Storage element, storage method, storage array and application method

Abstract

The invention relates to the field of data storage and calculation based on light regulation and control, and discloses a storage and calculation element, a storage and calculation method, a storage and calculation array and an application method of the storage and calculation array. The storage and calculation element comprises at least one dielectric layer with light polarization or light intensity modulation control preparatory property, and further comprises a data detection module, and the data detection module is used for detecting the light intensity change after passing through the dielectric layer and converting the light intensity change into digital and / or analog signals to be output. The storage and calculation element accurately realizes light regulation and control and storage and calculation functions. According to the technology, weight storage and calculation are integrated in a storage and calculation element, optical / electric signals are used for direct calculation, and calculation and storage efficiency is improved. The matrix storage and calculation scheme can be used for efficiently realizing data storage and calculation and signal detection, and is adaptive to application scenes such as large-scale integrated operation circuits and neural network nodes.

Description

Technical Field

[0001] The present invention relates to the field of storage and computing technology, and in particular to a storage and computing element, a storage and computing method, a storage and computing array, and an application method based on light regulation. Background Art

[0002] With the rise of artificial intelligence, the demand for computing power is increasing, and computing power is often positively correlated with electricity consumption. Although the energy consumption of computing equipment can be reduced to a certain extent by improving the manufacturing process, it is still a drop in the bucket compared to the increasing electricity consumption. In the field of chips involving storage and computing integration, memristors are currently the most commonly used nonlinear circuit elements with memory functions to achieve storage and computing integration. Their resistance value changes according to the current or voltage history passing through it, and they can still maintain the current resistance state after the external stimulus disappears, thereby "memorizing" past electrical history information. From a physical mechanism point of view, the resistance change of a memristor usually stems from microscopic processes such as the migration of internal ions or phase change of the material. For example, in some metal oxide-based memristors, applying voltage causes the movement of oxygen ions, thereby changing the conductive channel inside the material and thus changing the resistance value. Although memristors and similar components have many advantages, they also have some disadvantages in practical applications, which are mainly reflected in the following aspects:

[0003] 1. Device Characteristics: Reliability is limited. After repeated read and write operations, memristors may experience performance degradation, such as resistance drift and threshold voltage variations. This reduces the accuracy of stored data and the reliability of computational results, limiting their application in mission-critical applications requiring extremely high reliability. Furthermore, memristors often exhibit complex nonlinear resistance variations, and their actual behavior deviates from the ideal model. This makes accurate modeling and control of memristors difficult, increasing the complexity of circuit design and system optimization.

[0004] 2. Limitations in material selection: Most of the materials currently used to manufacture memristors have certain limitations.

[0005] For example, the preparation process of some materials is complex and costly, making them difficult to achieve large-scale production; while other materials may have problems with stability and compatibility, which limits the performance improvement and application expansion of memristors.

[0006] 3. Process compatibility challenges: The compatibility of memristors with traditional complementary metal oxide semiconductor (CMOS) processes needs to be improved. When integrating memristors with CMOS circuits, issues such as process temperature and material compatibility may arise. This increases the difficulty and cost of chip manufacturing, hindering the widespread application of memristors in existing integrated circuit systems.

[0007] The present invention will attempt to use liquid crystal or electrochromic glass as the storage and computing medium, avoiding the ion migration degradation problem of the memristor, and has higher compatibility with CMOS technology. It can be manufactured through existing flat panel display technology, reducing the difficulty of integration.

[0008] To this end, the inventors have conducted useful explorations and attempts and found a solution to the above-mentioned problem. The solution to be introduced below is produced under this background. Summary of the Invention

[0009] The technical problem to be solved by the present invention is to provide a storage and computing element, a storage and computing method, a storage and computing array, and an application method of the storage and computing array in response to the deficiencies and defects of the existing technology.

[0010] The technical problem solved by the present invention can be achieved by adopting the following technical solutions:

[0011] A storage and computing component, characterized by comprising:

[0012] At least one light-regulating medium layer, which is a liquid crystal layer and / or an electrochromic layer; used to realize data storage and calculation through changes in optical properties;

[0013] A transparent conductive layer is provided on both sides of the dielectric layer and is used to apply an electric field to regulate the polarization rotation angle or transmittance of the dielectric layer;

[0014] The liquid crystal layer type storage and computing element is provided with a polarizing layer, which is provided on the incident side and / or the exit side of the liquid crystal layer and is used to control the polarization state of light and cooperate with the dielectric layer to achieve light polarization state modulation;

[0015] A photoelectric detection module, including a photodiode or a photodetector, is used to detect the intensity of light passing through the dielectric layer and convert it into an electrical signal;

[0016] The liquid crystal layer realizes the output light intensity = incident light intensity × rotation angle / 90 or incident light intensity × ((90-rotation angle) / 90) through the polarization rotation angle; the electrochromic layer realizes the output light intensity = incident light intensity × transmittance through the transmittance.

[0017] In a preferred embodiment of the present invention, the dielectric layer is connected to an external control circuit, and the control circuit regulates the polarization rotation angle of the dielectric layer or the transmittance of the electrochromic layer by applying voltage to the transparent conductive layer. The data detection module is a light detection element such as a photodiode or a photodetector.

[0018] In a preferred embodiment of the present invention, the polarizing layer is a grating structure, forming a polarizer that only allows light of a specific polarization state to pass through, and is attached to both sides of the liquid crystal layer.

[0019] A storage and calculation method, wherein the storage and calculation component comprises the following steps:

[0020] Step 1: If the memory computing element is a liquid crystal layer type, an initial light beam is emitted from a light source, filtered by a polarizing layer to a single polarization state, and then incident on the liquid crystal layer; if the memory computing element is an electrochromic layer type, the polarization step is omitted and the initial light beam is directly incident on the electrochromic layer;

[0021] Step 2: The control circuit applies a set voltage to the transparent conductive layer to adjust the polarization rotation angle or transmittance of the dielectric layer, thereby achieving a multiplication operation of the incident light intensity and the control coefficient;

[0022] Step 3: The data detection module converts the outgoing light intensity into an electrical signal, which is then output through analog-to-digital conversion or directly as an analog signal.

[0023] In a preferred embodiment of the present invention, the polarization rotation angle of the control medium layer is 0-90 degrees, and the light transmittance of the control medium layer is 0-100%.

[0024] In a preferred embodiment of the present invention, the storage and computing elements are arranged in a two-dimensional matrix and connected to the row selector and the column selector via vertically intersecting row lines and column lines;

[0025] A light splitting device can be set between adjacent storage and computing element layers to split the light into a calculation beam and a feedback detection beam;

[0026] The storage and calculation element realizes vector-matrix multiplication through light intensity multiplication operation by regulating the polarization state or transmittance of light and storing weight values.

[0027] In a preferred embodiment of the present invention, the storage and computing elements are arranged in a two-dimensional matrix, and are vertically cross-wired by metal wires. The row lines and column lines are respectively connected to the row selector and column selector to achieve precise addressing of a single storage and computing element.

[0028] In a preferred embodiment of the present invention, an accumulation circuit is further included, which is connected to the column selector or the storage and calculation elements of the same group, and is used to accumulate the electrical signals output by the storage and calculation elements of the same column or the same group to realize the accumulation operation of the matrix multiplication results.

[0029] An application method of a storage and computing array, wherein the storage and computing array comprises the following steps:

[0030] Data input: The input data is converted into a light intensity signal by the light source. If it is a liquid crystal layer type, it is incident on the storage and computing array through the polarizing layer;

[0031] Weight calculation: The row selector and column selector select the target storage element and realize the multiplication of input light intensity and weight by adjusting the transmittance;

[0032] Result accumulation: The output electrical signals of the same column or the same group of storage and calculation elements are summed through the accumulation circuit to obtain the intermediate result of matrix multiplication;

[0033] Non-linear activation: Through independent activation layer storage elements, the intermediate results are non-linearly transformed using comparators (such as ReLU) or lookup tables (such as Sigmoid);

[0034] Signal output: The final electrical signal is converted from digital to analog or directly output as an analog signal.

[0035] In a preferred embodiment of the present invention, the nonlinear activation step is specifically:

[0036] ReLU function: Compares the linear calculation result with the preset threshold through a comparator circuit. When the result is greater than the threshold, the original value is output. The corresponding formula is ReLU(x) = max(0,x);

[0037] Sigmoid function: Pre-store the input-output correspondence through a lookup table (LUT), use the linear calculation result as an index to quickly query and output the corresponding nonlinear transformation value, and achieve approximate calculation of σ(x) = 1 / (1+ex).

[0038] Due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0039] 1. The present invention has a long service life. The liquid crystal and the color-changing glass have good fatigue resistance, can quickly write and erase information, and can maintain stable performance after multiple cycles. For example, the color-changing glass can undergo more than 300,000 changes in light and dark without failing. Liquid crystal materials can usually withstand more than 100,000 or even higher orders of magnitude of state changes driven by electric fields. In the field of storage technology, flash memory is widely used for its excellent performance, and its cycle times can reach 10 3 -10 7 Therefore, from the perspective of the storage field, the life of the material used in the present invention has the characteristics of an excellent storage device.

[0040] 2. Stable equipment performance: Data storage and processing functions are achieved through current or voltage regulation, ensuring long-term and reliable operation of the equipment.

[0041] 3. The manufacturing process of the present invention is relatively mature, simple and low-cost: the manufacturing process of the storage unit can be realized by slightly improving the existing technology and is suitable for large-scale production.

[0042] Fourth, the present invention's memory-computing element accurately implements both light control and memory-computing functions. This memory-computing technology integrates weight storage and computation within the memory-computing element, utilizing optical / electrical signals for direct computation, improving computational and storage efficiency. This memory-computing matrix solution can be used to efficiently implement data storage, computation, and signal detection, and is suitable for applications such as large-scale integrated computing circuits and neural network nodes. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0044] Figure 1 This is a schematic diagram of the principle of implementing the storage and computing functions based on optical signal processing of the present invention.

[0045] Figure 2 This is a schematic diagram of the working principle of the storage and computing element based on the liquid crystal structure of the present invention.

[0046] Figure 3 This is a schematic diagram of the optical path principle of the electrically controlled color-changing glass storage and computing element of the present invention.

[0047] Figure 4 Schematic diagram of the single-layer vertical cross wiring scheme of the storage and computing matrix of the present invention.

[0048] Figure 5 Schematic diagram of the data reading architecture of the storage and computing element of the present invention.

[0049] Figure 6 Schematic diagram of the data reading architecture of the storage and computing element rows and columns jointly selected in the present invention.

[0050] Figure 7 Schematic diagram of the multi-layer vertical cross wiring scheme of the storage and computing matrix of the present invention.

[0051] Figure 8 Schematic diagram of the principle of the two-layer inter-light splitting device of the present invention for storage and calculation operations.

[0052] Figure 9 Schematic diagram of the vertical cross wiring and light splitting hybrid solution of the storage and computing matrix of the present invention.

[0053] Figure 10 Schematic diagram of the light supplementation principle based on the spectroscopic device of the present invention.

[0054] Figure 11 This is a diagram of the control principle of the storage and computing element of the present invention under the action of dual control light beams.

[0055] Figure 12 Schematic diagram of the integrated principle of vertical cross wiring control and detection of storage and computing elements in the present invention.

[0056] Figure 13 Schematic diagram of the light splitting detection principle based on the vertical cross wiring matrix of the present invention.

[0057] Figure 14 The figure shows the light intensity detection and calculation principle of the liquid crystal storage and calculation element combining light splitting and vertical cross wiring of the present invention.

[0058] Figure 15 The figure is a schematic diagram of the principle of realizing matrix multiplication based on a vertical cross wiring matrix of the present invention. DETAILED DESCRIPTION

[0059] In order to make the technical means, creative features, objectives and effects achieved by the present invention easier to understand, the present invention is further described below.

[0060] See Figure 1 As shown, the present invention is based on the fundamental principles of optical computing, controlling the storage and computing elements through which light passes. These elements act on the light, affecting its polarization or intensity. This changing optical signal is detected by other elements and can be implemented in the following ways: 1. Directly reflecting the magnitude of an analog quantity in the form of an analog signal for analog storage or calculation. 2. Alternatively, it can be converted into a digital signal through analog-to-digital conversion technology for digital storage or calculation. 3. Switching between digital and analog can be achieved at any time according to different application scenarios.

[0061] The present invention manipulates storage and computational elements to process data represented by optical signals (representing computational activity). By detecting optical signals (generally, the intensity of the light is used to represent the amount of processed data), storage and / or computation, or computation and / or storage, is achieved. The middle layer represents the superposition of storage and computational elements, increasing or decreasing the number of serially distributed storage and computational elements based on different computational scenarios.

[0062] In the embodiment of the present invention, the storage and computing component implements the above functions through the following two parallel schemes:

[0063] Example 1: Liquid crystal layer type memory computing element

[0064] See Figure 2 As shown, the schematic diagram of the liquid crystal storage and computing element shows the structural composition of the liquid crystal storage and computing element, including incident light, a polarizing layer, a transparent conductive layer, a liquid crystal layer and a voltage applying device (marked with "V" in the figure).

[0065] Liquid crystal molecules change their orientation under the influence of an electric field due to their dielectric anisotropy. In nematic liquid crystals, when an electric field is applied to the transparent conductive layers on either side of the liquid crystal layer, the liquid crystal molecules rotate under the influence of the electric field. The angle of rotation of the liquid crystal is determined by the desired polarization state or transmittance, thereby changing its optical properties.

[0066] The incident light first passes through the first polarizing layer composed of gratings. This polarizing layer only allows light with a specific polarization state to pass through, giving the light a uniform polarization state. Subsequently, the light enters the liquid crystal layer. Under the control of the electric field, the orientation of the liquid crystal molecules changes, thereby regulating the polarization state of the light. Finally, the light passes through the second polarizing layer, which is perpendicular to the grating of the first polarizing layer. The angle between the polarization state of the light and the grating of the second polarizing layer determines the intensity of the transmitted light. By detecting the intensity of the outgoing light and converting it into an electrical signal or digital signal, the computing function can be realized. By precisely controlling the presence and strength of the electric field at both ends of the transparent conductive layer, the orientation of the liquid crystal molecules can be accurately adjusted to achieve different light polarization state control to adapt to different computing tasks.

[0067] In addition, the memory function of the liquid crystal storage and computing element is based on maintaining the position of the liquid crystal unchanged. When the liquid crystal is operated, if its initial position is maintained, the liquid crystal position records the operation information, thus realizing the memory function. However, the liquid crystal position may drift over time, affecting the memory stability. To solve this problem, when the memory state needs to be restored, the stored liquid crystal position information is read from the storage device, and the electric field at both ends of the transparent conductive layer is adjusted by the control circuit based on this information to restore the liquid crystal position to the state at the time of recording, ensuring that the memory function is reliably realized. For example, if the stored position information indicates that the angle between the liquid crystal and the second-level grating is 30 degrees, the liquid crystal molecular orientation is restored to the state corresponding to the 30-degree angle by adjusting the electric field.

[0068] Specifically, the transparent conductive layer in the embodiment of the present invention can be made of a transparent conductive material such as indium tin oxide. The transparent conductive layer can be uniformly deposited on the substrate material through methods such as magnetron sputtering or chemical vapor deposition to ensure good conductivity and optical transparency. The liquid crystal layer can be made of a suitable nematic liquid crystal material, such as 5CB (4-n-pentyl-4'-cyanobiphenyl). The liquid crystal material is filled between the two transparent conductive layers through methods such as spin coating or perfusion to form a liquid crystal layer with a thickness of 5-8 microns. The polarizing layer can utilize grating fabrication techniques such as photolithography or holographic exposure to create a grating structure on a polymer material, forming a polarizing layer that only allows light of a specific polarization state to pass through, and then adhered to both sides of the liquid crystal layer.

[0069] The transparent conductive layer is connected to an external control circuit, which precisely adjusts the magnitude and direction of the voltage applied across it. The voltage adjustment range is determined by the computational task, and the corresponding electric field is applied through the control circuit. For example, for simple logic operations, a voltage of 2 volts can be applied; for complex matrix operations, the voltage can be adjusted based on the scale and accuracy required.

[0070] Example 2: Memory and calculation element of electrically controlled color-changing glass

[0071] See Figure 3 As shown, the electrochromic glass storage and computing element realizes the storage and computing function based on the characteristics of the electrochromic glass. The transmittance of the electrochromic layer is regulated by controlling the changes in the electrical signals at both ends of the transparent conductive layer. It should be noted that in the embodiment of the present invention, the transparent base layer mainly plays a protective role for the transparent conductive layer and is not an essential component. The light transmittance represents different weights and values ​​and can be used to realize computing and storage functions. However, since the transmittance of most photochromic glasses does not have linear control characteristics, it is necessary to go through a calibration program to realize its control function and improve the accuracy of the solution. This is also applicable to liquid crystal storage and computing elements.

[0072] The control schemes are divided into two types: direct control scheme and proofreading control scheme:

[0073] Direct control scheme: According to the parameters that have been verified in advance, the storage and computing elements are directly controlled to achieve the target transmittance. That is, through theoretical calculation and experimental verification, the relationship between the transmittance and the electrical signal (voltage or current) parameters required in different application scenarios is determined. For example, for a specific computing task, after testing and calculation, it is found that when a voltage of 2 volts is applied for 300 milliseconds, the transmittance of the storage and computing element can reach 0.6, which meets the computing requirements. That is, the direction, magnitude and duration of the applied voltage or current are controlled to achieve different transmittances. After calculating the parameter relationship, the transparent conductive layer is connected to the external power supply and the control circuit, and the corresponding voltage or current is applied through the control circuit according to the determined parameters. As in the above example, the control circuit applies a voltage of 2 volts to the transparent conductive layer and maintains it for 300 milliseconds to achieve the control of the transmittance of the storage and computing element. In the embodiment of the present invention, a logic control unit is used to control the storage and computing element to affect the transmittance of light, and the logic control unit receives the external data signal and makes a control instruction for the transmittance of the storage and computing element. For example, a positive instruction can control the transmittance to increase, and a negative instruction can control the transmittance to decrease. There are two specific control schemes:

[0074] The proofreading and control plan is divided into the testing phase and the functional phase:

[0075] Testing phase: The first step is the emission and detection of light. A stable light source is used to emit a front beam of light, which passes through the storage and computing element to form a rear beam of light. The intensity of the rear beam is detected using a light detection element, and the light intensity signal is converted into an electrical signal and fed back to the logic control unit.

[0076] The second step is comparison and regulation. The logic control unit compares the received light intensity electrical signal with a pre-set expected value. If the electrical signal corresponding to the detected backbeam intensity is greater than the expected value, the logic control unit sends a command to the control circuit of the transparent conductive layer to change the magnitude or direction of the applied voltage or current, thereby reducing the transmittance of the storage element. If it is less than the expected value, the logic control unit sends a command to change the magnitude or direction of the voltage or current to increase the transmittance. This process is repeated until the detected light intensity approaches the expected value.

[0077] Functional Stage: Once the transmittance of the storage and calculation element is adjusted to the desired value and stabilized, the electrical signal input at that point is fixed. For example, if the transmittance is adjusted to 0.5 and stabilized, it can be considered to store the value "0.5." When the front beam is input, the multiplication function is performed by detecting the back beam intensity according to the formula "back beam = front beam * 0.5", completing the storage and calculation operation.

[0078] In the embodiment of the present invention, the light detection part can be a photosensitive element, such as a photoresistor, a photosensitive semiconductor or other light detection element. For example, if the photoresistor is integrated with the storage and calculation element, the resistance of the photoresistor changes with the intensity of the received light. By detecting the change in current or voltage across the photoresistor, the numerical change in the calculation result of the storage and calculation element can be reflected. For example, after the storage and calculation element completes the calculation, the change in the output light intensity causes the resistance of the photoresistor to change. By measuring its current change, the numerical information corresponding to the calculation result can be obtained. By integrating the photoresistor with the storage and calculation element, the photoresistor has the function of a traditional memristor and realizes the memory function. For example, a photosensitive semiconductor, such as a photodiode or a phototransistor, is integrated with the storage and calculation element. When a photosensitive semiconductor is irradiated with light, photogenerated carriers are generated, and its electrical properties change. By detecting the change in electrical parameters such as current, voltage or resistance of the photosensitive semiconductor, the calculation result information of the storage and calculation element can be obtained, so that the photosensitive semiconductor becomes a semiconductor element with memory function, further improving the storage and calculation integrated function.

[0079] Since the functions of a single storage and computing element are relatively limited, it is difficult to meet the needs of large-scale data processing. In order to achieve important functions such as efficient data storage, parallel computing, and serving as a neural network node, it is necessary to use a crossbar array to form a large-scale integrated computing circuit. In the process of constructing such a crossbar array, the crossbar array assembly method, control method, and data access mechanism in the storage and computing element become key issues. The embodiment of the present invention also discloses a storage and computing matrix wiring scheme for constructing a large-scale integrated computing circuit to realize data storage, parallel computing, and application in neural networks of storage and computing elements.

[0080] Vertical cross wiring implementation scheme:

[0081] See Figure 4 As shown, a single-layer vertical cross-wiring matrix shows that the storage and computing elements are connected to the row selectors and column selectors via row and column lines to achieve element addressing. Metal wires are carefully wired in a vertical cross pattern to connect the storage and computing elements, row selectors, and column selectors, ensuring the accuracy and stability of signal transmission. The control signal inputs of the row and column selectors are connected to the external control circuit. The control circuit can use a high-performance microcontroller (MCU). Through precise control programs, it can control the selectors in real time and accurately, achieving flexible operation of the storage and computing elements in the storage and computing matrix.

[0082] Please combine Figure 5 The schematic diagram of the row data reading architecture for storage and computational elements shows the row selection process for a single-layer matrix operation: the control circuit sends a row selection signal to the row selector based on the specific needs of data storage, parallel computing, or neural network operations. For example, to operate the storage and computational elements in the mth row, the control circuit sends a selection signal corresponding to the mth row. The transistor switch array within the row selector switches accordingly, connecting the storage and computational elements in the mth row to subsequent data processing circuits, such as samplers and amplifiers, providing a path for subsequent data reading, writing, or computation operations.

[0083] Please combine Figure 6 The schematic diagram of the data reading architecture of the joint selection of rows and columns of storage computing elements shown in the figure shows the column selection process of the single matrix operation: after completing the row selection, if it is necessary to further select the storage computing element of a specific column, the control circuit sends a column selection signal to the column selector. For example, if the storage computing element of the nth column is to be selected, the column selector connects the storage computing element of the nth column of the row to the subsequent data processing circuit according to the received signal to achieve precise positioning of a single storage computing element. At this point, the selected storage computing element can write data (such as storing the weight value of the neural network, etc.), read (such as reading the stored data for parallel computing) or participate in calculations (such as in matrix multiplication operations) and other operations.

[0084] Please combine Figure 7As shown, when the storage and computation matrix has a multi-layer structure and serial data processing is required between layers, the control circuit first sends a layer selection signal to the layer selector to determine the target layer to be operated. The layers are arranged in a two-dimensional matrix, connected to the row and column selectors via vertically intersecting row and column lines. Specifically, for example, if the storage and computation matrix has k layers and an operation is to be performed on the i-th layer, the layer selector, based on the received signal, connects the storage and computation element array of the i-th layer to the row and column selection circuits, allowing subsequent row and column selection operations to be performed on that specific layer. After selecting a layer, the row and column selectors sequentially select the storage and computation elements of the target row and column, following the operational flow of a single-layer matrix. The control circuit precisely controls the signals of the row and column selectors based on the specific computation task, achieving accurate positioning and operation of specific storage and computation elements in the multi-layer matrix. For example, in multi-layer computation in a neural network, this method can sequentially operate on the storage and computation elements of each layer, enabling data transfer and computation between multiple layers.

[0085] The rows and columns of each computing unit layer are selected in sequence through the row selector and column selector. If the data is processed serially between the computing layers, a layer selector needs to be added, and the above related operations are performed in sequence.

[0086] After the storage and computing components are positioned, data operations are performed based on the specific application scenario. When data needs to be stored, such as storing trained weights in a neural network, the data signal to be stored is input into the selected storage and computing component via a connecting line. The storage and computing component stores data based on its internal storage mechanism, such as changes in liquid crystal molecular orientation or changes in the transmittance of electrochromic glass in the embodiments of the present invention.

[0087] When performing parallel computing tasks, such as matrix multiplication, the data involved in the calculation (such as matrix element values) is input into the corresponding storage and calculation element in the form of optical or electrical signals. The storage and calculation element performs the calculation according to a pre-set calculation rule (such as multiplication based on the corresponding relationship between light transmittance and data). The calculation result is then output through the circuit and subsequent operations such as accumulation are performed to finally obtain the calculation result.

[0088] In the development of integrated storage and computing technology, how to efficiently operate and regulate storage and computing elements is a key issue. The embodiments of the present invention disclose embodiments of a splitting solution and a hybrid solution.

[0089] See Figure 8The figure shows the structure of the spectrometer between the two layers, showing the changes in the optical path of the light after it passes through the spectrometer into beam 1 and beam 2, as well as the positional relationship between the two layers (the front layer and the back layer). A spectrometer can be set between adjacent storage and computing element layers to split the light into a computing beam and a feedback detection beam. Specifically, by introducing a spectrometer, such as a spectroscope or a spectrometer film, between two layers of the storage and computing matrix, parameters such as the film thickness and refractive index are precisely controlled so that light in a specific wavelength range is reflected and projected in a certain proportion, the light is divided into two beams, namely beam 1 and beam 2. Beam 1 continues to enter the next layer of vertical cross-wiring matrix for storage and computing, and beam 2 enters the light detection element to detect the light intensity or enters other functional elements to obtain feedback data. At the same time, the control light emitted by the logic control unit is returned through the reverse optical path of beam 2 to control the storage and computing elements of the layer where the original light passes.

[0090] Please combine Figure 9 The figure shows the optical splitter device between vertical cross-wiring matrix layers, demonstrating its application in the matrix structure and the propagation paths of beams 1 and 2 between matrix layers. Combining the optical splitter solution with other related solutions, such as vertical cross-wiring, can partially or fully leverage the advantages of both to achieve more efficient storage-computing matrix operations and meet the performance requirements of integrated storage-computing systems in different application scenarios.

[0091] It should be noted that between the corresponding two layers of the storage and computing matrix (the specific number of layers can be determined according to actual needs, such as installing a spectrometer every three layers), the spectrometer should be installed according to the designed optical path layout. To ensure the collimation and stability of the optical path, use equipment such as an optical adjustment frame to precisely adjust the angle and position of the spectrometer so that light accurately enters the spectrometer.

[0092] As light passes through multi-layer storage and computation matrices, its intensity gradually decreases due to various optical losses. When the intensity drops below a certain level, it can adversely affect the normal regulation and detection of signals, thereby reducing the performance and accuracy of the storage and computation matrices. This embodiment of the present invention also discloses an embodiment of a light compensation solution.

[0093] See Figure 10As shown, by introducing a supplementary light beam 3 after a certain layer of matrix, it is magnified in a certain proportion to the intensity of the detected light beam 2, and after reflection by the spectrometer, the light is supplemented along the same optical path as the light beam 1 to maintain the light intensity and ensure normal signal regulation and detection. Specifically, a light detection element with high sensitivity and fast response speed is selected, such as a photodetector or a photodiode (suitable for visible light band). According to the light intensity detection requirements of the storage matrix, the sensitivity, response bandwidth and other parameters of the detection element are determined. The light detection element is installed in a position where the light intensity of the light beam 2 can be accurately detected to ensure that the detection element is accurately aligned with the optical path. Use equipment such as an optical adjustment frame to accurately adjust the angle and position of the detection element so that it can efficiently convert the light intensity signal into an electrical signal output.

[0094] After the storage and computation matrix begins operating, the light detection element detects the intensity of beam 2 in real time and converts the intensity signal into an electrical signal for output. The output signal of the detection element is processed by signal processing circuits such as amplifiers and analog-to-digital converters, converting the analog electrical signal into a digital signal for subsequent analysis and processing.

[0095] The processed light intensity digital signal is fed back to a control system, such as a microcontroller, via a data transmission line. The control system analyzes the fed-back light intensity data to determine whether the light intensity is below a preset threshold.

[0096] When the control system detects that the intensity of Beam 2 falls below a threshold, it calculates the ratio of the intensity of supplementary Beam 3 to that of Beam 2 according to a pre-set light compensation algorithm. For example, if compensation is set to occur when the intensity of Beam 2 drops to 80% of its initial value, and the ratio of the intensity of supplementary Beam 3 to that of Beam 2 is set to 1.2, the required intensity of supplementary Beam 3 is calculated based on the actual intensity of Beam 2.

[0097] Based on the calculated intensity of the supplementary beam 3, the control system sends a control signal to the light source of the supplementary beam generating device. By adjusting parameters such as the light source's drive current or voltage, the light source's output intensity is precisely controlled to achieve the calculated supplementary intensity value.

[0098] Controlling the transmission of light along a specific reverse optical path to the storage and computing element, and based on the photovoltaic effect of the semiconductor corresponding to the storage and computing element, regulating the action of the light beam to generate a voltage difference, achieving arbitrary control of the transmittance of the storage and computing element, and collecting the control data for deep learning training to optimize accuracy are important factors in achieving precise control of the transmittance of the storage and computing element.

[0099] Please combine Figure 11 、 12 shown. Figure 11 This demonstrates the control principle of the splitting scheme. Figure 12The connection between the row selector, column selector, sampler, function selector, and the memory and computing element array is shown, as well as the structure of the photodiode in the memory and computing element, including the front transparent conductive layer, rear transparent conductive layer, unit column lines, and unit row lines, intuitively demonstrating the function switching and signal transmission path. The mechanism of controlling the transmittance by generating a voltage difference between the corresponding semiconductors of the memory and computing element by modulating the light beam is also presented.

[0100] Please combine Figure 13 As shown, the optical path is set as follows: a spectrometer is installed at a specific position of the storage and computing matrix (such as between two layers), and the control light beam transmission path is set according to the designed optical path layout, including optical elements such as reflectors and lenses, to ensure that the control light is accurately transmitted to the storage and computing elements.

[0101] Semiconductor Configuration: N-type and P-type semiconductors are placed on either side of the storage and computing element. Based on the photovoltaic effect, they generate a voltage difference when irradiated by a controlled light beam. Semiconductor materials with high photoelectric conversion efficiency, such as silicon and gallium arsenide, are used.

[0102] Detection and data collection equipment: Install light intensity detection elements (such as photodiodes), frequency meters, timers, and other equipment to detect parameters such as the intensity, frequency, and duration of the regulated light beam, and collect and store component transmittance data. The equipment is connected to the control system through a data acquisition circuit.

[0103] Control operation process:

[0104] Light transmission: The logic control unit emits a controlled light beam, which is transmitted to the memory-computing element along the reverse path of beam 2. Controlled beam 1 illuminates the semiconductor on one side of the memory-computing element, generating a voltage difference V1. Controlled beam 2 illuminates the semiconductor on the other side, generating a reverse voltage difference V2.

[0105] Transmittance Control: Based on control requirements, parameters such as the intensity, frequency, and duration of control beams 1 and 2 are adjusted, and the voltage differences V1 and V2 are varied to achieve arbitrary control of the transmittance of the storage and computing element. For example, increasing the intensity of control beam 1 increases V1, which in turn increases the transmittance of the storage and computing element.

[0106] Data Collection and Optimization: Detection equipment collects real-time data on control beam parameters and the transmittance of the memory and computing components, transmitting it to the control system. Deep learning algorithms analyze and train this data to establish a model for the relationship between control parameters and transmittance. Based on this model, the control strategy is optimized to improve control accuracy. For example, after repeated training, it is determined that the transmittance of the memory and computing component can accurately reach the target value under a specific light intensity and frequency combination.

[0107] It should be noted that the total number of layers composed of liquid crystal storage and computing elements is divided into several groups (can be divided into greater than or equal to 1 group), each group is composed of multiple layers (can be divided into greater than or equal to 1 layer), and each group divides computing resources according to the number of layers involved in the calculation. The total rotation angle within each group is 90 degrees. If there are n layers involved in the calculation, the number of angles allocated to each layer in each group is: 90 / n, or divided according to a certain proportion rule.

[0108] Please combine Figure 14 As shown, after each layer of calculation is completed, the light is split into beams 1 and 2 by the spectrometer. Before beam 2 reaches the light detector, a polarizing layer (polarizer) is placed between them. This polarizer must be perpendicular to the incident polarizer (polarizer) relative to the original polarization of the light. The intensity of this layer is ultimately detected, and the coefficient of this intensity is recorded (for example, the intensity represents the rotation angle, the maximum intensity represents the unrotated angle, and the actual rotation angle is (90 - rotation angle) divided by the maximum rotation angle of 90 to obtain the coefficient. Different formulas are used depending on the scenario). In summary, the liquid crystal layer rotates the polarization angle to achieve the following equation: outgoing light intensity = incident light intensity × rotation angle / 90, or incident light intensity × ((90 - rotation angle) / 90). The coefficients of each layer are multiplied cumulatively to obtain the final calculated value. Beam 1 continues forward for the next stage of calculation. By designing the spectrometer, we can minimize the energy loss by ensuring that the light reflected by beam 2 is detected by the detector while minimizing the energy of beam 2.

[0109] The embodiments of the present invention also disclose the cross-application of storage-computation integrated technology and the field of deep learning, specifically the implementation method of storage-computation matrix in matrix multiplication operations, activation function calculations and pooling operations, which is used to improve the efficiency of deep learning calculations.

[0110] Matrix multiplication, activation function calculations, and pooling operations are core components of deep learning computations. Traditional computing architectures, due to the separation of storage and computation, result in high data transmission overhead and low computational efficiency. Integrated storage and computing technology integrates storage and computation functions, reducing data transmission latency. However, it is urgent to clarify its specific implementation in typical deep learning operations to promote the development of efficient deep learning computing chips.

[0111] The present invention also discloses an application method of a storage and computing array, which comprises the following steps:

[0112] Data input: The input data is converted into a light intensity signal by the light source. If it is a liquid crystal layer type, it is incident on the storage and computing array through the polarizing layer;

[0113] Weight calculation: The row selector and column selector select the target storage element and realize the multiplication of input light intensity and weight by adjusting the transmittance;

[0114] Result accumulation: The output electrical signals of the same column or the same group of storage and calculation elements are summed through the accumulation circuit to obtain the intermediate result of matrix multiplication;

[0115] Non-linear activation: Through independent activation layer storage elements, the intermediate results are non-linearly transformed using comparators (such as ReLU) or lookup tables (such as Sigmoid);

[0116] Signal output: The final electrical signal is converted from digital to analog or directly output as an analog signal.

[0117] The nonlinear activation step is specifically as follows:

[0118] ReLU function: Compares the linear calculation result with the preset threshold through a comparator circuit. When the result is greater than the threshold, the original value is output. The corresponding formula is ReLU(x) = max(0,x);

[0119] Sigmoid function: Pre-store the input-output correspondence through a lookup table (LUT), use the linear calculation result as an index to quickly query and output the corresponding nonlinear transformation value, and achieve approximate calculation of σ(x) = 1 / (1+ex).

[0120] For details, please refer to the previous diagram and combine Figure 15 As shown, the specific implementation method is to use the principle of light polarization state or transmittance, and each storage element in the storage array stores a weight value. For example, for a 4×4 storage matrix, each unit stores a weight value W in the matrix multiplication. ij (i, j = 1, 2, 3, 4), with a transmittance range of 0-100%, corresponding to a weight value of 0-1 (the original matrix value must be divided by the same coefficient to convert it to a number ≤ 1). Weight writing: The target storage element is selected through the row selector and column selector, and a specific voltage is applied to control the transmittance to write the weight value. For example, if the storage weight is 0.6, the transmittance of the storage element is controlled to 60%.

[0121] Input light intensity setting: Convert input data into light intensity signal X i (i=1,2,3,4), for example, if the input data is [0.2,0.3,0.4,0.5], the corresponding light intensity signal strength is set proportionally (such as 0.2 corresponds to light intensity I1, 0.3 corresponds to light intensity I2, etc.)

[0122] Multiplication and accumulation operation: The input light intensity signal is incident on the storage matrix, and the output light intensity Y of each storage element is ij =X i *W ij The light detection element (such as a photodiode) detects the light intensity output by each unit and converts it into an electrical signal. The electrical signals in the same column are accumulated by the accumulation circuit to obtain the matrix multiplication result Y j =∑ 7i＝1 =X i *W ij , realizing vector-matrix multiplication operation. That is, the accumulation circuit is connected to the column selector or the storage and calculation elements in the same group, and is used to accumulate the electrical signals output by the storage and calculation elements in the same column or the same group to realize the accumulation operation of the matrix multiplication results.

[0123] In summary, input data is input in the form of a light intensity signal. Using the principle of light transmittance (which is similar to the principle of light polarization), the input light intensity (X) is multiplied by the weight (W) represented by the transmittance of the storage and computing element. The output light intensity (Y = X * W) is converted into an electrical signal by the light detection element. The electrical signal is a current or voltage. The electrical signals of multiple storage and computing elements are accumulated to perform the accumulation operation of matrix multiplication. This method of performing matrix multiplication and addition directly on the array of storage and computing elements avoids the data transmission process from memory to computing units in traditional architectures, greatly improving the computing speed.

[0124] Activation function calculation: Activation functions are used to introduce nonlinear characteristics in deep learning. After completing linear calculations (such as matrix multiplication and convolution), integrated memory and computation chips need to perform activation function calculations. For simple activation functions, such as the ReLU (Rectified Linear Unit) function, where ReLU(x) = max(0, x), integrated memory and computation chips can implement this function by controlling the memory and computation components. For example, after the light intensity in the memory and computation component is converted into an electrical signal, a comparator circuit compares the output of the linear calculation with a set threshold. If the output value exceeds the threshold, the value is directly output; if it is less than the threshold, the output is 0. For complex activation functions, such as the Sigmoid function, where σ(x) = 1 / (1 + ex), the chip can use a lookup table to pre-store the Sigmoid function output values ​​corresponding to different input values ​​in a specific memory area on the chip. During calculations, the linear calculation output value is used as an index to quickly retrieve the corresponding activation function output value Y from the lookup table. This improves the efficiency of the activation function while ensuring accuracy. In addition, Y can also be directly obtained through peripheral calculation, and then the polarization state or light intensity of the light can be directly controlled to achieve the output control target Y.

[0125] Different X values ​​will produce different output values ​​Y, thus establishing a corresponding relationship between X and Y. First, Y / X is converted to a number where Y = < 1. If Y / X does not satisfy this, all Y / X values ​​are further divided by the same coefficient. This allows the activation function value to be expressed through the transmittance of the computational element. This requires a separate computational matrix layer to implement the activation function layer.

[0126] Pooling is a common technique in deep learning, primarily used to downsample data to reduce the amount of data and computation while preserving important feature information. The following describes different scenarios, each requiring a separate layer of matrix storage to implement the pooling layer's functionality. In practice, this should not be limited to the following scenarios; flexible adjustments should be made based on the application scenario. Any process that achieves equivalent results to pooling should be considered acceptable.

[0127] Max Pooling: In each pooling window, the maximum value within the window is taken as the output of the window. This method can highlight significant features in the data, such as edges and corners in the image, because these features usually correspond to large values.

[0128] In this case, the position of the maximum value in each pooled unit can be found, and only this position is transparent (this transmittance should be kept the same for all pooling), and other positions are not transparent.

[0129] Average Pooling: Calculates the average value of the elements within each pooling window as the output. It can smooth the data and reduce the impact of noise. It is more effective for data with uniform distribution characteristics.

[0130] In this case, the transmittance in all pooling units can be set to 50%.

[0131] Stochastic Pooling: Randomly selects elements as output within the pooling window according to a certain probability distribution. It introduces randomness during training and helps improve the generalization ability of the model.

[0132] In this case, a position in the pooling unit is randomly selected and light is transmitted to it, while other positions are not.

[0133] In other pooling situations, based on the target pooling value that needs to be achieved, the polarization state or light intensity of the light is directly adjusted after peripheral calculation to achieve the output target pooling value.

[0134] The present invention has a long service life and stable equipment performance: According to relevant literature records, liquid crystal and photochromic glass have good fatigue resistance and can quickly write and erase information. According to relevant research, the characteristic of photochromic glass is that it is not easy to fatigue and still does not fail after experiencing more than 300,000 changes in light and dark. Liquid crystal materials can usually withstand more than 100,000 or even higher orders of magnitude of state changes driven by electric fields. In the field of storage technology, flash memory is widely used for its excellent performance, and its cycle number can reach 10 3 -10 7Therefore, from the perspective of storage, the life of the material used in this invention has the characteristics of an excellent storage device. These are all not available in traditional memristor elements. It is reported that currently, there is research on the use of zinc oxide nanorods (ZnO-NRs) to introduce an oxygen vacancy layer into the Au / ZnO-NRs / AZO structure, which effectively improves the switching ratio (from ≈10 3 Increase to ≈10 4 ) and switching stability. The memristor field has long strived to match the number of operations achieved by flash memory, and achieving the same endurance achieved by the materials used in this invention remains a significant challenge. As the number of operations increases, memristor components suffer from poor cycling stability, a major drawback of memristors. This problem can cause the memristor to degenerate to a near-fixed resistor function, ultimately losing its memory resistance function.

[0135] The manufacturing process of the present invention is relatively mature, simple and low-cost: the manufacturing process of the storage unit can be realized by slightly improving the existing technology, and is suitable for large-scale production.

[0136] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A storage and computing element, characterized in that: include: At least one light-regulating medium layer, which is a liquid crystal layer and / or an electrochromic layer; used to realize data storage and calculation through changes in optical properties; A transparent conductive layer is provided on both sides of the dielectric layer and is used to apply an electric field to regulate the polarization rotation angle or transmittance of the dielectric layer; A photoelectric detection module, including a photodiode or a photodetector, is used to detect the intensity of light passing through the dielectric layer and convert it into an electrical signal; The liquid crystal layer type storage and computing element is provided with a polarizing layer, which is provided on the incident side and / or the exit side of the liquid crystal layer and is used to control the polarization state of light and cooperate with the dielectric layer to realize light polarization state modulation; The liquid crystal layer realizes the output light intensity = incident light intensity × rotation angle / 90 or incident light intensity × ((90-rotation angle) / 90) through polarization rotation angle, and the electrochromic layer realizes the output light intensity = incident light intensity × transmittance through transmittance.

2. The storage and computing element according to claim 1, characterized in that: The dielectric layer is connected to an external control circuit, which regulates the polarization rotation angle of the dielectric layer or the transmittance of the electrochromic layer by applying voltage to the transparent conductive layer. The data detection module is a light detection element such as a photodiode or a photodetector.

3. The storage and computing element according to claim 1, wherein: The polarizing layer is a grating structure, forming a polarizer that only allows light of a specific polarization state to pass through, and is attached to both sides of the liquid crystal layer.

4. A storage and calculation method, based on the storage and calculation element according to claim 1, characterized in that: The following steps are involved: Step 1: If the memory computing element is a liquid crystal layer type, an initial light beam is emitted from a light source, filtered by a polarizing layer to a single polarization state, and then incident on the liquid crystal layer; if the memory computing element is an electrochromic layer type, the polarization step is omitted and the initial light beam is directly incident on the electrochromic layer; Step 2: The control circuit applies a set voltage to the transparent conductive layer to adjust the polarization rotation angle or transmittance of the dielectric layer, thereby achieving a multiplication operation of the incident light intensity and the control coefficient; Step 3: The data detection module converts the outgoing light intensity into an electrical signal, which is then output through analog-to-digital conversion or directly as an analog signal.

5. A storage and calculation method according to claim 4, characterized in that: The polarization rotation angle of the medium layer is regulated to be 0-90 degrees, and the light transmittance of the medium layer is regulated to be 0-100%.

6. A storage and computing array, characterized in that: comprising a plurality of the memory-computation elements according to claim 1, wherein the plurality of memory-computation elements are arranged in a two-dimensional matrix and connected to a row selector and a column selector via vertically intersecting row lines and column lines; A light splitting device can be set between adjacent storage and computing element layers to split the light into a calculation beam and a feedback detection beam; The storage and calculation element realizes vector-matrix multiplication through light intensity multiplication operation by regulating the polarization state or transmittance of light and storing weight values.

7. The storage and computing array according to claim 6, characterized in that: The storage and computing elements are arranged in a two-dimensional matrix and are vertically cross-wired through metal wires. The row lines and column lines are connected to the row selector and column selector respectively to achieve precise addressing of a single storage and computing element.

8. The storage and computing array according to claim 7, characterized in that: It also includes an accumulation circuit, which is connected to the column selector or the storage and calculation elements of the same group, and is used to accumulate the electrical signals output by the storage and calculation elements of the same column or the same group to achieve the accumulation operation of the matrix multiplication results.

9. An application method of a storage and computing array, based on the storage and computing array according to claim 6, characterized in that The following steps are involved: Data input: The input data is converted into a light intensity signal by the light source. If it is a liquid crystal layer type, it is incident on the storage and computing array through the polarizing layer; Weight calculation: The row selector and column selector select the target storage element and realize the multiplication of input light intensity and weight by adjusting the transmittance; Result accumulation: The output electrical signals of the storage and calculation elements in the same column are summed through the accumulation circuit to obtain the intermediate result of matrix multiplication; Non-linear activation: Through independent activation layer storage elements, the intermediate results are non-linearly transformed using comparators (such as ReLU) or lookup tables (such as Sigmoid); Signal output: The final electrical signal is converted from digital to analog or directly output as an analog signal.

10. The application method of the storage and computing array according to claim 9, characterized in that: The nonlinear activation step is specifically as follows: ReLU function: Compares the linear calculation result with the preset threshold through a comparator circuit. When the result is greater than the threshold, the original value is output. The corresponding formula is ReLU(x) = max(0,x); Sigmoid function: Pre-store the input-output correspondence through a lookup table (LUT), use the linear calculation result as an index to quickly query and output the corresponding nonlinear transformation value, and achieve approximate calculation of σ(x) = 1 / (1+ex).

Images