Nonvolatile, linear and bidirectional synaptic weight elements based on magnetic domain walls

Abstract

A magnetic double tunnel junction (MDTJ) (preferably with a large aspect ratio, where the length L of the MDTJ>>the width w of the MDTJ) has a magnetic domain wall or DW in the free layer of the MDTJ, wherein the controlled movement of the DW across the free layer is affected in response to the polarity, amplitude and duration of a voltage pulse across the MDTJ. The motion and relative position of the DW causes the conductivity of the MDTJ (which is measured across the MDTJ) to vary in a symmetrical and linear manner. By reversing the polarity of the bias voltage, the creation and / or direction of the DW motion can be reversed, allowing a bidirectional response to an input pulse.

Description

Background Art Technical Field

[0001] The present invention relates generally to the field of synaptic weight elements (eg, in a crossbar array). More specifically, the present invention relates to non-volatile, linear and bidirectional synaptic weight elements based on magnetic domain walls.

[0002] Related technical discussions

[0003] Deep learning networks or neural networks or other machine learning algorithms rely on multiply-accumulate functions (related to matrix multiplication, etc.) This can be most efficiently implemented in crossbar arrays of nominally identical synaptic elements that exhibit preferentially linear but necessarily bidirectional and symmetric conductivity responses to input signals.

[0004] The synaptic elements that have been used in the prior art generally include phase change materials (PCM), resistive random access memory (RRAM), ferroelectric RAM (FeRAM), etc. Another prior art embodiment disclosed in a paper such as Lequeux et al. (“A Magnetic Synapse: Multilevel Spin-Torque Memristor with Perpendicular Anisotropy,” Scientific Reports, 6:31510, 2016) uses a “memristor” as a non-volatile nano-resistor, wherein the resistance of such a nano-resistor can be adjusted by an applied current or voltage and can be set to a large number of levels.

[0005] Embodiments of the present invention are improvements over prior art systems and methods. Summary of the invention

[0006] In one embodiment, the present invention provides a method for adjusting the conductivity of a synaptic weight element, wherein the synaptic weight element includes: (i) an upper electrical contact; (ii) a first magnetic reference layer directly below the upper electrical contact; (iii) a first insulating tunnel barrier directly below the first magnetic reference layer; (iv) a magnetic free layer directly below the first insulating tunnel barrier; (v) a second insulating tunnel barrier directly below the magnetic free layer; (vi) a second magnetic reference layer directly below the second insulating tunnel barrier, and (vii) a lower electrical contact directly below the second magnetic reference layer, the first insulating tunnel barrier and the second insulating tunnel barrier having different resistances. In one aspect, there is a method comprising the steps of: (a) when the conductivity of a synaptic weight element is to be increased, applying a first bias voltage V between an upper electrical contact and a lower electrical contact, such that the first bias voltage V induces a first voltage V1 across a first insulating tunnel barrier and induces a second voltage V2 across a second insulating tunnel barrier, wherein V=V1+V2 and V1≠V2, wherein the first bias voltage V creates and / or moves at least one magnetic domain wall in a magnetic free layer along a first direction, wherein the creation and / or movement of at least one magnetic domain wall along the first direction increases the conductivity of the synaptic weight element; and (b) when the synaptic weight element is to be decreased When the conductivity of the synaptic weight element is reduced, a second bias voltage V' is applied between the upper electrical contact and the lower electrical contact, the polarity of the second bias voltage V' is opposite to the polarity of the first bias voltage V, the second bias voltage V' causes a third voltage V3 across the first insulating tunnel barrier, and causes a fourth voltage V4 across the second insulating tunnel barrier, wherein V'=V3+V4 and V3≠V4, and the second bias voltage V' creates and / or moves at least one magnetic domain wall in the magnetic free layer in a second direction opposite to the first direction, and wherein the creation and / or movement of at least one magnetic domain wall in the second direction reduces the conductivity of the synaptic weight element.

[0007] In another aspect of the present invention, a method for adjusting the conductivity of a synaptic weight element is provided, wherein the synaptic weight element includes: (i) an upper electrical contact; (ii) a first magnetic reference layer directly below the upper electrical contact; (iii) a first insulating tunnel barrier directly below the first magnetic reference layer; (iv) a magnetic free layer directly below the first insulating tunnel barrier; (v) a second insulating tunnel barrier directly below the magnetic free layer; (vi) a second magnetic reference layer directly below the second insulating tunnel barrier, and (vii) a lower electrical contact directly below the second magnetic reference layer, the first insulating tunnel barrier and the second insulating tunnel barrier having different resistances. The method comprises the following steps: (a) when the conductivity of the synaptic weight element is to be reduced, applying a first bias voltage V between the upper electrical contact and the lower electrical contact, so that the voltage V causes a first current I1 to travel from the upper electrical contact to the lower electrical contact, the first bias voltage V creates and / or moves at least one magnetic domain wall in the magnetic free layer in a first direction, wherein the creation and / or movement of the at least one magnetic domain wall in the first direction reduces the conductivity of the synaptic weight element; and (b) when the conductivity of the synaptic weight element is to be increased, applying a second bias voltage V between the upper electrical contact and the lower electrical contact, the polarity of the second bias voltage V' being opposite to the polarity of the first bias voltage V, the second bias voltage V' causing a second current I2 to travel from the lower electrical contact to the upper electrical contact, wherein the second bias voltage V' creates and / or moves at least one magnetic domain wall in the magnetic free layer in a second direction opposite to the first direction, and wherein the increase and / or movement of the at least one magnetic domain wall in the second direction increases the conductivity of the synaptic weight element.

[0008] In one embodiment, the present invention provides a device comprising: (i) a first magnetic layer; (ii) a first tunnel barrier located below the first magnetic layer; (iii) a second magnetic layer located below the tunnel barrier; (iv) a second tunnel barrier located below the second magnetic free layer, the first tunnel barrier and the second tunnel barrier having different resistances; and (v) a third magnetic layer located below the second tunnel barrier, wherein the first magnetic layer, the first tunnel barrier, the second magnetic layer, the second tunnel barrier and the third magnetic layer are close to each other to form a layer stack, so that when an electric current passes through the stack, the electric current creates and / or moves at least one magnetic domain wall in the second magnetic layer.

[0009] In another embodiment, the present invention provides a device comprising: (i) a first magnetic layer; (ii) a third tunnel barrier located below the first magnetic layer; (iii) a second magnetic layer located below the tunnel barrier; (iv) a second tunnel barrier located below the second magnetic free layer, the first tunnel barrier and the second tunnel barrier having different resistances; and (v) a third magnetic layer located below the second tunnel barrier, wherein the first magnetic layer, the first tunnel barrier, the second magnetic layer, the second tunnel barrier and the third magnetic layer are close to each other to form a layer stack, so that a tunnel current passes through the first magnetic layer, the first tunnel barrier, the second magnetic layer, the second tunnel barrier and the third magnetic layer in sequence, and the tunnel current creates and / or moves a magnetic domain wall or multiple domain walls in the second magnetic layer. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] According to one or more different examples, the present disclosure is described in detail with reference to the following drawings. The drawings are provided for illustrative purposes only and depict only examples of the present disclosure. These drawings are provided to facilitate the reader's understanding of the present disclosure and should not be considered as limitations on the breadth, scope or applicability of the present disclosure. It should be noted that for clarity and ease of description, these drawings are not necessarily drawn to scale.

[0011] Figure 1A-1B A schematic cross-sectional view and a top view of a magnetic free layer of an MDTJ (magnetic double tunnel junction) of the present invention are depicted respectively.

[0012] Figure 2A-2B An embodiment of an MDTJ is depicted with parallel top and bottom reference layers during write and read operations, respectively.

[0013] Figure 3A-3D Bidirectional / symmetric operation of one embodiment of the MDTJ of the present invention is shown, with parallel top and bottom reference layers, where the spin transfer torque from the thinner tunnel junction determines the direction of the magnetic DW (domain wall) motion.

[0014] Figure 4A-4B A schematic cross-sectional view and a top view of the magnetic free layer of the MDTJ of the present invention are depicted, respectively.

[0015] Figure 5A-Figure 5B An embodiment of an MDTJ is depicted with anti-parallel top and bottom reference layers during write and read operations, respectively.

[0016] Figure 6A-6D The bidirectional / symmetric operation of one embodiment of the MDTJ of the present invention is shown, with antiparallel top and bottom reference layers, where the spin transfer torques from the two tunnel barriers act in a synchronized manner and determine the direction of the magnetic DW motion.

[0017] Figure 7 Depicted is a general example of a crossbar array that can be used, for example, in neuromorphic computing.

[0018] Figure 8A-8C Depicted is a non-limiting example of a crossbar array having an array of annular synaptic MDTJ devices in accordance with the teachings of the present invention, where the devices may be closed annular structures (eg, Figure 5B as shown) or an open annular structure (as Figure 8C shown).

[0019] Figure 9A-9C Other non-limiting examples of crossbar arrays having arrays of square synaptic MDTJ devices according to the teachings of the present invention are depicted, where the square devices may be closed square structures (e.g., Fig. 9B as shown) or an open square structure (as Figure 8C shown).

[0020] Figures 10A-10D Top views of a closed ring MDTJ device, an open ring MDTJ device, a closed square MDTJ device, and an open square MDTJ device are depicted, respectively.

[0021] Figure 11A-11B A non-limiting example of a method of the present invention for regulating the conductivity associated with a synaptic weight element is depicted. DETAILED DESCRIPTION

[0022] Although the present invention is illustrated and described in preferred embodiments and various aspects, the present invention can be produced in many different configurations. Preferred embodiments and various aspects of the present invention are depicted in the accompanying drawings and will be described in detail herein, it should be understood that the disclosure should be considered as an example of the principles of the present invention and the related functional descriptions for its construction, rather than limiting the present invention. Those skilled in the art will foresee many other possible changes within the scope of the present invention.

[0023] Note that in this specification, reference to "one embodiment" or "an embodiment" means that the referenced features are included in at least one embodiment of the present invention. In addition, separate references to "one embodiment" in this specification do not necessarily refer to the same embodiment; however, unless so stated and unless obvious to one of ordinary skill in the art, the embodiments are not mutually exclusive. Therefore, the present invention can include any various combinations and / or integrations of the embodiments described herein.

[0024] The present invention provides a magnetic double tunnel junction (MDTJ) (in a preferred embodiment, the MDTJ has a large aspect ratio, where the length L of the MDTJ >> the width w of the MDTJ) so as to have a magnetic domain wall or DW in the free layer of the MDTJ, wherein the creation and / or movement of the DW on the free layer is controlled in response to the polarity, amplitude and duration of a voltage pulse across the MDTJ. The movement and relative position of the DW causes the conductivity of the MDTJ (measured across the MDTJ) to change in a symmetrical and linear manner. By reversing the polarity of the bias voltage, the direction of the DW movement can be reversed, allowing a bidirectional and symmetrical response to an input pulse.

[0025] Figure 1A-1B A schematic cross-sectional view and a top view of an MDTJ of one embodiment are depicted, respectively. The magnetic free layer of the MDTJ is sandwiched by two magnetic reference layers with fixed magnetizations. The magnetic DW is shown in the magnetic free layer at a distance d from the left edge. Figure 1A The cross-section of the MDTJ shown in FIG. 1 includes: (i) an upper electrical contact 102, (ii) a first magnetic reference layer 104 directly below the upper electrical contact 102; (iii) a first insulating tunnel barrier 106 directly below the first magnetic reference layer 104; (iv) a magnetic free layer 108 directly below the first insulating tunnel barrier 106; (v) a second insulating tunnel barrier 110 directly below the magnetic free layer 108; (vi) a second magnetic reference layer 112 directly below the second insulating tunnel barrier 110, and (vii) a lower electrical contact directly below the second magnetic reference layer.

[0026] As a non-limiting example, the magnetic reference layer 104 may be formed of synthetic antiferromagnetic multilayers such as (from bottom to top, where the numbers for the various elements indicate their respective thicknesses in Angstroms, and the numbers before 'x' imply repetition of the corresponding bilayer indicated within {}):

[0027] 6x{6Co / 6Pt}|6Co|9Ru|6Co|5x{6Pt|6Co}|{2Ta, 2Mo or 2W}|12Co 20 Fe 60 B 20

[0028] Similarly, 112 may be formed of synthetic antiferromagnetic multilayers such as (from bottom to top, where the numbers for the various elements indicate their respective thicknesses in angstroms, and the numbers before 'x' imply repetition of the corresponding bilayer indicated within {}):

[0029] 12Co 20 Fe 60 B 20|{2.5Ta,2.5Mo or 2.5W}|4x{6Co / 6Pt}|6Co|9Ru|6Co|6x{6Pt|6Co}

[0030] The magnetic free layer 108 may be formed of:

[0031] 9Co 20 Fe 60 B 20 |{2Ta,2Mo or 2W}|8Co 20 Fe 60 B 20 |

[0032] The thinner and thicker insulating tunnel barriers 106 and 110 may be formed from 11 Angstroms of MgO and 14 Angstroms of MgO, respectively.

[0033] It should be noted that the above film stacks are illustrative and non-limiting examples. It is also possible to obtain a reference layer without any synthetic antiferromagnetic exchange.

[0034] As indicated, the two tunnel barriers have different thicknesses t1 and t2, so that for the current flowing through the MDTJ, the voltages v1 and v2 across the two tunnel barriers are different, respectively. Figure 1A-1B ), which ensures that the spin transfer torques of the two tunnel barriers will have different magnitudes, so that the remaining spin transfer torque causes a magnetic DW or multiple DWs to be created and / or moved along the magnetic free layer.

[0035] For the configuration where the two reference layers are antiparallel to each other ( Figure 4A and Figure 4B ), the selection of two tunnel barriers having different thicknesses t1 and t2 as indicated ensures that the spin transfer torque on the two tunnel barriers uniformly causes a magnetic DW or multiple DWs to be created and / or moved along the magnetic free layer.

[0036] As described above, in preferred embodiments, the MDTJ device of the present invention has a large aspect ratio, where the length L of the MDTJ >> the width w of the MDTJ. In one non-limiting example, for illustration purposes only, L may be 100 nm and w may be 10 nm. In another non-limiting example, for illustration purposes only, L may be 66.67 nm and w may be 15 nm. It should be noted that in the illustrative and non-limiting examples, L and w are selected so that the cross-sectional area in both cases, i.e., L×w, is similar, i.e., 10 3 nm 2 .

[0037] Figure 2A-2BThe MDTJ during write and read operations are depicted, respectively. The DW moves laterally in the magnetic free layer of the MDTJ under a nanosecond (ns) long voltage pulse V MDTJ .exist Figure 2A During the write operation depicted in , the spin transfer torque on the thinner tunnel barrier (i.e., in this non-limiting example, t1) will determine the direction of the magnetic DW motion because the spin transfer torque at the low voltage V1 will have a larger magnitude and, therefore, be more effective than the spin transfer torque at the higher voltage V2.

[0038] During a write operation, the large V MDTJ (in a non-limiting example, it may be ~0.5V-1V) is applied to the MDTJ. The DW motion is guided by the voltage across the thinner tunnel barrier (in this non-limiting example, t1) because the spin transfer torque efficiency decreases with increasing voltage across the tunnel barrier.

[0039] exist Figure 2B During the read operation depicted in FIG. 1 , since the resistance of the thicker tunnel barrier (in this non-limiting example, t2) will be higher than the resistance of the thinner tunnel barrier (in this non-limiting example, t1), the resistance difference, i.e., the tunnel magnetoconductance (TMG) [TMG = (G P -G AP ) / G AP ] or tunnel magnetoresistance (TMR) [TMR = (R AP -R P ) / R P ] will be determined by the thicker tunnel barrier. Figure 2B During the read operation depicted in the figure, the smaller V MDTJ (in a non-limiting example, may be ~0.1V) is applied across the MDTJ. Thus, during read, the voltage drop across the two junctions is smaller, and by extension, a higher TMR (or TMG) ​​is observed from both junctions.

[0040] The conductivity across the tunnel junction is given by the following equation,

[0041] G=G P (1-TMG*d / L)

[0042] Among them, G P (G AP ) is the parallel (antiparallel) conductivity of the MDTJ, tunnel magnetic permeance (TMG) = (G P -G AP ) / G P The above equation implies that the conductivity of the DMTJ varies linearly with the magnetic DW position d.

[0043] Note that since during read, the TMG (1 / TMR) from the thicker tunnel barrier (in this non-limiting example, t2) dominates (see Figure 2B ); therefore, in another embodiment, the reference layer proximate the thinner tunnel barrier is formed of an in-plane magnetic material oriented along the length L of the device or perpendicular to the surface defined by L and t. During a write operation, the combination of the torques from the two tunnel barriers will determine the motion of the magnetic domain wall or domain walls in the magnetic free layer.

[0044] Figure 3A-3C The bidirectional / symmetric operation of the MDTJ of the present invention is shown, where the spin transfer torque from the thinner tunnel junction determines the direction of the magnetic DW motion. MDTJ Polarity makes the electron flow I electron When the magnetic DW flows from the bottom electrode to the top electrode, it will pass through the DW in FIG. 3A to FIG. 3B This causes the conductivity to decrease (i.e., the resistance to increase), as shown in Figure 3D 302 in the graph. Reversing the voltage polarity causes DW to move in the opposite direction (ie, to the left). Switching V MDTJ The polarity causes the magnetic DW to move in opposite directions, thereby showing a bidirectional and symmetrical response, as shown in FIG. Figure 3C This results in an increase in conductivity (i.e., a decrease in resistance), as shown by Figure 3D The portion of the graph pointed by arrow 304 is shown.

[0045] Table 1 below summarizes the Figure 3A-3D In the example shown, the relationship between the electron flow polarity, the DW movement direction and the conductivity (i.e., the increase or decrease in conductivity) is shown. In the present invention, the magnetic domain wall ( Figure 1A ) can have two polarities: up-down (↓↑) or down-up (↓↑). Specifically, Figure 1A The magnetic free layer shown in FIG. 1 comprises down-up (↓↑) domain walls. It should be noted that for a given current polarity, domain walls with opposite polarities will move in opposite directions, resulting in the same change in conductivity, i.e., no matter which polarity of domain walls is created and / or moved, the change in conductivity is determined by the polarity of the current.

[0046] Current polarity DW Sports Electrical conductivity (G) bottom up right Reduce Top to bottom Left Increase

[0047] Table 1

[0048] Although Figure 1A Magnetic reference layers 104 and 112 are depicted as being parallel to each other, but other embodiments are contemplated. For example, Figure 4A and Figure 5A-Figure 5BAnother embodiment is depicted in which the magnetic reference layers 502 and 504 are designed to be antiparallel to each other. In this embodiment, one or more magnetic domain walls in the magnetic free layer are created and / or moved by two torques from two tunnel barriers acting in unison. During a read operation, TMG (or TMR) is still dominated by the thicker tunnel barrier.

[0049] Figure 5A-Figure 5B The MDTJ is depicted during write and read operations, respectively, where the magnetic reference layers are antiparallel to each other. One or more DWs move laterally in the free layer of the MDTJ under a nanometer-long voltage pulse (V MDTJ ) is created and / or moved. During write ( Figure 5A ), the larger V MDTJ (~0.5-1 V) is applied, and the DW movement is guided by the spin transfer torque (STT) from the two tunnel barriers. Figure 5B ), the smaller V MDTJ (~0.1 V) is applied, and TMG (or TMR) is dominated by the thicker tunnel barrier.

[0050] Figure 6A-6D The bidirectional / symmetric operation of the MDTJ of the present invention is shown, where the spin transfer torques from the two tunnel barriers act in a synchronized manner to determine the direction of the magnetic DW motion. Table 2 below summarizes the spin transfer torques for Figure 6A-6D The relationship between the electron flow polarity, DW movement direction and conductivity (i.e., increase or decrease in conductivity) of the example shown in FIG. Figure 4A ) can have two polarities: up-down (↓↑) or down-up (↓↑). Specifically, Figure 4A The magnetic free layer shown in FIG. 1 comprises down-up (↓↑) domain walls. It should be noted that for a given current polarity, domain walls with opposite polarities will move in opposite directions, resulting in the same change in conductivity, i.e., regardless of which polarity of domain walls are created and / or moved, the change in conductivity is determined by the polarity of the current.

[0051] When V MDTJ Polarity makes the electron flow from the top electrode to the bottom electrode, and the magnetic DW will move to the left, such as FIG. 6A to FIG. 6B This results in a decrease in conductivity (increase in resistance), as shown by Fig.6D The portion of the graph indicated by arrow 602 is shown in FIG. 6 . Reversing the voltage polarity causes magnetic DW to move in the opposite direction (i.e., to the right). Switching V MDTJ The polarity causes the magnetic DW to move in opposite directions, thus presenting a bidirectional and symmetrical response, such as FIG. 6B to FIG. 6C This results in an increase in conductivity (a decrease in resistance), as shown by Figure 3DThe portion of the graph pointed by arrow 604 is shown.

[0052] Current polarity DW Sports Electrical conductivity (G) bottom up right Reduce Top to bottom Left Increase

[0053] Table 2

[0054] Fig.11A A non-limiting example of a method of the present invention for adjusting the conductivity associated with a synaptic weight element is depicted. When the conductivity of the synaptic weight element is to be increased, the upper electrical contact ( Figure 1A layer 102) and the lower electrical contact ( Figure 1A A first bias voltage V is applied between the layers 114 in the middle, wherein the first bias voltage V causes a first insulating tunnel barrier ( Figure 1A The first voltage V1 across the layer 106) causes the second insulating tunnel barrier ( Figure 1A The first bias voltage V is applied to the magnetic free layer (110) along the first direction. Figure 1A At least one magnetic domain wall is created and / or moved in a layer 108-108 in a first direction, wherein the creation and / or movement of at least one magnetic domain wall in a first direction linearly increases the conductivity of the synaptic weight element.

[0055] exist Fig.11A In the case of decreasing the conductivity of the synaptic weight element, the upper electrical contact ( Figure 1A layer 102) and the lower electrical contact ( Figure 1A A second bias voltage V' is applied between the layer 114 in the middle, wherein the polarity of the second bias voltage V' is opposite to the polarity of the first bias voltage V and the second bias voltage V, causing the first insulating tunnel barrier ( Figure 1A The third voltage V3 across the layer 106 in the middle) causes the second insulating tunnel barrier ( Figure 1A The fourth voltage V4 is applied to both ends of the magnetic free layer ( 110 in the embodiment of the present invention). Please note that, as described above, V'=V3+V4 and V3≠V4. The second bias voltage V' is applied to the magnetic free layer ( 110 in the embodiment of the present invention) in a second direction opposite to the first direction. Figure 1A The method further comprises creating and / or moving at least one magnetic domain wall in a layer 108 in the second direction, wherein the creation and / or movement of the at least one magnetic domain wall in the second direction linearly reduces the conductivity of the synaptic weight element.

[0056] Fig. 11B A non-limiting example of a method of the present invention for adjusting the conductivity associated with a synaptic weight element is depicted. When the conductivity of the synaptic weight element is to be increased, the upper electrical contact ( Figure 1A layer 102) and the lower electrical contact ( Figure 1A A first bias voltage V is applied between the layers 114 in the middle, wherein the first bias voltage V guides the upper electrical contact ( Figure 1A layer 102) to the lower electrical contact ( Figure 1A The first bias voltage V is applied in the first direction to the magnetic free layer ( Figure 1A Creating and / or moving at least one magnetic domain wall in a layer 108 in a first direction, wherein the creation and / or movement of at least one magnetic domain wall in a first direction linearly increases the conductivity of the synaptic weight element.

[0057] exist Fig. 11B In the case of decreasing the conductivity of the synaptic weight element, the upper electrical contact ( Figure 1A layer 102) and the lower electrical contact ( Figure 1A A second bias voltage V' is applied between the layers 114 in the lower portion, wherein the second bias voltage V' has the same polarity as the first bias voltage V and causes a current to flow from the lower electrical contact ( Figure 1A layer 114) to the upper electrical contact ( Figure 1A The second bias voltage V' is applied to the magnetic free layer ( Figure 1A The method further comprises creating and / or moving at least one magnetic domain wall in a layer 108 in the second direction, wherein the creation and / or movement of the at least one magnetic domain wall in the second direction linearly reduces the conductivity of the synaptic weight element.

[0058] It should be noted that Fig.11A and Fig. 11B In , there are double arrows linking the boxes for increasing and decreasing conductance, because the value of conductance may have to be decreased (from a higher value to a lower value) or increased (from a lower value to a higher value) depending on the desired synaptic weight.

[0059] Example Device Metrics: Non-limiting examples of devices fabricated according to the teachings of the present invention are provided below. The specific numbers provided are provided by way of example only.

[0060] Example requirements (per device):

[0061] Technical constraints:

[0062] 1.V write ~1.0V (v1=0.1V, v2=0.9V), and Max(I write )~10uA→R P ~100kΩ or G P ~10uS

[0063] 2.V read ~0.1V(v1=0.01V,v2=0.09V), and Max(I read )~0.2uA→R P~500kΩ or GP~2uS

[0064] Physical constraints (during write operations):

[0065] 1. Minimum current density for DW motion, j min ~10 6 A / cm 2

[0066] 2. {v1 / t1,v2 / t2}<1V / nm

[0067] Cross-sectional area (MTJ): A device (w×L)=10 3 nm 2 (For example, w = 10nm & L = 100nm)

[0068] →RA(MTJ)~300Ω 2 ; (tunnel barrier t1 ~ 1.1nm and t2 ~ 1.5nm thick)

[0069] →TMR~200%,G P / G AP ~3

[0070] Listed here are several possibilities for the resistance range of the synaptic device, depending on the chosen RA of the device.

[0071] 100 resistance states: 300kΩ to 900kΩ in 6kΩ steps, or

[0072] 100 resistance states: 200kΩ to 600kΩ in 4kΩ steps, or

[0073] 100 resistance states: 100 kΩ to 300 kΩ in 2 kΩ steps

[0074] For w = 10nm, L = 100nm

[0075] Controlled DW motion with 1-2nm resolution → 100-50 discrete linear & bidirectional steps DW speed near threshold current ~10nm / ns → pulse length ~0.1-0.5ns = 2-10GHz Energy per write (@1GHz): 10fJ; Power per device: 10uW

[0076] In one embodiment, the MDTJ device is part of a 2D crossbar array, wherein the MDTJ device of the present invention can be in various shapes. Figure 7 Depicted is a general example of a crossbar array that can be used, for example, in neuromorphic computing.

[0077] Figure 8A-8CDepicted is a non-limiting example of a crossbar array having an array of annular synaptic MDTJ devices in accordance with the teachings of the present invention, where the devices may be closed annular structures (eg, Figure 8B as shown) or an open annular structure (as Figure 8C Opening device (see Figure 8C ) can have any number of magnetic DWs (i.e., an odd or even number of magnetic DWs), and the closing device (see Figure 8B ), for example a ring device will have an even number of magnetic DWs.

[0078] It should be noted that although Figure 8A-8C Only two variations are shown (closed and open ring devices), but other shapes of MDTJ devices of the present invention are also within the scope of the present invention.

[0079] For example, Figure 9A-9C Another non-limiting example of a crossbar array having an array of square-shaped synaptic MDTJ devices according to the teachings of the present invention is depicted, wherein the square-shaped devices may be closed square-shaped structures (such as Fig. 9B as shown) or an open square shaped structure (as shown Fig. 9C shown).

[0080] Figures 10A-10D Top views of a closed ring MDTJ device, an open ring MDTJ device, a closed square MDTJ device, and an open square MDTJ device are depicted, respectively.

[0081] The MDTJ device approach of the present invention meets all the key attributes of 2-terminal synaptic devices necessary for next generation neuromorphic devices and provides symmetrical, continuous and linear weight updates with ultrafast GHz operation and 10fJ energy per device per part write and 0.1fJ per device read.

[0082] in conclusion

[0083] In the above aspects and embodiments, a system and method for effectively implementing non-volatile, linear, symmetric and bidirectional synaptic weight elements based on magnetic domain walls is shown. Although various preferred embodiments have been shown and described, it should be understood that this is not intended to limit the invention by such disclosure, but is intended to cover all modifications that fall within the spirit and scope of the invention as defined by the appended claims.

Claims

1. A method for regulating the conductivity of a synaptic weight element, The synaptic weight element comprises: upper electrical contacts; a first magnetic reference layer directly beneath the upper electrical contact; a first insulating tunnel barrier directly beneath the first magnetic reference layer; a magnetic free layer directly below the first insulating tunnel barrier; a second insulating tunnel barrier directly below the magnetic free layer; a second magnetic reference layer directly below the second insulating tunnel barrier, and a lower electrical contact directly below the second magnetic reference layer, the first and second insulating tunnel barriers having different resistances; The method comprises: When the conductivity of the synaptic weight element is to be increased, a first bias voltage V is applied between the upper electrical contact and the lower electrical contact, so that the first bias voltage V causes a first voltage V1 across the first insulating tunnel barrier and causes a second voltage V2 across the second insulating tunnel barrier, wherein V=V1+V2 and V1≠V2, wherein the first bias voltage V creates and / or moves at least one magnetic domain wall in the magnetic free layer in a first direction, wherein the creation and / or movement of the at least one magnetic domain wall in the first direction increases the conductivity of the synaptic weight element; as well as When the conductivity of the synaptic weight element is to be reduced, a second bias voltage V' is applied between the upper electrical contact and the lower electrical contact, the polarity of the second bias voltage V' being opposite to the polarity of the first bias voltage V, the second bias voltage V' causing a third voltage V3 across the first insulating tunnel barrier and causing a fourth voltage V4 across the second insulating tunnel barrier, wherein V'=V3+V4 and V3≠V4, and the second bias voltage V' creates and / or moves at least one magnetic domain wall in the magnetic free layer in a second direction opposite to the first direction, and wherein the creation and / or movement of the at least one magnetic domain wall in the second direction reduces the conductivity of the synaptic weight element.

2. The method of claim 1, wherein the synaptic weight elements are used in an artificial neural network (ANN). The method of claim 1 , wherein the synaptic weight elements are used for data storage. The method of claim 1 , wherein the synaptic weight elements are used for neuromorphic computing. The method of claim 1 , wherein the synaptic weight element is cylindrical.

6. The method according to claim 1, wherein the conductivity G is given by G=G P (1-TMG*d / L) where Gp is ​​the parallel conductivity of the device and TMG is given by (G P -G AP ) / G P The tunnel permeance is given by G P is the parallel conductivity, G AP is the antiparallel conductivity of the device, L is the length of the device, and d is the position of the at least one magnetic domain.

7. The method of claim 1, wherein the synaptic weight element is part of a crossbar array. 8 . The method of claim 1 , wherein the synaptic weight element is made into at least one of the following shapes: a ring, an open ring, a square, or an open square.

9. The method of claim 1 , wherein the first insulating tunnel barrier is thinner than the second insulating tunnel barrier and the first magnetic reference layer is formed of an in-plane magnetic material, or the second insulating tunnel barrier is thinner than the first insulating tunnel barrier and the second magnetic reference layer is formed of an in-plane magnetic material. 10 . The method of claim 1 , wherein magnetization states of the first magnetic reference layer and the second magnetic reference layer are antiparallel to each other. 11 . The method of claim 1 , wherein magnetization states of the first magnetic reference layer and the second magnetic reference layer are parallel to each other.

12. A device comprising a synaptic weight element, the synaptic weight element comprising: upper electrical contacts; a first magnetic layer directly beneath the upper electrical contact; a first tunnel barrier disposed below the first magnetic layer; a second magnetic free layer located below the first tunnel barrier; a second tunnel barrier located below the second magnetic free layer, the first tunnel barrier and the second tunnel barrier having different resistances; a third magnetic layer located below the second tunnel barrier; a lower electrical contact directly below the third magnetic layer, wherein the upper electrical contact, the first magnetic layer, the first tunnel barrier, the second magnetic free layer, the second tunnel barrier, the third magnetic layer, and the lower electrical contact are proximate to each other to form a layer stack, wherein a first bias voltage applied between the upper electrical contact and the lower electrical contact induces a first voltage across the first tunnel barrier and induces a second voltage across the second tunnel barrier, the first bias voltage creating and / or moving at least one magnetic domain wall in the second magnetic free layer in a first direction, wherein the creation and / or movement of the at least one magnetic domain wall in the first direction increases the conductivity of the synaptic weight element, and wherein a second bias voltage applied between the upper electrical contact and the lower electrical contact causes a third voltage across the first tunnel barrier and causes a fourth voltage across the second tunnel barrier, the second bias voltage creates and / or moves at least one magnetic domain wall in the second magnetic free layer in a second direction opposite to the first direction, wherein the creation and / or movement of the at least one magnetic domain wall in the second direction reduces the conductivity of the synaptic weight element, and the polarity of the second bias voltage is opposite to the polarity of the first bias voltage.

13. The device according to claim 12, wherein the device is a linear, bidirectional, symmetric synaptic weight element in an artificial neural network (ANN).

14. The device of claim 12, wherein the device stores data. The device of claim 12 , wherein the layer stack is cylindrical. 16 . The device of claim 12 , wherein the first tunnel barrier is thinner than the second tunnel barrier and the first magnetic layer is formed of an in-plane magnetic material, or the second tunnel barrier is thinner than the first tunnel barrier and the third magnetic layer is formed of an in-plane magnetic material. 17 . The device of claim 12 , wherein magnetization states of the first magnetic layer and the third magnetic layer are antiparallel to each other. 18 . The device of claim 12 , wherein magnetization states of the first magnetic layer and the third magnetic layer are parallel to each other.

Images