An in-memory computing array structure based on spin wave units and its control method

Abstract

The present application discloses an in-memory computing array structure based on spin wave units and a control method thereof, relating to the field of integrated circuits. The in-memory computing array structure based on spin wave units comprises: a plurality of interconnected in-memory computing units; each in-memory computing unit comprises a spin wave unit and a transistor unit connected to the spin wave unit; the spin wave unit comprises a spin wave body, the spin wave body comprising a spin wave signal generating end and a spin wave signal detecting end disposed opposite each other; a first magnetic domain wall driving unit and a second magnetic domain wall driving unit disposed on adjacent sides of the spin wave body, respectively, away from the spin wave signal generating end and the spin wave signal detecting end; magnetic domain walls within the spin wave body; and achieving preset weights in Hopfield neural network calculations. Due to the scattering and reflection of spin waves during propagation and the fluctuation of the magnetization of the magnetic domain walls themselves, the reliability and stability of spintronic devices and in-memory computing are improved.

Description

Technical Field

[0001] The present application relates to the field of integrated circuits, and in particular to an in-memory computing array structure based on spin wave units and a control method thereof. Background Art

[0002] The Hopfield network is a single-layer feedback neural network, commonly used in combinatorial optimization and associative memory methods. As a recursive neural network, the Hopfield network can usually converge stably, but it may also converge to a non-global minimum, making it impossible to obtain a global optimal solution.

[0003] With the advancement of technology, research on using spintronics devices to achieve integrated storage and computing capabilities has become increasingly mature. By combining new spintronic devices such as spin-orbit torque (SOT) or spin-transfer torque (STT) magnetic random access memory (MRAM) with transistors, basic Boolean logic calculations can be achieved. Due to the excellent speed and durability of spintronics devices, hybrid circuits combining spintronics and transistors, with the assistance of external circuits, can perform logical calculations in memory, integrating storage and computing functions and avoiding the memory wall problem in traditional von Neumann architectures.

[0004] However, since a single spintronic device can only store data in a binary form, STT-MRAM requires a large current to flip the magnetic moment, and SOT-MRAM also requires an external magnetic field to switch the resistance state. This makes spintronic devices unfavorable for large-scale integration, reducing their reliability and stability. The existing Hopfield network model cannot achieve a globally optimal solution, reducing the reliability and application scenarios of in-memory computing. Summary of the Invention

[0005] The purpose of this application is to provide an in-memory computing array structure based on spin wave units and its control method to solve the problems that existing spin electronic devices are not conducive to large-scale integration, reduce the reliability and stability of spin electronic devices, and the existing Hopfield network model cannot achieve a global optimal solution, which reduces the reliability and application scenarios of in-memory computing.

[0006] In a first aspect, the present application provides an in-memory computing array structure based on a spin wave unit and a control method thereof, wherein the in-memory computing array structure comprises:

[0007] Multiple in-memory computing units connected to each other;

[0008] Each of the in-memory computing units includes a spin wave unit and a transistor unit connected to the spin wave unit;

[0009] The spin wave unit includes a spin wave body, and the spin wave body includes a spin wave signal generating end and a spin wave signal detecting end that are arranged opposite to each other;

[0010] The spin wave unit further includes: a first magnetic domain wall driving unit and a second magnetic domain wall driving unit respectively disposed on two adjacent sides of the spin wave body away from the spin wave signal generating end and the spin wave signal detecting end, wherein the first magnetic domain wall driving unit and the second magnetic domain wall driving unit are disposed opposite to each other;

[0011] The spin wave unit further includes: a magnetic domain wall disposed between the first magnetic domain wall driving unit and the second magnetic domain wall driving unit along a direction of the first magnetic domain wall driving unit, and within the spin wave body;

[0012] The spin wave unit is configured to determine a target position at which the magnetic domain wall stops moving between the first magnetic domain wall drive unit and the second magnetic domain wall drive unit, and generate a corresponding preset weight based on a preset weight control signal sent by the transistor unit to the spin wave signal generating end, the spin wave signal detecting end, the first magnetic domain wall drive unit, and the second magnetic domain wall drive unit, when the first magnetic domain wall drive unit and the second magnetic domain wall drive unit drive the magnetic domain wall to move.

[0013] In the case of adopting the above technical solution, an embodiment of the present application provides an in-memory computing array structure based on a spin wave unit, the in-memory computing array structure comprising: a plurality of interconnected in-memory computing units; each of the in-memory computing units comprising a spin wave unit and a transistor unit connected to the spin wave unit; the spin wave unit comprising a spin wave body, the spin wave body comprising a spin wave signal generating end and a spin wave signal detecting end disposed oppositely; the spin wave unit further comprising: a first magnetic domain wall driving unit and a second magnetic domain wall driving unit disposed on adjacent sides of the spin wave body away from the spin wave signal generating end and the spin wave signal detecting end, respectively, the first magnetic domain wall driving unit and the second magnetic domain wall driving unit being disposed oppositely; the spin wave unit further comprising: a magnetic domain wall disposed between the first magnetic domain wall driving unit and the second magnetic domain wall driving unit along a direction of the first magnetic domain wall driving unit and within the spin wave body; the spin wave unit is configured to, when the first magnetic domain wall driving unit and the second magnetic domain wall driving unit drive the magnetic domain wall to move, Based on preset weight control signals sent by the transistor unit to the spin wave signal generating terminal, the spin wave signal detecting terminal, the first magnetic domain wall driving unit, and the second magnetic domain wall driving unit, the target position at which the magnetic domain wall stops moving between the first magnetic domain wall driving unit and the second magnetic domain wall driving unit is determined, and a corresponding preset weight is generated. This ensures that the weights in the Hopfield neural network calculation reach the preset weights. Due to the scattering and reflection of spin waves during propagation, as well as the inherent fluctuations in the magnetization of the magnetic domain walls, the weights of the in-memory computing array structure can fully utilize the coupling perturbation when reading. This, in turn, avoids the drawback of seeking local optimal solutions in the Hopfield neural network calculation during the iterative process, allowing the global optimal solution to be determined. Furthermore, the weight representation achieved through magnetic domain wall magnetization perturbations has low energy consumption. The in-memory computing array structure described in this application has simple peripheral circuitry, high integration, compatibility with existing mainstream CMOS processes, and ease of large-scale manufacturing and integrated application to solve specific in-memory computing optimization tasks, thereby improving the reliability and stability of spintronic devices and in-memory computing.

[0014] In a possible implementation, the number of the in-memory computing units is determined based on the computational scale of the traveling salesman problem.

[0015] In one possible implementation, the plurality of in-memory computing units are configured to use the corresponding preset weights as weights of the Hopfield neural network when performing calculations on the traveling salesman problem, and determine the target optimal solution based on the preset weights.

[0016] In a possible implementation, the transistor unit includes a first transistor, a second transistor, and a third transistor, wherein a source of the first transistor is connected to the first magnetic domain wall driving unit, a drain of the first transistor is connected to a bit line, and a gate of the first transistor is connected to a first word line;

[0017] The source of the second transistor is connected to the spin wave signal generating terminal, the drain of the second transistor is connected to the bit line, and the gate of the second transistor is connected to the second word line;

[0018] A source of the third transistor is connected to the second magnetic domain wall driving unit, a drain of the third transistor is connected to the bit line, and a gate of the third transistor is connected to a third word line;

[0019] The spin wave signal detection end is connected to the source line.

[0020] In a possible implementation, current, spin waves, or an external magnetic field are applied to the first magnetic domain wall driving unit and the second magnetic domain wall driving unit to enable the first magnetic domain wall driving unit and the second magnetic domain wall driving unit to drive the magnetic domain wall to move.

[0021] In a possible implementation, the material composition of the spin wave body that stabilizes the magnetic domain wall includes iron, cobalt, nickel and / or iron-cobalt-nickel alloy.

[0022] In a second aspect, the present application further provides a method for controlling an in-memory computing array structure based on a spin wave unit, characterized in that the method is applied to any in-memory computing array structure based on a spin wave unit according to the first aspect, and the method comprises:

[0023] When the first magnetic domain wall driving unit and the second magnetic domain wall driving unit drive the magnetic domain wall to move, the spin wave unit determines a target position at which the magnetic domain wall stops moving between the first magnetic domain wall driving unit and the second magnetic domain wall driving unit based on a preset weight control signal sent by the transistor unit to the spin wave signal generating end, the spin wave signal detecting end, the first magnetic domain wall driving unit, and the second magnetic domain wall driving unit, and generates a corresponding preset weight.

[0024] In a possible implementation, the method further includes:

[0025] When performing calculations on the traveling salesman problem, the plurality of in-memory computing units use the corresponding preset weights as weights of the Hopfield neural network, and determine the target optimal solution based on the preset weights.

[0026] In one possible implementation, before the spin wave unit determines a target position at which the magnetic domain wall stops moving between the first magnetic domain wall driving unit and the second magnetic domain wall driving unit based on a preset weight control signal sent by the transistor unit to the spin wave signal generating end, the spin wave signal detecting end, the first magnetic domain wall driving unit, and the second magnetic domain wall driving unit, and generates the corresponding preset weight, the method further includes:

[0027] The spin wave unit determines an input intensity value of the spin wave signal by inputting a spin wave signal of a preset frequency and a preset intensity to the spin wave signal generating end when the first magnetic domain wall driving unit and the second magnetic domain wall driving unit drive the magnetic domain wall to move;

[0028] The spin wave signal detection end determines an output intensity value corresponding to the spin wave signal;

[0029] The spinning wave unit determines a current weight based on the input intensity value and the output intensity value;

[0030] The spinning wave unit determines the preset weight control signal based on the current weight.

[0031] In one possible implementation, after the spin wave unit determines a target position at which the magnetic domain wall stops moving between the first magnetic domain wall driving unit and the second magnetic domain wall driving unit based on a preset weight control signal sent by the transistor unit to the spin wave signal generating end, the spin wave signal detecting end, the first magnetic domain wall driving unit, and the second magnetic domain wall driving unit, and generates a corresponding preset weight, the method further includes:

[0032] Controlling the plurality of in-memory computing units to store the corresponding current weights;

[0033] The target weight is determined based on the multiple current weights to complete the regulation of the weight size.

[0034] The beneficial effects of the control method of the in-memory computing array structure based on spin wave units provided in the second aspect are the same as the beneficial effects of the in-memory computing array structure based on spin wave units described in the first aspect or any possible implementation of the first aspect, and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0036] Figure 1A circuit diagram of an in-memory computing array structure based on a spin wave unit provided in an embodiment of the present application is shown;

[0037] Figure 2 A schematic structural diagram of a spin wave unit provided in an embodiment of the present application is shown;

[0038] Figure 3 A schematic diagram showing an optimization result of an in-memory computing array structure based on spin wave units provided in an embodiment of the present application when used to solve the traveling salesman problem;

[0039] Figure 4 A schematic diagram of a simulation of the propagation of spin wave signals in a device at different magnetic domain wall positions using multi-physics field software provided in an embodiment of the present application is shown;

[0040] Figure 5 A schematic diagram showing the fast Fourier transform results of a spin wave signal received by a receiving end when a 1 MHz input signal and magnetic domain wall positions are different, provided by an embodiment of the present application.

[0041] Reference numerals

[0042] In-memory computing unit 10; spin wave unit 101; transistor unit 102; spin wave body 1011; spin wave signal generating terminal 1011A; spin wave signal detecting terminal 1011B; spin wave body component 1011C; first magnetic domain wall driving unit 1012; second magnetic domain wall driving unit 1013; magnetic domain wall 1014; first transistor 1021; second transistor 1022; third transistor 1023; bit line BL; first word line WL 11 ; Second word line - WL 12 ; Third word line - WL 13 ; Source line-SL. DETAILED DESCRIPTION

[0043] To facilitate the clear description of the technical solutions of the embodiments of the present application, in the embodiments of the present application, words such as "first" and "second" are used to distinguish between identical or similar items with substantially the same functions and effects. For example, the first threshold and the second threshold are merely used to distinguish between different thresholds and do not limit their order. Those skilled in the art will understand that words such as "first" and "second" do not limit the quantity or execution order, and words such as "first" and "second" do not necessarily mean that they are different.

[0044] It should be noted that, in this application, words such as "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described in this application as "exemplary" or "for example" should not be construed as being preferred or advantageous over other embodiments or designs. Rather, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0045] In this application, "at least one" means one or more, and "more" means two or more. "And / or" describes the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can mean: the existence of A alone, the existence of A and B at the same time, and the existence of B alone, where A and B can be singular or plural. The character " / " generally indicates that the previous and next associated objects are in an "or" relationship. "At least one of the following items" or similar expressions refers to any combination of these items, including any combination of single items or plural items. For example, at least one of a, b or c can mean: a, b, c, the combination of a and b, the combination of a and c, the combination of b and c, or the combination of a, b and c, where a, b, c can be single or multiple.

[0046] Figure 1 A circuit diagram of an in-memory computing array structure based on a spin wave unit provided in an embodiment of the present application is shown. The in-memory computing array structure includes:

[0047] a plurality of in-memory computing units 10 connected to each other;

[0048] Each of the in-memory computing units 10 includes a spin wave unit 101 and a transistor unit 102 connected to the spin wave unit;

[0049] The spin wave unit 101 includes a spin wave body 1011, and the spin wave body 1011 includes a spin wave signal generating end 1011A and a spin wave signal detecting end 1011B arranged opposite to each other, and also includes a spin wave body component 1011C;

[0050] The spin wave unit 101 further includes: a first magnetic domain wall driving unit 1012 and a second magnetic domain wall driving unit 1013 respectively disposed on two adjacent sides of the spin wave body 1011 away from the spin wave signal generating end 1011A and the spin wave signal detecting end 1011B, wherein the first magnetic domain wall driving unit 1012 and the second magnetic domain wall driving unit 1013 are disposed opposite to each other;

[0051] The spin wave unit 101 further includes: a magnetic domain wall 1014 disposed between the first magnetic domain wall driving unit 1012 and the second magnetic domain wall driving unit 1013 along the direction of the first magnetic domain wall driving unit 1012 , and within the spin wave body 1011 .

[0052] The spin wave unit 101 is used to determine the target position where the magnetic domain wall 1014 stops moving between the first magnetic domain wall driving unit 1012 and the second magnetic domain wall driving unit 1013 based on the preset weight control signal sent by the transistor unit 102 to the spin wave signal generating terminal 1011A, and generate a corresponding preset weight when the first magnetic domain wall driving unit 1012 and the second magnetic domain wall driving unit 1013 drive the magnetic domain wall 1014 to move.

[0053] In the present application, current, spin waves, or an external magnetic field can be applied to the first and second domain wall drive units to drive the domain walls to move. A corresponding preset weight is generated by increasing or decreasing the intensity ratio between the spin wave signal generating end and the spin wave signal detecting end. This means that the weights in the Hopfield neural network calculation can reach the preset weights. Due to the scattering and reflection of spin waves during propagation and the fluctuations in the magnetization of the domain walls themselves, the weights of the in-memory computing array structure can fully utilize the coupling disturbance when reading. This allows the iterative process to avoid the drawback of obtaining a local optimal solution in the Hopfield neural network calculation and determine a global optimal solution. The in-memory computing array structure described in the present application is compatible with existing mainstream complementary metal oxide semiconductor (CMOS) processes and is easily adaptable to large-scale manufacturing and integrated applications to solve specific in-memory computing optimization tasks.

[0054] The first and second magnetic domain wall drive units can be driven to move the magnetic domain walls by applying current, spin waves, or an external magnetic field to the first and second magnetic domain wall drive units. When spin waves are used to control the positions of the magnetic domain walls, the spin wave units do not generate Joule heat, which can greatly reduce the power consumption of the device.

[0055] An embodiment of the present application provides an in-memory computing array structure based on a spin wave unit, the in-memory computing array structure comprising: a plurality of interconnected in-memory computing units; each of the in-memory computing units comprising a spin wave unit and a transistor unit connected to the spin wave unit; the spin wave unit comprising a spin wave body, the spin wave body comprising a spin wave signal generating end and a spin wave signal detecting end disposed opposite to each other; the spin wave unit further comprising: a first magnetic domain wall driving unit and a second magnetic domain wall driving unit disposed on adjacent sides of the spin wave body away from the spin wave signal generating end and the spin wave signal detecting end, the first magnetic domain wall driving unit and the second magnetic domain wall driving unit being disposed opposite to each other; the spin wave unit further comprising: a magnetic domain wall disposed between the first magnetic domain wall driving unit and the second magnetic domain wall driving unit along a direction of the first magnetic domain wall driving unit and within the spin wave body; the spin wave unit being configured to, when the first magnetic domain wall driving unit and the second magnetic domain wall driving unit drive the magnetic domain wall to move, The tube unit sends preset weight control signals to the spin wave signal generating end, the spin wave signal detecting end, the first magnetic domain wall driving unit, and the second magnetic domain wall driving unit to determine the target position at which the magnetic domain wall stops moving between the first magnetic domain wall driving unit and the second magnetic domain wall driving unit, and generates corresponding preset weights. This ensures that the weights in the Hopfield neural network calculation reach the preset weights. Due to the scattering and reflection of spin waves during propagation, as well as the inherent fluctuations in the magnetization of the magnetic domain walls, the weights of the in-memory computing array structure can fully utilize the coupling perturbation when reading. This can avoid the drawback of seeking a local optimal solution in the Hopfield neural network calculation during the iterative process, and can determine a global optimal solution. The weight representation achieved through magnetic domain wall magnetization perturbation has low energy consumption. The in-memory computing array structure described in this application has simple peripheral circuitry, high integration, compatibility with existing mainstream CMOS processes, and ease of large-scale manufacturing and integrated application to solve specific in-memory computing optimization tasks, thereby improving the reliability and stability of spintronic devices and in-memory computing.

[0056] Figure 2 FIG. 1 shows a schematic structural diagram of a spin wave unit provided in an embodiment of the present application, such as Figure 2 As shown, the spin wave unit 101 includes a spin wave body 1011, and the spin wave body 1011 includes a spin wave signal generating end 1011A and a spin wave signal detecting end 1011B that are arranged opposite to each other;

[0057] The spin wave unit 101 further includes: a first magnetic domain wall driving unit 1012 and a second magnetic domain wall driving unit 1013 respectively disposed on two adjacent sides of the spin wave body 1011 away from the spin wave signal generating end 1011A and the spin wave signal detecting end 1011B, wherein the first magnetic domain wall driving unit 1012 and the second magnetic domain wall driving unit 1013 are disposed opposite to each other;

[0058] The spin wave unit 101 further includes: a magnetic domain wall 1014 disposed between the first magnetic domain wall driving unit 1012 and the second magnetic domain wall driving unit 1013 along the direction of the first magnetic domain wall driving unit 1012 , and within the spin wave body 1011 .

[0059] In this application, the magnetic domain wall is a magnetic structure with a certain width. There is no limitation on the specific width value, and it can be adjusted according to the actual application scenario.

[0060] Magnetic domain walls can serve as channels for spin wave propagation. Within these channels, spin waves propagate in a directionally controlled manner. The position of the domain walls can be controlled by a first and second magnetic domain wall driving units, allowing specific spin waves generated by a spin wave signal generator to propagate to varying degrees. For example, when the domain wall is located between the spin wave signal generator and the spin wave signal detector, the spin waves emitted by the spin wave signal generator can propagate directionally along the domain wall. The spin wave signals emitted by the spin wave signal generator can be received by the spin wave signal detector, minimizing spin wave signal loss. For another example, when the domain wall is located far from the spin wave signal generator and the spin wave signal detector, a channel for directional spin wave propagation is lacking between the two terminals. The spin wave signal of a specific frequency emitted by the spin wave signal generator will propagate non-directionally within the entire spin wave body, and scattering and reflection will occur during the propagation process. The spin wave signal with the same frequency as the excitation source signal will suffer a large degree of loss, and the intensity of the frequency signal received by the spin wave signal detection end will also be very weak.

[0061] Therefore, the position of the magnetic domain wall can affect the spin wave intensity received by the spin wave signal detection end. The ratio between the intensity value of the spin wave signal of a specific frequency excited by the spin wave signal generation end and the intensity value of the spin wave signal of the same frequency received by the spin wave signal detection end can be defined as the weight represented by the spin wave unit. The device weight can then be controlled by changing the position of the magnetic domain wall.

[0062] Optionally, current, spin waves or external magnetic fields are applied to the first magnetic domain wall driving unit and the second magnetic domain wall driving unit to enable the first magnetic domain wall driving unit and the second magnetic domain wall driving unit to drive the magnetic domain wall to move.

[0063] In the present application, the material composition of the spin wave body that maintains the stability of the magnetic domain wall includes iron, cobalt, nickel and / or iron-cobalt-nickel alloy. The material parameters corresponding to the magnetic domain wall material can be modulated to ensure that the magnetic domain wall can exist stably in the spin wave body.

[0064] Optionally, the iron-cobalt-nickel alloy may be Permalloy.

[0065] Optionally, the material system of the magnetic domain wall is a ferromagnet or antiferromagnet with magnetic order, and the lower the damping coefficient, the better the effect, such as YIG (yttrium iron garnet) and other materials.

[0066] Optionally, the number of the in-memory computing units is determined based on the computational scale of the traveling salesman problem, and each in-memory computing unit may be configured to store a preset weight to facilitate solving the problem.

[0067] The plurality of in-memory computing units are used to use the corresponding preset weights as weights of the Hopfield neural network when performing calculations on the traveling salesman problem, and to determine the target optimal solution based on the preset weights.

[0068] Optional, see Figure 1 The transistor unit 102 includes a first transistor 1021, a second transistor 1022, and a third transistor 1023. The source of the first transistor 1021 is connected to the first magnetic domain wall driving unit 1012, and the connection is a 4-port. The drain of the first transistor 1021 is connected to the bit line BL, and the gate of the first transistor 1021 is connected to the first word line WL. 11 connect;

[0069] The source of the second transistor 1022 is connected to the spin wave signal generating terminal 1011A, which is a port 1. The drain of the second transistor 1022 is connected to the bit line BL. The gate of the second transistor 1022 is connected to the second word line WL. 12 connect;

[0070] The source of the third transistor 1023 is connected to the second magnetic domain wall driving unit 1013, and the connection is a two-port connection. The drain of the third transistor 1023 is connected to the bit line BL, and the gate of the third transistor 1023 is connected to the third word line WL. 13 connect;

[0071] The spin wave signal detection terminal 1011B is connected to the source line SL, and the connection point is a 3-port.

[0072] Among them, port 1 and port 3 act as converters between electrical signals and spin wave signals. When the spin wave signal is detected, the bit line (BL) is grounded and the second word line WL is grounded. 12 Set the level high, the transistor is turned on, and when port 1 receives an electrical signal under the control of the transistor, it converts the electrical signal into a spin wave signal.

[0073] When the source line SL is at a high level, port 3 receives the spin wave signal, converts the spin wave signal into an electrical signal, and reads the electrical signal through an external device. The electrical signal is then aggregated and compared with the spin wave signal from port 1. The ratio of the spin wave signal intensities at a certain frequency between the spin wave signal generation end and the spin wave signal detection end is defined as the device weight.

[0074] When the bit line (BL) is grounded, the first word line WL 11 When set to a high level, the transistor is turned on and the four ports start working, converting the electrical signal into a suitable current or spin wave signal in order to achieve the control of the position of the magnetic domain wall.

[0075] Similarly, port 2 can be on the bit line BL and the third word line WL 13 The same function can be achieved under the control of . The regulation of 2 ports and 4 ports can make the magnetic domain wall reach the corresponding position.

[0076] Figure 3 A schematic diagram shows the optimization results of an in-memory computing array structure based on spin wave units provided by an embodiment of the present application when used to solve the traveling salesman problem. When using the Hopfield neural network to solve the traveling salesman problem, if there is no disturbance, it will often fall into a local optimal solution, such as Figure 3 (a) is shown. After a certain number of iterations, the energy tends to be stable and does not change, that is, it falls into a local optimal solution and cannot obtain a global optimal solution. If random fluctuations are added during the device reading process, then during the iteration process, using the device array for solution can make Hopfield jump out of the local optimal solution and continue to iterate to obtain the lowest point of the Hopfield neural network energy, and obtain the global optimal solution. The optimal result is shown in Figure 3 (b) shows the optimal solution path for the traveling salesman problem of 8 cities.

[0077] Furthermore, when the spin wave signal propagates from the generator to the detector, it will be constantly reflected and scattered in the medium, causing some signal loss and disturbance. In addition, when solving the traveling salesman problem, since the magnetization direction itself also has disturbances, there will be random disturbances in the reading process of the device weight. Figure 3 , Figure 3 The result of (c) is the relationship between the number of iterations and energy in the process of solving the traveling salesman problem after introducing Gaussian fluctuations with a mean of 0 and a standard deviation of 2.

[0078] Figure 4 FIG1 shows a schematic diagram of a simulation of the propagation of spin wave signals in a device at different magnetic domain wall positions using multi-physics field software provided by an embodiment of the present application. Figure 4 (a) shows the distribution of the magnetic moment Z component in the device before the spin wave detection signal is input into the device. At this time, the magnetic domain wall is located in the middle of the device (y=200 nm). Figure 4 (b) shows the process of the spin wave signal propagating in a directional manner along the magnetic domain wall when the magnetic domain wall is in the middle of the device. It can be clearly seen that the crests and troughs of the spin wave propagating in a directional manner appear alternately along the propagation path. Figure 4 (c) shows the distribution of the magnetic moment Z component in the device before the spin wave detection signal is input into the device, where the magnetic domain wall y = 100 nm. Figure 4 Figure (d) shows the overall magnetic moment distribution of the device when a spin wave is input from spin wave signal generator 1011A and received by spin wave signal detector 1011B, when the domain wall distance y = 100 nm. It can be seen that there are no distinct spin wave peaks or troughs between spin wave signal generator 1011A and spin wave signal detector 1011B. The extreme value of the y-direction component of the magnetic moment occurs near the domain wall. The spin wave signal received by spin wave signal detector 1011B is not significant.

[0079] Figure 5 The figure shows a schematic diagram of the fast Fourier transform result of the spin wave signal received by the receiving end when the magnetic domain wall position is different when a 1 MHz signal is input, provided by an embodiment of the present application. Figure 5 (a) It can be seen that when the magnetic domain wall is exactly between the spin wave signal generation end and the input end, the frequency of the received signal is basically around 1 GHz, and the intensity is about 0.02. Figure 5 (b) As can be seen, when the domain wall is located far from the spin wave generator and input, the received 1 GHz signal is significantly lost, with the intensity being approximately 0.001. By comparing the ratio of the spin wave signal intensity at the input and output for two different cases, we find that the ratios are 0.7484 and 0.0436 for domain wall positions of y = 200 nm and y = 100 nm, respectively. In summary, the domain wall position can effectively control the weight of a spin wave device.

[0080] The in-memory computing array structure based on spin wave units described in this application can use the Hopfield network as a computing model, the spin wave unit as a hardware carrier, and the classic traveling salesman problem as the solution target. By increasing the weight fluctuation through the fluctuation of the magnetic moment itself and the scattering and reflection of the spin wave, the optimal solution to the traveling salesman problem is obtained.

[0081] The in-memory computing array structure based on spin wave units described in this application has a controllable initial maze domain magnetic texture structure, that is, a spin wave unit including a magnetic domain wall. The magnetic domain wall is driven to move by spin waves, current or voltage, and the weight increase and decrease update mapping in the Hopfield network calculation is realized by increasing or decreasing the intensity ratio between the spin wave signal generating end and the spin wave signal detecting end. When the integrated device constitutes the Hopfield weight array, that is, the in-memory computing array structure, due to the scattering and reflection of the spin wave during the propagation process and the fluctuation of the random magnetization of the magnetic domain and domain wall itself, the weight of the device fully utilizes the coupling disturbance when reading, and then gets rid of the disadvantage of obtaining the local optimal solution in the Hopfield network calculation during the iterative process, and obtains the global optimal solution. The device and array of this application are compatible with the existing mainstream CMOS process, and are easy to manufacture and integrate on a large scale to solve specific in-memory computing optimization tasks.

[0082] An embodiment of the present application provides an in-memory computing array structure based on a spin wave unit, the in-memory computing array structure comprising: a plurality of interconnected in-memory computing units; each of the in-memory computing units comprising a spin wave unit and a transistor unit connected to the spin wave unit; the spin wave unit comprising a spin wave body, the spin wave body comprising a spin wave signal generating end and a spin wave signal detecting end disposed opposite to each other; the spin wave unit further comprising: a first magnetic domain wall driving unit and a second magnetic domain wall driving unit disposed on adjacent sides of the spin wave body away from the spin wave signal generating end and the spin wave signal detecting end, the first magnetic domain wall driving unit and the second magnetic domain wall driving unit being disposed opposite to each other; the spin wave unit further comprising: a magnetic domain wall disposed between the first magnetic domain wall driving unit and the second magnetic domain wall driving unit along a direction of the first magnetic domain wall driving unit and within the spin wave body; the spin wave unit being configured to, when the first magnetic domain wall driving unit and the second magnetic domain wall driving unit drive the magnetic domain wall to move, The tube unit sends preset weight control signals to the spin wave signal generating end, the spin wave signal detecting end, the first magnetic domain wall driving unit, and the second magnetic domain wall driving unit to determine the target position at which the magnetic domain wall stops moving between the first magnetic domain wall driving unit and the second magnetic domain wall driving unit, and generates corresponding preset weights. This ensures that the weights in the Hopfield neural network calculation reach the preset weights. Due to the scattering and reflection of spin waves during propagation, as well as the inherent fluctuations in the magnetization of the magnetic domain walls, the weights of the in-memory computing array structure can fully utilize the coupling perturbation when reading. This can avoid the drawback of seeking a local optimal solution in the Hopfield neural network calculation during the iterative process, and can determine a global optimal solution. The weight representation achieved through magnetic domain wall magnetization perturbation has low energy consumption. The in-memory computing array structure described in this application has simple peripheral circuitry, high integration, compatibility with existing mainstream CMOS processes, and ease of large-scale manufacturing and integrated application to solve specific in-memory computing optimization tasks, thereby improving the reliability and stability of spintronic devices and in-memory computing.

[0083] The embodiment of the present application provides a control method for an in-memory computing array structure based on a spin wave unit, characterized in that it is applied to Figure 1 In the in-memory computing array structure based on spin wave units, the method includes:

[0084] When the first magnetic domain wall driving unit and the second magnetic domain wall driving unit drive the magnetic domain wall to move, the spin wave unit determines a target position at which the magnetic domain wall stops moving between the first magnetic domain wall driving unit and the second magnetic domain wall driving unit based on a preset weight control signal sent by the transistor unit to the spin wave signal generating end, the spin wave signal detecting end, the first magnetic domain wall driving unit, and the second magnetic domain wall driving unit, and generates a corresponding preset weight.

[0085] It should be noted that the preset weight corresponding to the spin wave unit can be determined by the weight control function of the spin wave unit in advance, so that a preset weight control signal can be generated, and finally the spin wave unit can be controlled to be under the preset weight.

[0086] Specifically, in the present application, the implementation process of the weight control function of the spin wave unit may include the following sub-steps:

[0087] Sub-step S1: When the first magnetic domain wall driving unit and the second magnetic domain wall driving unit drive the magnetic domain wall to move, the spin wave unit determines an input intensity value of the spin wave signal by inputting a spin wave signal of a preset frequency and a preset intensity to the spin wave signal generating end.

[0088] A driving signal can be input to the first and second magnetic domain wall driving units via current, magnetic field, or spin waves to drive the movement of the magnetic domain walls. Simultaneously, a spin wave signal of a preset frequency and intensity can be input to the spin wave signal generating end to determine the input intensity value of the spin wave signal.

[0089] The embodiment of the present application does not limit the specific values ​​of the preset frequency and the preset intensity, and they can be adjusted according to the actual application scenario.

[0090] Sub-step S2: The spin wave signal detection end determines an output intensity value corresponding to the spin wave signal.

[0091] Sub-step S3: the spin wave unit determines a current weight based on the input intensity value and the output intensity value;

[0092] Sub-step S4: The spinning wave unit determines the preset weight control signal based on the current weight.

[0093] In the present application, the magnetic domain wall can serve as an effective channel for the directional propagation of spin waves. Therefore, when the position of the magnetic domain wall is different, the ratio of the intensity values ​​of the spin wave signal generation end and the spin wave signal detection end is different. In the present application, this ratio is positioned as a weight, and the device weight can be modulated by moving the position of the magnetic domain wall.

[0094] In the present application, when performing calculations on the traveling salesman problem, the plurality of in-memory computing units use the corresponding preset weights as weights of the Hopfield neural network, and determine the target optimal solution based on the preset weights.

[0095] Specifically, due to the disturbance in the magnetization direction of the magnetic domain wall itself, the spin wave will be reflected and scattered when propagating in the device. Therefore, there will be a certain degree of fluctuation when reading the preset weight. The corresponding preset weight can be used as the weight of the Hopfield neural network, so that the iteration of the Hopfield neural network can get rid of the local energy minimum point and obtain the target optimal solution, which is also the global optimal solution.

[0096] In the present application, multiple in-memory computing units can be controlled to store the corresponding current weights, and a target weight can be determined based on the multiple current weights to complete the regulation of the weight size.

[0097] The embodiment of the present application provides a control method for an in-memory computing array structure based on a spin wave unit. When the first magnetic domain wall driving unit and the second magnetic domain wall driving unit drive the magnetic domain wall to move, the spin wave unit determines the target position where the magnetic domain wall stops moving between the first magnetic domain wall driving unit and the second magnetic domain wall driving unit based on the preset weight control signal sent by the transistor unit to the spin wave signal generating end, the spin wave signal detecting end, the first magnetic domain wall driving unit, and the second magnetic domain wall driving unit, and generates the corresponding preset weight, that is, it can achieve the preset weight in the Hopfield neural network calculation. Since spin waves are scattered and reflected during propagation, and the magnetization of the magnetic domain wall itself fluctuates, the weights of the in-memory computing array structure can make full use of the coupling disturbance when reading, thereby getting rid of the disadvantage of obtaining a local optimal solution in the Hopfield neural network calculation during the iterative process, and determining the global optimal solution. In addition, the expression of the weights achieved through the magnetization disturbance of the magnetic domain wall has low energy consumption. The in-memory computing array structure described in the present application has simple peripheral circuits and high integration, is compatible with existing mainstream CMOS processes, and is easy to manufacture and integrate on a large scale to solve specific in-memory computing optimization tasks, thereby improving the reliability and stability of spin electronic devices and in-memory computing.

[0098] Although the present application is described herein in conjunction with various embodiments, in the process of implementing the claimed application, those skilled in the art may understand and implement other variations of the disclosed embodiments by reviewing the drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple situations. A single processor or other unit may implement several functions listed in the claims. Certain measures are recorded in different dependent claims, but this does not mean that these measures cannot be combined to produce good results.

[0099] Although the present application has been described with reference to specific features and embodiments thereof, it is apparent that various modifications and combinations may be made thereto without departing from the spirit and scope of the present application. Accordingly, this specification and the drawings are merely illustrative of the present application as defined by the appended claims and are deemed to cover any and all modifications, variations, combinations or equivalents within the scope of the present application. Obviously, those skilled in the art may make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, the present application is intended to include such modifications and variations as fall within the scope of the claims of the present application and their equivalents.

Claims

1. An in-memory computing array structure based on spin wave units, characterized in that: The in-memory computing array structure includes: a plurality of in-memory computing units connected to each other; Each of the in-memory computing units includes a spin wave unit and a transistor unit connected to the spin wave unit; The spin wave unit includes a spin wave body, and the spin wave body includes a spin wave signal generating end and a spin wave signal detecting end that are arranged opposite to each other; The spin wave unit further includes: a first magnetic domain wall driving unit and a second magnetic domain wall driving unit respectively disposed on two adjacent sides of the spin wave body away from the spin wave signal generating end and the spin wave signal detecting end, wherein the first magnetic domain wall driving unit and the second magnetic domain wall driving unit are disposed opposite to each other; The spin wave unit further includes: a magnetic domain wall disposed between the first magnetic domain wall driving unit and the second magnetic domain wall driving unit along a direction of the first magnetic domain wall driving unit, and within the spin wave body; The spin wave unit is configured to determine a target position at which the magnetic domain wall stops moving between the first magnetic domain wall drive unit and the second magnetic domain wall drive unit, and generate a corresponding preset weight based on a preset weight control signal sent by the transistor unit to the spin wave signal generating end, the spin wave signal detecting end, the first magnetic domain wall drive unit, and the second magnetic domain wall drive unit, when the first magnetic domain wall drive unit and the second magnetic domain wall drive unit drive the magnetic domain wall to move.

2. The in-memory computing array structure based on spin wave units according to claim 1, characterized in that: The number of the in-memory computing units is determined based on the computational scale of the traveling salesman problem.

3. The in-memory computing array structure based on spin wave units according to claim 2, characterized in that: The plurality of in-memory computing units are used to use the corresponding preset weights as weights of the Hopfield neural network when performing calculations on the traveling salesman problem, and to determine the target optimal solution based on the preset weights.

4. The in-memory computing array structure based on spin wave units according to claim 1, characterized in that: The transistor unit includes a first transistor, a second transistor, and a third transistor, wherein a source of the first transistor is connected to the first magnetic domain wall driving unit, a drain of the first transistor is connected to a bit line, and a gate of the first transistor is connected to a first word line; The source of the second transistor is connected to the spin wave signal generating terminal, the drain of the second transistor is connected to the bit line, and the gate of the second transistor is connected to the second word line; A source of the third transistor is connected to the second magnetic domain wall driving unit, a drain of the third transistor is connected to the bit line, and a gate of the third transistor is connected to a third word line; The spin wave signal detection end is connected to the source line.

5. The in-memory computing array structure based on spin wave units according to claim 1, characterized in that: The first magnetic domain wall driving unit and the second magnetic domain wall driving unit drive the magnetic domain wall to move by applying current, spin wave or external magnetic field to the first magnetic domain wall driving unit and the second magnetic domain wall driving unit.

6. The in-memory computing array structure based on spin wave units according to claim 1, characterized in that: The material composition of the spin wave body that stabilizes the magnetic domain wall includes iron, cobalt, nickel and / or iron-cobalt-nickel alloy.

7. A control method for an in-memory computing array structure based on a spin wave unit, characterized in that: Applied to the in-memory computing array structure based on spin wave units according to any one of claims 1 to 6, the method comprises: When the first magnetic domain wall driving unit and the second magnetic domain wall driving unit drive the magnetic domain wall to move, the spin wave unit determines a target position at which the magnetic domain wall stops moving between the first magnetic domain wall driving unit and the second magnetic domain wall driving unit based on a preset weight control signal sent by the transistor unit to the spin wave signal generating end, the spin wave signal detecting end, the first magnetic domain wall driving unit, and the second magnetic domain wall driving unit, and generates a corresponding preset weight.

8. The control method of the in-memory computing array structure based on spin wave units according to claim 7, characterized in that: The method further comprises: When performing calculations on the traveling salesman problem, the plurality of in-memory computing units use the corresponding preset weights as weights of the Hopfield neural network, and determine the target optimal solution based on the preset weights.

9. The control method of the in-memory computing array structure based on spin wave units according to claim 7, characterized in that: Before the spin wave unit determines, based on the preset weight control signal sent by the transistor unit to the spin wave signal generating end, the spin wave signal detecting end, the first magnetic domain wall driving unit, and the second magnetic domain wall driving unit, a target position at which the magnetic domain wall stops moving between the first magnetic domain wall driving unit and the second magnetic domain wall driving unit, and generates the corresponding preset weight, the method further includes: The spin wave unit determines an input intensity value of the spin wave signal by inputting a spin wave signal of a preset frequency and a preset intensity to the spin wave signal generating end when the first magnetic domain wall driving unit and the second magnetic domain wall driving unit drive the magnetic domain wall to move; The spin wave signal detection end determines an output intensity value corresponding to the spin wave signal; The spinning wave unit determines a current weight based on the input intensity value and the output intensity value; Wherein, the current weight is the intensity ratio of the spin wave signal at a preset frequency; The spinning wave unit determines the preset weight control signal based on the current weight.

10. The control method of the in-memory computing array structure based on spin wave units according to claim 9, characterized in that: After the spin wave unit determines a target position at which the magnetic domain wall stops moving between the first magnetic domain wall driving unit and the second magnetic domain wall driving unit based on the preset weight control signal sent by the transistor unit to the spin wave signal generating end, the spin wave signal detecting end, the first magnetic domain wall driving unit, and the second magnetic domain wall driving unit, and generates the corresponding preset weight, the method further includes: Controlling the plurality of in-memory computing units to store the corresponding current weights; The target weight is determined based on the multiple current weights to complete the regulation of the weight size.

Images