental mode.

[0045] Figure 3 The graphs show the relationship between the effective refractive index and the waveguide width of the sub-wavelength grating waveguide with different duty cycles. It can be seen that the rate of change of the effective refractive index curve of the sub-wavelength grating waveguide is lower than that of the ordinary silicon wire waveguide, that is, the effective refractive index curve is flatter, so that phase matching can still be maintained in a larger range of waveguide width changes, realizing a higher-order mode converter with higher robustness.

[0046] The parameters of some mode converters constituting the mode multiplexer and demultiplexer in the optical convolution operator are shown in Table 1. In the TE1 to TE3 mode demultiplexers, the access waveguides used are conventional silicon waveguides with the same physical parameters; M-1 In the mode demultiplexer, the access waveguide used is a sub-wavelength grating waveguide with consistent physical parameters.

[0047] Table 1 Parameters of some mode converters cascaded in the mode multiplexer / demultiplexer

[0048]

[0049] Figure 2 (c) is a schematic diagram of a mode demultiplexer composed of cascaded mode converters. The mode demultiplexer is composed of M-1 different mode converters cascaded through a wedge-shaped waveguide. When a multi-mode multiplexed optical signal is input from the left end of the mode demultiplexer, TE M-1 The optical signal will be converted into a TE0 optical signal in the lower access waveguide in the coupling region between the waveguide M and the access waveguide structure. M-2 The optical signal will be converted into TE0 optical signal in the lower access waveguide in the coupling region of waveguide M-1 and access waveguide structure, and so on, until all mixed-mode optical signals in the bus waveguide are demultiplexed into TE0 optical signal in the access waveguide.

[0050] The M independent outputs of the mode demultiplexer are respectively output to M groups of ring waveguide arrays. The ring waveguide array consists of two parallel input and output waveguides, N weighted waveguides comprising GSST materials arranged between and perpendicular to the two input and output waveguides, and N pairs of two sections of ring waveguides with completely identical physical parameters arranged at the input and output ends of the weighted waveguides.

[0051] Figure 4 (a) shows each pair of ring waveguides and weighted waveguides. A section of the ring waveguide is placed between the array input waveguide and the weighted waveguide to form a coupling region to constitute a first microring resonator (MR), and a section of the ring waveguide is placed between the array output waveguide and the weighted waveguide to form a coupling region to constitute a second ring resonator (MR).

[0052] Figure 4 (b) shows the relationship between the transmittance of the Drop and Through ports of MR and the wavelength. MR is a device used for wavelength selection and wavelength filtering. Its basic working principle is that the light wave propagates in a micro-ring to form an interference effect. When the light propagation path and the phase matching conditions in the ring are met, a resonance phenomenon of a specific wavelength will occur, causing the input signal to be enhanced at this wavelength, while other wavelengths are suppressed. The phase matching condition of MR is shown in the following formula:

[0053]

[0054] Where λ is the wavelength of light, m ​​is the order of the resonant mode, and n eff is the effective refractive index, and R is the radius of the microring. When a specific wavelength in the input waveguide is consistent with the resonant wavelength of the MR, the wavelength will be demultiplexed by the first MR to the weighted waveguide and weighted with the GSST, and then remultiplexed to the output waveguide through the second MR, and the remaining wavelengths will continue to be guided by the input waveguide and weighted by the subsequent MR resonance to achieve the selection of a specific wavelength.

[0055] The GSST is a non-volatile phase change material (PCM), which is a type of material that can switch between different crystal structure states and is widely used in non-volatile memory devices, optical storage, thermal management and other fields. The main feature of PCM is its reversible transformation between different phases, and this transformation can "memorize" the state, which makes PCM non-volatile. In addition, PCMs in different phases have different physical properties (conductivity, optical transmittance, etc.), which is the basis for PCM to work in various fields.

[0056] In the present invention, GSST is used as a weight storage block, and different weight information is stored by controlling the phase change of GSST. When GSST is in a crystalline state, the light in the waveguide is absorbed by GSST due to the evanescent coupling effect, resulting in a significant decrease in the output optical power; while in the molten state, GSST has no effect on the light in the waveguide, thereby achieving storage of weights between 0 and 1 without the need for a continuous external power supply. In addition, GSST also has a reconfigurable property. When current is injected into the electrodes at both ends of the GSST, the temperature rises to the GSST phase change threshold. At this time, GSST will undergo a phase transition controlled by the input current, including any state from the molten state to the fully crystalline state. When sufficient current is applied, the PCM material undergoes melting and rapid cooling to form a molten state. Conversely, applying a lower current can cause the material to transform from a disordered state to an ordered crystalline state.

[0057] The transmittance curve of the optical waveguide coupled with GSST is shown in Figure 5 As shown in the figure, when GSST is in any phase, the light passing through the weighted waveguide will be weakened by the evanescent coupling effect of GSST and waveguide, thus achieving weighted loading. Assuming that the optical power input into the GSST coupled waveguide is X and the weight of GSST is M, the output optical power after passing through the weighted waveguide is Y=M×X.

[0058] like Figure 1 As shown in the calculation module in FIG, N pairs of ring waveguides and weighted waveguides are cascaded to form a ring waveguide array, in which the multi-wavelength single-mode light input to the array will be wavelength-selected and weighted and then re-multiplexed and output. There are a total of M groups of such ring waveguide arrays, which are composed of M Figure 2 (a) and Figure 2 The mode multiplexer cascaded into the mode converter shown in (b) receives and performs mode multiplexing, and finally forms the output of the entire calculation unit. Assuming that the weight value of GSST is m, and the input light loaded by the input module is an M×N matrix, the calculation process of the entire convolution calculation unit is the dot product of two M×N size matrices, as shown in the following formula:

[0059] The computing module in the present invention is composed of P computing units with the same structure. The only difference between each computing unit is the change of the weight matrix stored in GSST. Each time the input vector will be convolved in parallel with P weight matrices of size M×N, which greatly improves the parallel computing throughput of the convolution.

[0060] The systems, devices, modules or units described in the above embodiments may be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, the computer may be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

[0061] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0062] A person skilled in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing the relevant hardware through a program, and the program can be stored in a computer-readable storage medium, which can include: ROM, RAM, disk or CD, etc.

[0063] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A high-throughput, low-energy optical convolution operator based on on-chip multidimensional multiplexing and phase change materials, comprising an input module, a calculation module and an output module, characterized in that: The input module includes a laser array, an N×1 wavelength division multiplexer, a 1×M first optical splitter, M×N modulators, an M×1 mode multiplexer, and a 1×P second optical splitter; the N different wavelengths of light generated by the laser array enter the N×1 wavelength division multiplexer for multiplexing and integration, and then are input into the 1×M first optical splitter; The 1×M first optical splitter divides the multiplexed optical power of N wavelengths into M equal parts, forming M multi-wavelength multiplexed outputs; Each of the M multi-wavelength multiplexing outputs passes through N optical modulators to load the input signal. There are a total of M×N optical modulators, ultimately forming an M×N input matrix. Then, the M optical signals loaded by the modulator are input into the M×1 mode multiplexer to be multiplexed into M different modes transmitted on the same waveguide; finally, the light after wavelength and mode multiplexing enters the computing module through the 1×P second optical splitter; The computing module includes P layers of computing units, each layer of computing units includes a mode demultiplexer, M groups of ring waveguide arrays and a mode multiplexer; the mode demultiplexer is composed of a group of cascaded mode converters, and the group of cascaded mode converters converts TE M-1 The mode light to the TE1 mode light is converted into the TE0 mode light in order from high to low mode orders; Each ring waveguide array is composed of N pairs of ring waveguides and weighted waveguides with PCM coupled between the ring waveguides; the mode multiplexer is composed of a group of cascaded mode converters, which convert TE0 mode light into TE1 to TE2 mode light with the order from low to high in the cascade sequence. M-1 mode light, and output the wavelength mode multiplexed signal; The output module outputs the wavelength and mode multiplexing output signals of the P-layer computing units.

2. According to claim 1, a high-throughput, low-energy optical convolution operator based on on-chip multi-dimensional multiplexing and phase change materials, characterized in that: The laser array in the input module includes N lasers generating different wavelengths. The different wavelengths of light generated by these N lasers enter the N×1 wavelength division multiplexer for multiplexing. The mixed multi-wavelength optical signal after multiplexing is evenly divided into M equal parts by the 1×M first optical splitter and output independently, wherein each equal part includes N wavelengths. Subsequently, each of the M optical signals is independently modulated by N modulators for each wavelength, thereby completing the loading of the N×M input matrix.

3. A high-throughput, low-energy optical convolution operator based on on-chip multi-dimensional multiplexing and phase change materials according to claim 1 or 2, characterized in that: The M optical signals loaded by the modulator in the input module are respectively connected to M-1 mode converters to realize the conversion from fundamental mode to TE M-1 Mode conversion from mode light to TE1 mode light, wherein there is one path that does not undergo mode conversion processing and is connected to subsequent devices as TE0 mode light.

4. According to claim 3, a high-throughput, low-energy optical convolution operator based on on-chip multi-dimensional multiplexing and phase change materials, characterized in that: The TE M-1 The TE0 mode light is connected to the M×1 mode multiplexer, and M different mode optical signals are multiplexed into the same waveguide.

5. According to claim 1, a high-throughput, low-energy optical convolution operator based on on-chip multi-dimensional multiplexing and phase change materials, characterized in that: The mode demultiplexer and the mode converter in the mode multiplexer are cascaded through a wedge-shaped waveguide; the mode demultiplexer forms M independent fundamental mode light outputs, each output containing N wavelengths; The mode multiplexer receives the optical signal output after being weighted by the M groups of computing arrays, and realizes the weighted wavelength mode mixing and multiplexing output.

6. According to claim 5, a high-throughput, low-energy optical convolution operator based on on-chip multi-dimensional multiplexing and phase change materials, characterized in that: The mode converter is composed of an access waveguide and a bus waveguide, and the access waveguide and the bus waveguide form a coupling region. When light of different modes meets the phase matching condition in the coupling region, complete energy conversion occurs.

7. A high-throughput, low-energy optical convolution operator based on on-chip multi-dimensional multiplexing and phase change materials according to claim 5 or 6, characterized in that: The access waveguide used by the mode converters in the mode demultiplexer and the mode multiplexer, the TE1 to TE3 mode converters, is a conventional silicon waveguide with the same physical parameters; the TE4 to TE M-1 The access waveguide used by the mode converter is a sub-wavelength grating waveguide with consistent physical parameters.

8. The high-throughput, low-energy optical convolution operator based on on-chip multi-dimensional multiplexing and phase change materials according to claim 7, characterized in that: A mode converter from Bloch mode to fundamental mode is cascaded to the output end of the sub-wavelength grating waveguide. The mode converter is composed of a series of discrete blocks whose widths gradually increase to the width of the access waveguide. The discrete blocks are embedded in discrete gaps of the sub-wavelength grating waveguide to form a conical mode conversion section to achieve mode conversion from Bloch mode to fundamental mode.

9. The high-throughput, low-energy optical convolution operator based on on-chip multi-dimensional multiplexing and phase change materials according to claim 1, characterized in that: The ring waveguide array in the computing module is composed of two parallel input and output waveguides, N weighted waveguides comprising GSST material arranged between and perpendicular to the two input and output waveguides, and N pairs of two ring waveguides arranged left and right at the input and output ends of the weighted waveguides with completely identical physical parameters; one ring waveguide is arranged between the array input waveguide and the weighted waveguide to form a coupling region to constitute a first microring resonator (MR), and one ring waveguide is arranged between the array output waveguide and the weighted waveguide to form a coupling region to constitute a second microring resonator (MR).

10. The high-throughput, low-energy optical convolution operator based on on-chip multi-dimensional multiplexing and phase change materials according to claim 9, characterized in that: When a specific wavelength in the input waveguide is consistent with the resonant wavelength of the MR, the wavelength will be demultiplexed by the first MR to the weighted waveguide and weighted with the GSST, and then re-multiplexed to the output waveguide through the second MR. The remaining wavelengths will continue to be guided by the input waveguide and weighted by the subsequent MR resonance to achieve the selection of a specific wavelength.

Images