Cellular neural network hardware based on spin hall effect and spin torque effect

Abstract

The application discloses cell neural network hardware based on spin Hall effect and spin torque effect, comprising a certain number of CNN cells which are uniformly distributed in horizontal and vertical directions, the CNN cell comprising a fixed layer, a tunnel layer, a free layer, a spin orbit coupling layer and a channel layer, the first four layers constituting a magnetic tunnel junction, the fixed layer, the tunnel layer, the free layer, the spin orbit coupling layer and the channel layer being stacked in a top-down order, the tunnel layer, the spin orbit coupling layer and the channel layer being parallel to the horizontal direction, the fixed layer and the free layer being at a certain angle with the horizontal direction, the material of the fixed layer and the free layer being nanomagnet, the material of the spin orbit coupling layer being a material with spin Hall effect, the material of the channel layer being metal, the material of the tunnel layer being metal oxide, and the magnetic tunnel junction of each CNN cell being connected with a voltage source and a current source. The application has the advantages of simple structure, small layout area, reduced power consumption and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of CNN hardware implementation, and particularly relates to a cellular neural network hardware based on spin Hall effect and spin torque effect. BACKGROUND

[0002] Cellular neural network (CNN) is a promising non-Boolean computing system due to its powerful parallel computing ability and excellent image processing performance. It is superior to the traditional von Neumann computing structure in the fields of parallel computing and image processing. CNN generally uses CMOS device technology to construct operational amplifiers and operational transconductance amplifiers to realize neurons and synapses, which has no advantage in power consumption and area compared with digital logic circuits.

[0003] There are three kinds of mainstream CNN hardware implementations at present, namely analog very large scale integration (VLSI) implementation, digital VLSI implementation and new neuromorphic circuit implementation. Among them, the analog VLSI implementation mainly uses operational amplifiers, transconductance operational amplifiers and resistors and capacitors to realize CNN. The digital VLSI implementation mainly uses digital signal processors or field programmable logic arrays to construct CNN. The new neuromorphic circuit implementation mainly uses new electronic devices such as memristors and spintronic devices to construct CNN.

[0004] The CNN implemented by analog VLSI needs a large number of operational amplifiers and capacitors, resulting in large power consumption and layout area in large-scale integration implementation. In recent years, with the development of FPGA, digital VLSI has become the main method of CNN hardware implementation, which has the advantages of stable device performance, convenient data storage and short development cycle. Although it has many advantages, it still uses the von Neumann computing structure, which is characterized by separation of storage and logic operation, which is not suitable for the structure characteristics of CNN that storage and operation are on the neurons, resulting in problems such as large interconnection delay, low operation efficiency and high power consumption. In addition, in general spintronic circuits, spin current is transmitted in the channel, and spin current is sharply dissipated with the increase of channel length, resulting in a spin channel length of several hundred nanometers. In order to effectively flip the magnetic moment, the spin current is enhanced, resulting in increased power consumption.

[0005] The CNN implemented by new neuromorphic circuit is based on new electronic devices with non-volatility to construct neural network, which can realize neurons and synapses in a single unit at the same time, and is more consistent with the structure characteristics of CNN. However, these new electronic devices have high requirements on manufacturing process, and it is difficult to implement under the current process conditions, and most of them are still at the theoretical research level. SUMMARY

[0006] The present application provides a cell neural network hardware based on spin Hall effect and spin torque effect to overcome the above technical problems.

[0007] The cell neural network hardware based on spin Hall effect and spin torque effect comprises a certain number of CNN cells uniformly distributed in horizontal and vertical directions, the CNN cell comprising a fixed layer, a tunnel layer, a free layer, a spin orbit coupling layer and a channel layer, the fixed layer, the tunnel layer, the free layer and the spin orbit coupling layer constituting a magnetic tunnel junction, the fixed layer, the tunnel layer, the free layer, the spin orbit coupling layer and the channel layer being stacked in a top-down order, the tunnel layer, the spin orbit coupling layer and the channel layer being parallel to the horizontal direction, the fixed layer and the free layer being at a certain angle with the horizontal direction, the angle being in a range of 0 degree to 90 degree, the fixed layer and the free layer being made of nanomagnet, the spin orbit coupling layer being made of a material having spin Hall effect, the channel layer being made of metal, the tunnel layer being made of metal oxide, the channel layer of each CNN cell being connected to four adjacent CNN cells, the CNN cell at the center of the four adjacent CNN cells being a central neuron, and the magnetic tunnel junction of each CNN cell being connected to a voltage source and a current source.

[0008] Preferably, the free layer in the CNN cell generates a horizontal channel current and a vertical channel current in the channel layer after the voltage source is turned on, the horizontal channel current and the vertical channel current being obtained according to formula (1),

[0009]

[0010] wherein θ SHE and λ sf are the spin Hall angle and the spin diffusion length of the spin orbit coupling layer material, σ is a unit vector, pointing to the direction of the easy magnetization axis of the free layer, w FM and t SOC are the free layer magnet width and the spin orbit coupling layer thickness respectively, I S is the spin current flowing into the spin orbit coupling layer, I Cx and I Cy represent the horizontal channel current and the vertical channel current respectively, and θ represents the angle between the free layer and the horizontal direction.

[0011] Preferably, the magnetization direction of the magnet layer of the central neuron is represented according to formula (2),

[0012]

[0013] wherein γ is the electron gyromagnetic ratio, μ0 is the vacuum permeability, α is the Gilbert damping coefficient of the magnetic material, H ijrepresents the total effective field experienced by the free layer magnet in cell C(i,j), m ij is the magnetic moment of the free layer magnet in cell C(i,j), used as the state variable of cell C(i,j), S SOC is the spin-orbit torque coefficient, M ij is the magnetization direction of the fixed layer magnetic moment, S MTJ is the spin transfer torque coefficient in the magnetic tunnel junction, I ij,kl is the current flowing from cell C(k,l) into cell C(i,j), A ij is the cross-sectional area of the spin-orbit coupling layer in cell C(i,j).

[0014] Preferably, the spin-orbit torque coefficient is obtained according to formula (3),

[0015]

[0016] wherein, is the reduced Planck constant, q is the elementary charge, M S is the free layer saturation magnetization, t FM is the free layer magnet thickness, theta SHE is the spin Hall angle of the spin-orbit coupling layer material, gamma is the electron gyromagnetic ratio.

[0017] Preferably, the spin transfer torque coefficient is obtained according to formula (4),

[0018]

[0019] wherein, is the spin torque amplitude, gamma is the electron gyromagnetic ratio, is the reduced Planck constant, q is the elementary charge, M S is the free layer saturation magnetization, t FM is the free layer magnet thickness, A ij is the cross-sectional area of the spin-orbit coupling layer in cell C(i,j), I ij is the current flowing into the MTJ.

[0020] The present application provides a spin Hall effect and spin torque effect-based cell neural network hardware, and proposes a spin Hall effect and spin torque effect-based cell neural network hardware implementation, which uses the magnetic moment of the nanomagnet of the fixed layer and the free layer as a state variable to store information, and has the advantages of non-volatility, simple structure, small layout area, etc. Meanwhile, the present application transmits charge current in the channel, and the loss in the channel is negligible. Finally, the spin Hall effect is used to flip the magnetic moment, which is more efficient than using the spin torque effect to flip the magnetic moment, so that the power consumption can be further reduced. BRIEF DESCRIPTION OF DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the accompanying drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, the accompanying drawings in the following description are some embodiments of the present application, and all other embodiments obtained by a person of ordinary skill in the art without creative work based on these drawings are within the protection scope of the present application.

[0022] Figure 1 is a top view of the overall structure of the CNN of the present application;

[0023] Figure 2 is a front view of a single CNN cell of the present application;

[0024] Figure 3 is a three-dimensional structure diagram of the CNN based on SHE and STT effects of the present application;

[0025] Figure 4 is a traditional CNN structure diagram of the present application;

[0026] Figure 5 is an MTJ circuit of the present application;

[0027] Explanation of the reference signs:

[0028] 1, fixed layer; 2, tunnel layer; 3, free layer; 4, spin orbit coupling layer; 5, channel layer. DETAILED DESCRIPTION

[0029] In order to make the purpose, technical solutions and advantages of the embodiments of the present application more clear, the technical solutions in the embodiments of the present application will be clearly and completely described below in combination with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are some embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work are within the protection scope of the present application.

[0030] Figure 1 is a top view of the overall structure of the CNN of the present application, as shown in Figure 1 The method of the present embodiment can include:

[0031] The cell neural network hardware based on spin Hall effect and spin transfer torque effect comprises a certain number of CNN cells which are uniformly distributed in horizontal and vertical directions, the CNN cell comprises a fixed layer, a tunnel layer, a free layer, a spin orbit coupling layer and a channel layer, the fixed layer, the tunnel layer, the free layer and the spin orbit coupling layer form a magnetic tunnel junction, i.e. MTJ, the fixed layer, the tunnel layer, the free layer, the spin orbit coupling layer and the channel layer are stacked in the order from top to bottom, the tunnel layer, the spin orbit coupling layer and the channel layer are parallel to the horizontal direction, the fixed layer and the free layer form a certain angle with the horizontal direction, the angle ranges from 0 degree to 90 degree, the material of the fixed layer and the free layer is nanomagnet, the material of the spin orbit coupling layer is a material with spin Hall effect, the material of the channel layer is metal, and the material of the tunnel layer is metal oxide, the channel layer of each CNN cell is connected with four adjacent CNN cells, the CNN cell at the center of the four adjacent CNN cells is the central neuron, and the magnetic tunnel junction of each CNN cell is connected with a voltage source and a current source.

[0032] Based on the above scheme, the magnetic moment of the nanomagnet of the fixed layer and the free layer is used as a state variable to store information, which is non-volatile, so that the operation and storage can be integrated on one neuron, and the structure is simple and the layout area is small. At the same time, the charge current is transmitted in the channel, and the loss in the channel is negligible. Finally, the spin Hall effect is used to flip the magnetic moment, which is more efficient than using the spin transfer torque effect to flip the magnetic moment, so that the power consumption can be further reduced.

[0033] In the embodiment, the meanings of SHE and STT are spin Hall effect and spin transfer torque effect respectively, the spin Hall effect (Spin Hall Effect, SHE) refers to that in heavy metal materials, the current will cause the spin polarization in the vertical direction. The spin transfer torque (Spin Transfer Torque, STT) effect refers to that the spin-polarized current will act on the local magnetic moment during the conduction process, so that the magnetization direction changes.

[0034] The cell neural network based on spin Hall effect and spin transfer torque effect comprises a certain number of CNN cells which are uniformly distributed in horizontal and vertical directions.

[0035] The CNN cell comprises a fixed layer 1, a tunnel layer 2, a free layer 3, a spin orbit coupling layer 4 and a channel layer 5. The front view of a single CNN cell based on SHE and STT effect is shown in Figure 2 .

[0036] The fixed layer, the tunnel layer, the free layer and the spin orbit coupling layer form a magnetic tunnel junction.

[0037] The material of the fixed layer and the free layer is nanomagnet, which is generally composed of ferromagnetic material with in-plane magnetic anisotropy, such as nickel, cobalt, permalloy and the like.

[0038] The material of the spin-orbit coupling layer is a material with spin Hall effect, such as Bi, Mo, Ru, Pd, Pt, Ta, W and the like.

[0039] The material of the channel layer is metal, which can be generally composed of silver, copper, aluminum and the like.

[0040] The material of the tunnel layer is metal oxide, such as MgO, Al2O3 and the like.

[0041] The fixed layer, the tunnel layer, the free layer, the spin-orbit coupling layer and the channel layer are stacked in the order from top to bottom, each CNN cell is connected with four adjacent CNN cells, the CNN cell at the center of the four adjacent CNN cells is a central neuron, the connection is achieved through the channel layer between the CNN cells, the tunnel layer, the spin-orbit coupling layer and the channel layer are parallel to the horizontal direction, the fixed layer and the free layer are at a certain angle with the horizontal direction, the angle ranges from 0 degree to 90 degree, and the magnetic tunnel junction of each CNN cell is connected with a voltage source and a current source.

[0042] Each cell is connected with the four most adjacent cells. It should be noted that the MTJ free layer and the fixed layer are not arranged along the x-axis or the y-axis direction, but in a direction at a certain angle with the x-axis. Figure 1 The angle shown in the figure is 45°, which can be adjusted. In the figure, the MTJ (Magnetic Tunnel Junction) is a magnetic tunnel junction stacked on the spin-orbit coupling layer. In order to more intuitively show the designed CNN structure, Figure 3 A three-dimensional structure diagram of CNN based on SHE and STT effect is given. In the CNN, the general neuron and cell can be used. Figure 3 In the figure, only the voltage source and the current source are given on the central cell, and the same voltage source and current source are also provided on other cells, which are omitted in the figure for simplicity. The basic working principle is as follows: first, an appropriate current I is applied to the MTJ, which changes the magnetic moment direction of the free layer magnet, thereby realizing the initialization of the CNN. Secondly, a voltage VDD is applied to the free layer, which generates a spin current. When the spin current enters the spin-orbit coupling layer, due to the influence of the inverse spin Hall effect, charge current will be generated in the x-axis and y-axis directions. The charge current is collected to the spin-orbit coupling layer of the central cell through the channel layer, and due to the influence of the spin Hall effect, the magnetization direction of the free layer of the central neuron is changed, thereby realizing the change of the CNN state variable.

[0043] The basic cell neural network technology principle is introduced as follows:

[0044] The traditional cellular neural network (CNN) is a large-scale nonlinear analog circuit with real-time signal processing function, has high-speed parallel processing function and continuous dynamics, and has wide application prospect. At present, the research results of the CNN have been widely applied to many fields such as biomedical science, image processing, automatic control, pattern recognition, signal processing, secret communication and the like.

[0045] The CNN adopts the structure of cellular automata, and contains a plurality of same calculation units inside, each calculation unit is called a cell, and each cell is connected only with the closest cells, so that the problem that each neuron in the Hopfield neural network is connected with other neurons is overcome, and the CNN is more easy to realize on VLSI. Figure 4 A 5x5 CNN structure schematic diagram is given. The block represents a cell in the CNN, and each cell can not only be connected with the cells directly adjacent to it, but also be connected with cells in a farther range. Therefore, the neighborhood of the CNN can be defined by the action range of the cells and the surrounding cells.

[0046] If r represents the action range of the cells in the CNN, for an M-row N-column CNN array, the neighborhood of the cell C(i,j) in the i-th row and the j-th column (i.e. all the cells interacting with the cell C(i,j)) can be defined as:

[0047] N r (i,j)={C(i,j)|max(|i-k|,|j-l|)≤r,1≤k≤M;1≤l≤N;r≥1} (1)

[0048] Figure 4 In the formula, r=1 represents a single-neighborhood CNN, i.e. the center cell only interacts with the surrounding 8 cells. r=2 represents a double-neighborhood CNN, i.e. the center cell not only interacts with the surrounding 8 cells, but also interacts with the 16 cells in the outermost periphery in the figure. By analogy, when r=n, there are (2n+1) 2 -1 cells interacting with the center cell (without considering the boundary condition).

[0049] The dynamics of each cell in the traditional CNN can be expressed as:

[0050]

[0051] In the formula, x ij is the state voltage of the cell C(i,j), R and C are the resistance and capacitance in the cell respectively, u kl and y kl are the input and output of the adjacent cells respectively, and I ijis the input current of each cell. f(x) is a nonlinear function used to describe the relationship between the state voltage x of the cell and the output voltage y ij ij A and B are generally called A template and B template, whose values are used to represent the connection weight between cell C(i,j) and cell C(k,l). ij,kl ij,kl A and B are generally called A template and B template, whose values are used to represent the connection weight between cell C(i,j) and cell C(k,l).

[0052] Due to its high symmetry and consistency, CNN is easy to use analog VLSI or digital VLSI for hardware implementation, but its main disadvantage is low operation efficiency, high power consumption and large layout area.

[0053] The present application proposes a CNN structure based on spin Hall effect and spin torque effect. Unlike the traditional CNN constructed by CMOS devices, the proposed CNN does not use voltage as the state variable, but uses the magnetic moment of a magnet as the state variable, having the advantages of non-volatility and ultra-low power consumption. In addition, compared with the CNN constructed by using operational amplifier and transconductance amplifier, the proposed CNN uses a heterojunction structure (as shown in Figure 2 ) to realize the neuron, which is simpler in structure than the CNN constructed by CMOS devices.

[0054] The CNN structure based on SHE and STT uses MTJ structure to realize the initialization of CNN state variables, uses channel layer to realize the connection between cells, and adjusts the connection weight by selecting the material and size of spin orbit coupling layer and channel layer. The following will be described in detail.

[0055] (1) Magnetic tunnel junction structure

[0056] The MTJ circuit is as shown in Figure 5 , wherein the fixed layer magnetic moment of the MTJ remains unchanged, and the free layer of the MTJ serves as the core operation unit in the CNN, and its magnetization is used as the state variable of the CNN. By flowing different direction currents I on the MTJ, the direction of the magnetic moment of the free layer of the MTJ can be changed, thereby realizing the initialization of the CNN.

[0057] (2) Working principle of the cell neural network proposed in the embodiment

[0058] In the present application, the easy axis of the MTJ free layer (that is, the direction of the long axis of the free layer) is at an angle θ (0° < θ < 90°) with the x axis, as shown in Figure 3 . When a positive voltage V DD is applied to the free layer, a spin current in the -z axis direction (from top to bottom) is generated by the magnet, and a transverse charge current is generated in the spin Hall layer due to the influence of the inverse spin Hall effect.

[0059] ​​

[0060] where θ SHE and λ sf are the spin Hall angle and the spin diffusion length of the spin orbit coupling layer material. σ is a unit vector pointing to the easy axis direction of the free layer, w FM and t SOC are the free layer magnet width and the spin orbit coupling layer thickness, respectively. I S is the spin current flowing into the spin orbit coupling layer. Since the spin current component in the z-axis direction can be neglected, the free layer in the CNN cell generates a horizontal channel current and a vertical channel current in the channel layer after the voltage source is turned on, which are obtained according to formula (4),

[0061]

[0062] where θ SHE and λ sf are the spin Hall angle and the spin diffusion length of the spin orbit coupling layer material, σ is a unit vector pointing to the easy axis direction of the free layer, w FM and t SOC are the free layer magnet width and the spin orbit coupling layer thickness, respectively, I S is the spin current flowing into the spin orbit coupling layer, I Cx and I Cy represent the horizontal channel current and the vertical channel current, respectively, and θ represents the angle between the free layer and the horizontal direction.

[0063] It can be seen that I Cx and I Cy are related to the angle θ between the easy axis of the free layer and the x-axis, and when the angle θ is adjusted, the current size in the horizontal channel and the vertical channel of the CNN will change accordingly. When the charge currents of the surrounding neurons converge to the spin Hall layer of the central neuron (as shown in Figure 3 ), the magnetization direction of the magnet layer of the central neuron will change due to the influence of the spin Hall effect, and the magnetization dynamics behavior can be described by the LLG equation, and the magnetization direction of the magnet layer of the central neuron is represented according to formula (5),

[0064]

[0065] where γ is the electron gyromagnetic ratio, μ0 is the vacuum permeability, α is the Gilbert damping coefficient of the magnetic material, H ij represents the total effective field received by the free layer magnet in the cell C(i,j), m ij is the magnetic moment of the free layer magnet in the cell C(i,j), which is used as the state variable of the cell C(i,j), SSOC is the spin-orbit torque coefficient, M ij is the magnetization direction of the fixed layer magnetic moment, S MTJ is the spin-transfer torque coefficient in the magnetic tunnel junction, I ij,kl is the current flowing from cell C(k, l) to cell C(i, j), A ij is the cross-sectional area of the spin-orbit coupling layer in cell C(i, j).

[0066] The first term on the right side of the above LLG equation is the precession term of the total effective field experienced by the magnetic body, the second term is the Gilbert damping term, which represents the precession of the magnetic moment towards the effective field direction during the precession, the third term is the magnetic moment precession term caused by the spin-orbit coupling moment, and the fourth term is the magnetic moment precession term caused by the spin-transfer torque.

[0067] Equation (5) is an implicit function. By making it an explicit function, we get:

[0068]

[0069] In the first and second terms on the right side of equation (6), γ is the electron gyromagnetic ratio, μ0 is the vacuum permeability, α is the Gilbert damping coefficient of the magnetic material, H ij represents the total effective field experienced by the free layer magnetic body in cell C(i, j), m ij is the magnetic moment of the free layer magnetic body in cell C(i, j), which is used as the state variable of cell C(i, j). Therefore, the first and second terms on the right side of equation (6) can be considered as the first term in the right side of equation (2).

[0070] The spin-orbit torque coefficient is obtained according to equation (7),

[0071]

[0072] wherein, is the reduced Planck constant, q is the elementary charge, M S is the saturation magnetization of the free layer, t FM is the thickness of the free layer magnetic body, θ SHE is the spin Hall angle of the spin-orbit coupling layer material, and γ is the electron gyromagnetic ratio.

[0073] I ij,kl is the current flowing from cell C(k, l) to cell C(i, j), which is influenced by the voltage V DD applied to the spin-orbit coupling layer, the angle θ between the easy axis of the free layer and the x-axis, and the material and size of the spin-orbit coupling layer. In addition, we take the magnetic moment of the free layer magnetic body in the adjacent cell as the output of the adjacent cell, which is equivalent to y kl in equation (2), and take the voltage V DDAs the input to the neuron, it is equivalent to u in equation (2). kl Therefore, the third and fourth terms on the right side of equation (6) can be considered as the second and third terms on the right side of equation (2).

[0074] In the fifth and sixth terms on the right side of equation (6), M ij S is the magnetization direction of the fixed layer magnetic moment. MTJ Here is the spin-transfer torque coefficient in MTJ, which is obtained according to formula (8).

[0075]

[0076] in, It is the spin moment amplitude, and γ is the electron gyromagnetic ratio. It is the reduced Planck constant, q is the elementary charge, and M is the reduced Planck constant. S It is the saturation magnetization of the free layer, t FM It is the thickness of the free layer magnet, A ij Let I be the cross-sectional area of ​​the spin-orbit coupling layer in cell C(i,j). ij Let I be the current flowing into MTJ. It can be seen that this current is related to the current I flowing into MTJ. ij And it is related to the angle between the magnetic moments of the fixed layer and the free layer. Therefore, the fifth and sixth terms on the right side of equation (6) can be considered as the fourth term on the right side of equation (2).

[0077] Furthermore, the change in the magnetization direction of a magnet is a precession process, a non-linear motion trajectory, equivalent to a non-linear function in a traditional CNN.

[0078] In summary, the differential equation described in equation (6) can realize the function of CNN. Generally speaking, we use the current flowing into the MTJ structure as the input bias current in the CNN. This current must exceed a certain threshold to change the magnetic moment of the MTJ free layer. Simultaneously, we use the voltage VDD on the free layer of the neuron as the neuron's input, and the magnetic moment of the free layer as the neuron's output.

[0079] Generally, the A template is associated with the output variable (i.e., the state variable), while the B template is associated with the input variable. As can be seen from equations (6), (7), and (4), changing the material and size of the spin-orbit coupling layer and the channel layer will change the current density flowing into the central cell, which is equivalent to changing the connection weights between cells, that is, changing the parameter values ​​in the A template.

[0080] The material, size, and magnetic moment direction of the free-layer magnet will affect the efficiency and polarization direction of the input voltage conversion into spin current, which is equivalent to changing the connection weight between the input and output, that is, changing the parameter values ​​in the B template.

[0081] It is necessary to point out that in the CNN structure, the center cell only interacts with the four most adjacent cells, and does not interact with the cells on the diagonal, only a quasi-single-field CNN structure can be realized, which makes it have certain limitations in selecting the parameters of the A template and the B template relative to the traditional CNN.

[0082] (3) Performance estimation of CNN based on SHE and STT effects

[0083] By Figure 1 It can be seen that the single cell in the CNN proposed in the present application is only a heterojunction composed of a fixed layer, a tunnel layer, a free layer and a spin orbit coupling layer. Under the current process conditions, the length of the heterojunction can be within 100 nm, the width can be within 50 nm, and the thickness can be within 30 nm. The typical value of the channel length connecting the cells can be set to 200 nm. Therefore, the layout area of a 3*3 scale CNN is about 0.5 um2. In addition, according to our previous research, the CNN based on SHE effect and STT effect has a power consumption of about picojoule level, a delay of nanosecond level, and a working voltage of millivolt level. Compared with the CNN realized by the traditional CMOS process, the CNN has obvious improvements in power consumption, layout area and working voltage.

[0084] Overall beneficial effects:

[0085] At present, there are generally three kinds of hardware implementations of CNN, namely analog VLSI implementation, digital VLSI implementation and new neuromorphic circuit implementation. The present application proposes a hardware implementation method different from the three kinds, that is, a CNN hardware implementation method based on SHE and STT effects. Since the present application uses the magnetic moment of a nanomagnet as a state variable to store information, which is non-volatile, the operation and storage can be integrated on one neuron, which has the advantages of simple structure and small layout area. In addition, in general spintronic circuits, spin current is transmitted in the channel, and the spin current is sharply dissipated with the increase of the channel length, resulting in that the spin channel length is generally within a few hundred nanometers. In order to effectively flip the magnetic moment, the spin current is enhanced, resulting in an increase in power consumption. The CNN structure proposed in the present application transmits charge current in the channel, which can be ignored. Finally, the spin Hall effect is used to flip the magnetic moment, which is more efficient than using the spin torque effect to flip the magnetic moment, so that the power consumption can be further reduced. Therefore, the CNN based on SHE and STT effects proposed in the present application will greatly improve the performance of the layout area and power consumption relative to the traditional CNN, and is expected to be a competitive direction for CNN hardware implementation in the post-CMOS era.

[0086] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions recorded in the above embodiments can be modified, or some or all of the technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A cellular neural network hardware based on spin Hall effect and spin torque effect, characterized in that, The CNN cells are uniformly distributed in horizontal and vertical directions, and each CNN cell comprises a fixed layer, a tunnel layer, a free layer, a spin-orbit coupling layer, and a channel layer, the fixed layer, the tunnel layer, the free layer, and the spin-orbit coupling layer constitute a magnetic tunnel junction, the fixed layer, the tunnel layer, the free layer, the spin-orbit coupling layer, and the channel layer are stacked in a top-down order, the tunnel layer, the spin-orbit coupling layer, and the channel layer are parallel to the horizontal direction, the fixed layer and the free layer are at an angle with the horizontal direction, the angle ranges from 0 degree to 90 degree, the fixed layer and the free layer are made of nanomagnets, the spin-orbit coupling layer is made of a material having a spin Hall effect, the channel layer is made of a metal, the tunnel layer is made of a metal oxide, the channel layer of each CNN cell is connected to four adjacent CNN cells, and the CNN cell at the center of the four adjacent CNN cells is a central neuron, and the magnetic tunnel junction of each CNN cell is connected to a voltage source and a current source. The magnetization direction of the magnet layer of the central neuron is represented by formula (2), (2) in, Electron gyromagnetic ratio, The permeability of free space, The Gilbert damping coefficient of the magnetic material is... Represents cells The total effective field experienced by the magnet in the middle free layer It is a cell The magnetic moment of the free-layer magnet is used as a cell. State variables, The spin orbital moment coefficient, The magnetization direction is fixed by the magnetic moment of the layer. The spin-transfer moment coefficient in the magnetic tunnel junction. To from cells Inflow into cells The current in For cells Cross-sectional area of ​​the spin-orbit coupling layer.

2. The spin Hall effect and spin torque effect based cellular neural network hardware of claim 1, wherein, The free layer in the CNN cell generates a horizontal channel current and a vertical channel current in the channel layer after the voltage source is turned on, and the horizontal channel current and the vertical channel current are obtained according to formula (1), (1) wherein, and is a spin Hall angle and a spin diffusion length of a spin orbit coupling layer material, is a unit vector pointing to a direction of an easy axis of the free layer, and are a free layer magnet width and a spin orbit coupling layer thickness, respectively, is a spin current flowing into the spin orbit coupling layer, and represent a current of a horizontal channel and a current of a vertical channel, respectively, represents an angle of the free layer with a horizontal direction.

3. The spin Hall effect and spin torque effect based cellular neural network hardware of claim 2, wherein, The spin-orbit moment coefficient is obtained according to formula (3), (3) wherein, is the reduced Planck constant, q is the elementary charge, M S is the free layer saturation magnetization, t FM is the free layer magnet thickness, is the spin Hall angle of the spin orbit coupling layer material, is the electron gyromagnetic ratio.

4. The spin-Hall effect and spin torque effect based cellular neural network hardware of claim 2, wherein, The spin transfer torque coefficient is obtained according to formula (4), (4) in, η ( φ ) is the spin moment amplitude. Electron gyromagnetic ratio, It is the reduced Planck constant. q For elementary charge, M S It is the saturation magnetization of the free layer. t FM It is the thickness of the free layer magnet. For cells Cross-sectional area of ​​the spin-orbit coupling layer This represents the current flowing into the MTJ.

Images