WO2018182694A1 - Procédés et appareil pour neurones magnétoélectriques dans des réseaux neuronaux - Google Patents

Procédés et appareil pour neurones magnétoélectriques dans des réseaux neuronaux Download PDF

Info

Publication number
WO2018182694A1
WO2018182694A1 PCT/US2017/025413 US2017025413W WO2018182694A1 WO 2018182694 A1 WO2018182694 A1 WO 2018182694A1 US 2017025413 W US2017025413 W US 2017025413W WO 2018182694 A1 WO2018182694 A1 WO 2018182694A1
Authority
WO
WIPO (PCT)
Prior art keywords
magnet
current
magnetization
meso
artificial neuron
Prior art date
Application number
PCT/US2017/025413
Other languages
English (en)
Inventor
Dmitri E. Nikonov
Sasikanth Manipatruni
Ian A. Young
Original Assignee
Intel Corporation
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Intel Corporation filed Critical Intel Corporation
Priority to PCT/US2017/025413 priority Critical patent/WO2018182694A1/fr
Publication of WO2018182694A1 publication Critical patent/WO2018182694A1/fr

Links

Classifications

    • H—ELECTRICITY
    • H01—ELECTRIC ELEMENTS
    • H01L—SEMICONDUCTOR DEVICES NOT COVERED BY CLASS H10
    • H01L29/00—Semiconductor devices specially adapted for rectifying, amplifying, oscillating or switching and having potential barriers; Capacitors or resistors having potential barriers, e.g. a PN-junction depletion layer or carrier concentration layer; Details of semiconductor bodies or of electrodes thereof ; Multistep manufacturing processes therefor
    • H01L29/66—Types of semiconductor device ; Multistep manufacturing processes therefor
    • H01L29/66984—Devices using spin polarized carriers
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
    • G06N3/063—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using electronic means
    • G06N3/065—Analogue means
    • H—ELECTRICITY
    • H10—SEMICONDUCTOR DEVICES; ELECTRIC SOLID-STATE DEVICES NOT OTHERWISE PROVIDED FOR
    • H10N—ELECTRIC SOLID-STATE DEVICES NOT OTHERWISE PROVIDED FOR
    • H10N50/00—Galvanomagnetic devices
    • H10N50/20—Spin-polarised current-controlled devices
    • H—ELECTRICITY
    • H10—SEMICONDUCTOR DEVICES; ELECTRIC SOLID-STATE DEVICES NOT OTHERWISE PROVIDED FOR
    • H10N—ELECTRIC SOLID-STATE DEVICES NOT OTHERWISE PROVIDED FOR
    • H10N59/00—Integrated devices, or assemblies of multiple devices, comprising at least one galvanomagnetic or Hall-effect element covered by groups H10N50/00 - H10N52/00
    • B—PERFORMING OPERATIONS; TRANSPORTING
    • B82—NANOTECHNOLOGY
    • B82Y—SPECIFIC USES OR APPLICATIONS OF NANOSTRUCTURES; MEASUREMENT OR ANALYSIS OF NANOSTRUCTURES; MANUFACTURE OR TREATMENT OF NANOSTRUCTURES
    • B82Y10/00—Nanotechnology for information processing, storage or transmission, e.g. quantum computing or single electron logic
    • B—PERFORMING OPERATIONS; TRANSPORTING
    • B82—NANOTECHNOLOGY
    • B82Y—SPECIFIC USES OR APPLICATIONS OF NANOSTRUCTURES; MEASUREMENT OR ANALYSIS OF NANOSTRUCTURES; MANUFACTURE OR TREATMENT OF NANOSTRUCTURES
    • B82Y25/00—Nanomagnetism, e.g. magnetoimpedance, anisotropic magnetoresistance, giant magnetoresistance or tunneling magnetoresistance

Definitions

  • This disclosure relates generally to integrated circuit devices, and, more particularly, to methods and apparatus for magnetoelectric neurons in neural networks.
  • Machine learning applications such as object recognition or deep learning, and other forms of big data processing often involve implementing algorithms using neural networks that mimic the problem solving functionality of a biological brain. While neural networks may be expressed mathematically, neural networks may also be physically implemented using multiple interconnected units of circuitry referred to as artificial neurons.
  • An artificial neuron typical includes one or more inputs to receive signals that undergo some mathematical processes (e.g., a summation) to produce an output signal that may serve as an input to one or more other artificial neurons in the neural network.
  • FIG. 1 is a block diagram of the operation of an example artificial neuron.
  • FIG. 2 is an example magnetoelectric spin orbit (MESO) device for use in an example artificial neuron in accordance with the teachings disclosed herein.
  • MEO magnetoelectric spin orbit
  • FIG. 3 is the example MESO device of FIG. 2 with the magnetization of the magnet having switched directions.
  • FIG. 4 is an example artificial neuron constructed using multiple ones of the MESO device of FIGS. 2 and 3.
  • FIG. 5 is a flowchart of an example method to manufacture the example artificial neuron of FIG. 4.
  • FIG. 6 is a block diagram of an example processor system associated with one or more semiconductor fabrication machines to execute example machine readable instructions represented at least in part by the example method of FIG. 5 to manufacture the example artificial neuron of FIG 4.
  • any part e.g., a layer, film, area, or plate
  • any part is in any way positioned on (e.g., positioned on, located on, disposed on, or formed on, etc.) another part
  • the referenced part is either in contact with the other part, or that the referenced part is above the other part with one or more intermediate part(s) located therebetween.
  • Stating that any part is in contact with another part means that there is no intermediate part between the two parts.
  • An artificial neuron is a building block of neural networks. Generally speaking, the function of an artificial neuron is to produce an output (y) that is a non-linear function n of one or more inputs (x ; ) as follows:
  • w ; - is stored weights corresponding to the associated inputs Xj of the neuron
  • n is the number of inputs Xj
  • ⁇ is a threshold or transfer function (also known as an activation function).
  • the transfer function ⁇ is a step-like transfer function with an S-shaped curve (e.g., a sigmoid curve) that produces an output of 1 for high positive inputs, an output of -1 for strongly negative inputs, and intermediate output values
  • the transfer function may be step-like, artificial neurons typically use continuously variable analog signals with a smooth transition between the high and low values.
  • FIG. 1 is a block diagram of the operation of an example artificial neuron 100 implemented according to Equation 1 described above.
  • the neuron 100 includes a number of inputs 102 (e.g., xi, x 2 , ... XN), each of which is multiplied by a corresponding weight 104 (e.g., wi, w 2 , ... WN).
  • a summation 106 is applied to the products of the inputs 102 and associated weights 104.
  • a bias 107 is provided as an additional input to the summation 106.
  • the bias 107 is a threshold offset bias for the neuron 100.
  • the summation total 108 is evaluated according to a transfer function 110 to produce a final output 112.
  • the inputs 102 correspond to outputs of other artificial neurons within the same neural network.
  • the output 1 12 of the example neuron 100 may serve as an input to one or more other neurons within the neural network.
  • CMOS complementary metal-oxide-semiconductor
  • CMOS complementary metal-oxide-semiconductor
  • ferromagnetic materials e.g., a nanomagnet
  • Spintronic-based neurons implemented using nanomagnets are advantageous in that they may be implemented with considerably fewer components than comparable CMOS-based neurons.
  • existing spin- based implementations of artificial neurons are relatively inefficient in terms of both speed and power.
  • some known spin-based artificial neurons rely on spin transfer torque switching of nanomagnets, which can have a switching time of more than 10 nanoseconds and use more than 100 femtojoules (10 5 J) of energy.
  • the signals between the inputs and outputs of such spin-based neurons are carried by spin polarized currents, which attenuate after a relatively short distance (e.g., less than 1 micrometer), thereby limiting the overall size and/or associated total number of inputs possible for such neurons.
  • Examples disclosed herein implement an artificial neuron using magnetoelectric switching of a magnet that exhibits switching times of less than 1 nanosecond. Furthermore, magnetoelectric switching can be accomplished with approximately 10 attojoules (10 "18 J) of energy, which is comparable to the switching energy of CMOS transistors and orders of magnitude less than the power requirements for the spin transfer torque switching techniques mentioned above. Further still, example artificial neurons disclosed herein carry signals between components of the neuron (e.g., between the inputs and the output) using electric charge currents rather than spin polarized currents. As a result, the distance between the inputs and the output (and, thus, the number of inputs possible for a single neuron and/or its associated size) is effectively unlimited within practical constraints.
  • FIG. 2 is an example magnetoelectric spin orbit (MESO) device 200 for use in an example artificial neuron (e.g., the artificial neuron 400 of FIG. 4).
  • the example MESO device 200 includes a magnet 202 (e.g., a nanomagnet) extending between a charge-to-spin conversion node 204 and a spin-to-charge conversion node 206 of the MESO device 200.
  • the MESO device 200 may be electrically coupled to other components (e.g., other MESO devices) via first and second electrical interconnects 208, 210 connected to respective ones of the charge-to-spin conversion node 204 and the spin-to-charge conversion node 206.
  • the first and second electrical interconnects 208, 210 may be formed of any suitable conductive material such as, for example, copper (Cu). As described more fully below, the first interconnect 208 may serve as an input to receive signals used to establish or set a magnetization of the magnet 202. The second interconnect 210 may serve as an output interconnect to carry an output signal to other components (e.g., another MESO device).
  • Cu copper
  • the magnet 202 is formed of a ferromagnetic material such as, for example, cobalt (Co), iron (Fe), nickel (Ni), gadolinium (Gd), their alloys, or a Heusler alloy of the form X2YZ or XYZ where X, Y, Z can be elements of cobalt (Co), iron (Fe), nickel (Ni), aluminum (Al), germanium (Ge), gallium (Ga), gadolinium (Gd), manganese (Mn), etc.
  • the magnet 202 may be magnetized under an electric field and retain the magnetization even after the electric field is removed.
  • the magnet 202 is elongate having a length that is between 2 and 5 times a width of the magnet 202. In some examples, the aspect ratio (length to width) is approximately 3.
  • the elongate nature of the magnet 202 results in two steady states for the magnet 202 with the direction of magnetization 212 extending along the length the magnet 202. That is, the direction of magnetization 212 of the magnet 202 is represented in a first steady state in FIG. 2 and in a second steady state (with an opposite direction of magnetization 212) in FIG. 3.
  • the charge-to-spin conversion node 204 of the MESO device 200 converts an electric charge current into spin (i.e., magnetization).
  • the dielectric material 216 is a multiferroic material that exhibits both ferroelectric properties (e.g., can be electrically polarized with or without an applied electric field) and magnetic properties (e.g., may exhibit surface spin polarization which can be switched under the application of an extemal electric field).
  • the dielectric material 216 may be formed of any suitable magnetoelectric material such as, for example, bismuth ferrite (BiFeCb, BFO), chromium(III) oxide ((3 ⁇ 403) or magnesium oxide (MgO).
  • the dielectric material 216 is formed of a single material that directly produces a magnetoelectric effect.
  • the material 216 may be formed of a combination of materials such as multiple layers of oxides and intermetallics that define a dielectric stack. Such combination of materials may achieve a magnetoelectric effect through, for example, cascading of two transductions or physical phenomena in materials (e.g., cascading of a voltage to strain transduction and a strain to magnetization transduction).
  • a magnet setting electrical current (represented by the arrow 218) carried by the first interconnect 208 produces a voltage across the magnetoelectric dielectric material 216 of the capacitor 214.
  • FIG. 2 illustrates a positive magnet setting current 218 resulting in
  • FIG. 3 illustrates a negative magnet setting current 218 (shown pointing in the opposite direction to FIG. 2) resulting in ferroelectric polarization in the dielectric material 216 with the negative charge 222 adjacent the interconnect 208 and the positive charge 220 adjacent the magnet 202.
  • the charge accumulates in the capacitor 214, the spin of electrons in the dielectric material 216 at the interface between the material 216 and the magnet 202 become aligned to form surface spin polarization, which is effectively a magnet.
  • the direction of magnetization (spin) of the electrons in the surface spin polarization is defined by the direction of ferroelectric polarization within the dielectric material 216.
  • the direction of magnetization 212 of the magnet 202 may be switched between a first direction as shown in FIG. 2 (associated with the positive magnet setting current 218) and a second direction as shown in FIG. 3 (associated with the negative magnet setting current 218).
  • the change in the direction of magnetization 212 of the magnet 202 induced by the charge-to-spin conversion node 204 affects the output of the spin-to-charge conversion node 206 located at the opposite end of the magnet 202.
  • the spin-to-charge conversion node 206 of the example MESO device 200 of FIG. 2 converts spin (e.g., the magnetization 212 of the magnet 202) to an electric charge current.
  • the example spin-to-charge conversion node 206 includes a spin orbit effect stack 224 coupled to the magnet 202. Further, as shown in the illustrated example of FIG.
  • the spin-to-charge conversion node 206 includes a supply electrode 226 adjacent the magnet 202 opposite the spin orbit effect stack 224 and a ground electrode 228 adjacent the spin orbit effect stack 224 opposite the magnet 202.
  • the supply and ground electrodes 226, 228 may be formed of copper or any other suitable interconnect material. In the illustrated example, the ground electrode 228 is connected to ground.
  • Each of the supply electrode 226, the magnet 202, the spin orbit effect stack 224, and the ground electrode 228 are conductive metals.
  • a supply charge current (represented by arrow 230) received at the supply electrode 226 may pass through the supply electrode 226, the magnet 202, the spin orbit effect stack 224, and the ground electrode 228 to ground.
  • the magnetization 212 in the magnet 202 will produce a spin polarized current in which a substantial majority (e.g., greater than 80%) of electrons associated with the supply charge current 230 will exhibit spin (i.e., magnetization) having a direction corresponding to the magnetization 212 of the magnet 202.
  • the strength of the spin polarized current (e.g., the proportion of electrons that align with the magnet 202) is proportional to the strength of the magnetization 212 of the magnet 202.
  • the spin orbit effect stack 224 includes a non-magnetic metal material 236 such as silver (Ag), aluminum (Al), gold (Au), or copper (Cu) that forms a heterostructure with a spin orbit coupling material 238 that exhibits a spin orbit effect in a metallic system.
  • Example materials for the spin orbit coupling material 238 include elements of group V of the Periodic Table of Elements and their alloys (e.g., bismuth (Bi), bismuth-silver alloys) or traditional interconnect materials (e.g., copper (Cu), gold (Au), silver (Ag), aluminum (Al)) doped with high atomic weight transition elements.
  • the spin orbit effect stack 224 may include a spacer formed of a non-magnetic material (e.g., silver or copper) between the magnet 202 and the non-magnetic metal material 236.
  • the non-magnetic material 236 is disposed or deposited on the spin orbit coupling material 238 to define an interface 240 therebetween.
  • the interface 240 of the non-magnetic metal material 236 and the spin orbit coupling material 238 defines a high density 2D electron gas that has the ability to produce a strong or high spin orbit effect (also referred to as spin orbit coupling).
  • the spin orbit coupling that occurs at the interface 240 of the two materials 236, 238 of the spin orbit effect stack 224 is referred to as the Rashba-Edel stein effect (or Rashba effect for short).
  • the Rashba effect results in different electrons associated with the supply charge current 230 being deflected in opposite directions along the 2D electron gas of the interface 240.
  • the direction which the electrons are deflected depends upon the direction of spin of the electrons.
  • the supply charge current 230 becomes a spin polarized current as it passes through the magnet 202 and enters the spin orbit effect stack 224, a majority of the electrons (with spins aligned with the magnetization 212 of the magnet 202) will be deflected in one direction with a minority of the electrons being deflected in the opposite direction.
  • the MESO device 200 is constructed so that the direction of deflection of the electrons due to the Rashba effect within the spin orbit effect stack 224 is either into or away from the output interconnect 210, which serves as an output for the MESO device 200. More particularly, in the illustrated example, the deflection of electrons produced by the Rashba effect is along an axis substantially perpendicular to both the electric current (e.g., the supply charge current 230) and the spin polarized current (e.g., corresponding to the direction of magnetization 212 of the magnet 202), the two of which are substantially perpendicular to each other.
  • the electric current e.g., the supply charge current 230
  • the spin polarized current e.g., corresponding to the direction of magnetization 212 of the magnet 202
  • the output interconnect 210 is positioned substantially perpendicular to the magnet 202 (and associated direction of magnetization 212) and substantially perpendicular to the direction of the supply charge current 230.
  • the spin orbit effect stack 224 deflects a majority of electrons either into or away from the output interconnect 210, thereby resulting in an output charge current 242 along the interconnect 210 that is proportional to the supply charge current 230.
  • a residual charge current 246 will pass through the spin orbit effect stack 224 to ground.
  • a positive supply charge current 230 in the positive Z direction
  • a positive spin current corresponding to the magnetization 212 in the positive Y direction
  • a positive output charge current 242 in the positive X direction
  • the output charge current 242 in the illustrated examples is represented as conventional current flow, which is opposite electron flow.
  • the spin orbit effect stack 140 deflects a maj ority of electrons (with spin aligned with the magnetization 212 of the magnet 202) into the output interconnect 210 in FIG. 2 (to produce a negative output charge current 242) and away from the output interconnect 210 in FIG. 3 (to produce a positive output charge current 242).
  • the magnet setting current 218 and the supply charge current 230 are provided during separate operations implemented at different times. More particularly, in the illustrated example, providing the magnet setting current 218 may be compared to a write operation that sets or adjusts the direction of magnetization 212 of the magnet 202. Further, in the illustrated example, providing the supply charge current 230 may be compared to a read operation that produces the output charge current 242, which is proportional to the magnetization 212 of the magnet 202 previously established during the write operation associated with the magnet setting current 218.
  • FIGS. 2 and 3 show the spin orbit effect stack 224 with one layer of the non-magnetic metal material 236 and one layer of the spin orbit coupling material 238, in other examples, multiple layers of materials may be included in a superlattice to define multiple interfaces that exhibit the Rashba effect.
  • the spin orbit effect stack 224 is formed of a single material that intrinsically exhibits a spin orbit effect to produce spin orbit coupling within its bulk. Spin orbit coupling occurring intrinsically within a single material is referred to as the bulk effect or the spin Hall effect. Example materials for such
  • implementations of the spin orbit effect stack 224 include materials with a high spin Hall effect coefficient (e.g., on the order of 0.01 to 10 or greater (e.g., 0.1 to 1)) such as tantalum (Ta), tungsten (W), bismuth (Bi), bismuth selenide (Bi2Se3), or platinum (Pt), or high atomic weight transition elements such as lutetium (Lu), hafnium (Hf), rhodium (Rh), osmium (Os), iridium (Ir), gold (Au), or mercury (Hg).
  • a high spin Hall effect coefficient e.g., on the order of 0.01 to 10 or greater (e.g., 0.1 to 1)
  • materials with a high spin Hall effect coefficient e.g., on the order of 0.01 to 10 or greater (e.g., 0.1 to 1)
  • a high spin Hall effect coefficient e.g., on the order of 0.01 to 10 or greater (e.g.,
  • the spin Hall effect produces the same result as the Rashba effect except that the electrons are deflected within the bulk of the material rather than at the interface 240 of a heterostructure of two materials as shown in FIGS. 2 and 3.
  • the spin orbit effect stack 224 may exhibit both the Rashba effect and the spin Hall effect.
  • FIG. 4 is an example artificial neuron 400 constructed from a plurality of interconnected MESO devices 402, 404, 406, 408 constructed substantially the same as the example MESO device 200 of FIGS. 2 and 3.
  • the illustrated example of FIG. 4 is depicted with all components within a two-dimensional plane for purposes of explanation. However, as mentioned above, in actual implementation the output interconnect 210a-d for each MESO device 402, 404, 406, 408 in the illustrated example would be oriented substantially perpendicular to the corresponding magnet 202a-d to produce a corresponding output charge current based on spin orbit coupling within the spin orbit effect stack 224a-d.
  • the supply electrode 226a-d of each MESO device 402, 404, 406, 408 is shown at a different location on the
  • FIGS. 2 and 3 to illustrate that the position of the electrode is not significant to the design of the MESO devices 402, 404, 406, 408.
  • each of the first three MESO devices 402, 404, 406 corresponds to a respective input of the artificial neuron 400.
  • the fourth MESO device 408 corresponds to the output of the artificial neuron 400.
  • the artificial neuron 400 may be formed with one more MESO device than the total number of inputs for the artificial neuron 400.
  • each of the MESO devices 402, 404, 406, 408 are fabricated to be substantially identical. That is, the dimensions, structure, and associated materials of the components are substantially consistent between different ones of the MESO devices 402, 404, 406, 408. This reduces the complexity and number of fabrication processes needed to produce such an artificial neuron 400.
  • some or all of the MESO devices 402, 404, 406, 408 are formed using the identical or similar fabrication processes. Furthermore, in some examples, some or all of the MESO devices 402, 404, 406, 408 may be fabricated at the same time. Of course, in other examples, specific ones of the MESO devices 402, 404, 406, 408 may be constructed with a different size, shape, or material than other ones of the MESO devices 402, 404, 406, 408 as appropriate with resulting differences in the fabrication processes for such devices.
  • the inputs (Xi, X 2 , .. . XN) to the artificial neuron 400 are charge currents (Ii, I 2 , ... IN) received at the supply electrode 226a-c associated with each input MESO device 402, 404, 406. That is, the inputs (Xi, X 2 , .. . XN) correspond to the supply charge current 230 shown and described in connection with FIGS. 2 and 3. In some examples, the inputs (Xi, X 2 , .. . XN) correspond to outputs produced by other artificial neurons in a neural network.
  • the magnetization 410, 412, 414 of each magnet 202a-c of each input MESO device 402, 404, 406 produces a spin polarized current as electrons associated with the supply charge currents (Ii, I 2 , . . . IN) magnetically align as they pass through the respective magnets 202a-c.
  • the strength of the spin polarized current e.g., the number of electrons that exhibit a spin (magnetization) aligned with the magnets 202a-c) is
  • the magnetization 410, 412, 414 of the magnet 202a-c in each input MESO device 402, 404, 406 serves as a weighting factor multiplier (e.g., the weights 104 of FIG.
  • each MESO device 402, 404, 406 corresponds to weighted charge currents (I ou ti, Iout2, ... IoutNr) that represent the inputs multiplied by respective weights.
  • the magnetization 410, 412, 414 of the magnet 202a-c in each input MESO device 402, 404, 406 may be controlled by an electric charge current (e.g., the magnet setting current 218 of FIGS. 2 and 3) applied at the corresponding interconnect 208a-c of each MESO device 402, 404, 406 to induce a magnetoelectric effect via the magnetoelectric dielectric material 216a-c as described above.
  • the magnetization 410, 412, 414 for each input MESO device 402, 404, 406 is set once and remains fixed during operation thereafter. This example corresponds to a neuron with fixed weights.
  • different magnet setting currents at the input interconnect 208a-c can cause the magnetization 410, 412, 414 to switch or change.
  • the value of the magnetization 410, 412, 414 for each input MESO device 402, 404, 406 may be binary in which the magnets 202a-c are switched between one of two stable states.
  • the magnets 202a-c are in stable states when the corresponding magnetization 410, 412, 414 of the magnet 202a-c is in the same direction throughout the magnets 202a-c.
  • a first stable state is represented in the first MESO device 402 of FIG. 4 with the magnetization 410 pointing in a first (e.g., positive) direction while the second stable state is represented in the third MESO device 406 of FIG. 4 with the magnetization 410 pointing in a second (e.g., negative) direction.
  • the value of the magnetization 410, 412, 414 is equal in both states but has an opposite sign (e.g., +1 in the first direction and -1 in the second direction).
  • the corresponding output charge current 416, 418, 420 is proportional to the input current except that the sign may be changed.
  • the magnetization 410 of the first input MESO device 402 has a value of +1 (steady state in the positive direction) resulting in a corresponding positive output charge current 416
  • the magnetization 414 of the last input MESO device 406 has a value of -1 (steady state in the negative direction) resulting in a corresponding negative output charge current 420.
  • intermediate values for the magnetization 410, 412, 414 of the magnets 202a-c may be achieved by controlling the position of a domain wall 422 dividing distinct magnetic domains within the magnets 202a-c of the input MESO devices 402, 404, 406.
  • the switch is not instantaneous.
  • a domain wall 422 will form adjacent the dielectric material 216 and then propagate across the entire magnet 202a-c (given a sufficient amount of time and/or a sufficient amount of energy) to reverse the magnetization of the magnet 202a-c.
  • the duration and/or energy provided to charge the dielectric material 216a-c is controlled so that the domain wall 422 does not reach the end of the magnet 202a-c but stops somewhere between the opposing ends.
  • Some magnets, such as those used in the binary switching example described above, may be limited to two stable states in which the magnetization is entirely in one direction or the other along the length of the magnet. In such examples, a domain wall formed in the middle region of the magnet is unlikely to remain in place but will drift toward one end of the magnet or the other. However, some example magnets may be designed to maintain the position of a domain wall with relatively high precision.
  • domain walls typically correspond to a region of a magnet that is approximately 10-15 nanometers wide such that magnets with dimensions greater than this range are needed to sustain a domain wall. Even if a domain wall can be formed within a magnet, there is still the possibility that the domain wall may drift, for example, due to thermal fluctuations. This is less of a concern for relatively large magnets but may also be at least partially mitigated against by creating pinning sites (e.g., notches) on the magnet at desired locations for the domain wall.
  • pinning sites e.g., notches
  • the domain wall 422 in the magnet 202b of the second input MESO device 404 results in a magnetization 412 that includes a first (negative) magnetization 412a in a first magnetic domain and a second (positive) magnetization 412b in a second magnetic domain.
  • magnetizations 412a, 412b are in opposite directions.
  • the opposing magnetic fields act against each other in producing the resulting spin polarized current that enters the spin orbit effect stack 224b.
  • the output charge current 418 may have a value somewhere between the other two output charge currents 416, 420 associated with the respective magnets 202a, 202c in steady states associated with a single magnetic domain.
  • the magnetization 412 of the second input MESO device 404 results in a weight (W2) that is an intermediate value between the weights (Wi and WN) of the other two input MESO devices 402, 406 corresponding to the respective magnetizations 410, 414.
  • the magnetization of the magnets 202a-d may be controlled to a range of different values.
  • the range may correspond to a plurality of discrete values associated with both steady states and one or more fixed positions for a domain wall 422 (e.g., defined by pinning site(s)).
  • the range of values for the magnetization may be substantially continuous between both steady states by precisely controlling the position of the domain wall 422 along the magnet 202a-d.
  • the position of the domain wall 422 is only relevant in defining the weight for each respective input when the domain wall 422 is within the area of the magnet 202a-d interfacing with the spin orbit effect stack 224.
  • the reason for this is that only the portions of the magnetic fields above the spin orbit effect stack 224a-d, as shown in the illustrated example of FIG. 4, will be involved in producing the spin polarized current provided to the spin orbit effect stack 224a-d.
  • the magnetization 412 in the second MESO device 404 shows a much larger domain with the negative magnetization 412a than the domain with the positive magnetization 412b, within the region directly above the spin orbit effect stack 224b, the magnetization is predominantly positive (i.e., predominantly corresponds to the positive magnetization 412b).
  • the weighting factor (W2) for the second input (X2) will be much closer to the weighting factor (Wi) for the first input (Xi) corresponding to a value of +1 than to the weighting factor (WN) for the last input (XN) corresponding to a value of -1.
  • the weight for an input corresponds to the average magnetization of the associated magnet 202a-c within the region of the magnet 202a-c associated with the interface between the magnet 202a-c and the spin orbit effect stack 224a-c.
  • the weight i.e., the average magnetization
  • the weight may be mathematically expressed as the sum of the products of the magnetization of each domain (+1 or -1) and the proportion of the spin orbit effect stack 224a-c the domain is directly adjacent.
  • the magnetization of the magnet 202b in the second MESO device 404 was the same as shown in the magnet 202d of the output MESO device 408 in FIG.
  • the weight (W2) for the second input (X2) would be -1 because the domain with the positive magnetization (pointing right) and the associated domain wall 422 are not within the area of the magnet 202d adjacent the interface between the magnet 202d and the spin orbit effect stack 224d.
  • the magnetization resulting from the domain wall 422 positioned as shown in the output MESO device 408 is effectively the same as the magnetization 414 of the third MESO device 407 that is in a steady state with only one magnetic domain.
  • different input MESO devices 402, 404, 406 may be constructed to provide output charge currents 416, 418, 420 (I ou ti, Iout2, ⁇ ⁇ ⁇ IoutNr) that are proportional to an input current (Ii, I 2 , ... IN) by a weighting factor (Wi, W 2 , . . . WN) corresponding to the magnetization 410, 412, 414 of the respective magnets 202a-c.
  • Each of the output charge currents 416, 418, 420 is produced along the output interconnect 210a-c associated with each input MESO device 402, 404, 406.
  • the output interconnects 210a-c of the input MESO devices 402, 404, 406 are connected in parallel to a combined conductive channel or
  • the value of the summation total 108 (based on a summation of the output charge currents provided by each input MESO device 402, 404, 406) is communicated from the input MESO devices 402, 404, 406 to the output MESO device 408 as an electric current.
  • the length of the conductive channel 424 and, thus, the number of input MESO devices that may be included in the artificial neuron 400 is effectively unlimited within practical constraints. This is a significant improvement over other known artificial neurons that rely on spin polarized currents to communicate weighted inputs to an output device because spin polarized currents are limited to relatively short distances (e.g., less than 1 micrometer) due to signal attenuation.
  • the MESO devices 402, 404, 406, 408 produce spin polarized currents that are used to implement the artificial neuron 400, the usage of the spin polarized current is locally limited to each specific MESO device 402, 404, 406, 408 (within the spin-to- charge conversion node 206 described in FIGS. 2 and 3).
  • designing large neurons with many input MESO devices (and/or inputs spaced far apart from each other and/or an output MESO device) is not a difficulty.
  • the magnet 202d of the output MESO device 408 is designed to be induced with a magnetization 428 that is substantially continuously variable across a range of values.
  • the magnetization 428 ranges between -1 when the magnet 202d is in a steady state (single domain) with a negative magnetization and +1 when the magnet 202d is in a steady state with a positive magnetization.
  • intermediate values of the magnetization may be achieved when a domain wall 422 is established within the magnet 202d adjacent the spin orbit effect stack 224d of the output MESO device 408 as described above.
  • the magnetization 428 of the magnet 202d of the output MESO device 408 may change (e.g., switch direction or at least change value) based on the magnetoelectric effect of the dielectric material 216d electrically polarized by the summation charge current 426. More particularly, the magnetoelectric effect depends upon the charge accumulated within the dielectric material 216d, which is based on the summation charge current 426 and the duration the current is applied.
  • the magnetization 428 within the magnet 202d will produce an output charge current with an intermediate value somewhere between the high (+1) and low (-1) limits.
  • the magnetization of the magnet 202d of the output MESO device 408 effectively represents a step-like transfer function for the artificial neuron 400 (e.g., corresponding to the transfer function 110 of FIG. 1).
  • an output charge current based on the magnetization 428 as shown in the illustrated example will have a value of -1 because the domain wall 422 is spaced apart from the spin orbit effect stack 224d.
  • switching (or adjusting) the magnetization of the magnet 202d of the output MESO device 408 does not itself produce an output signal for the artificial neuron 400.
  • the resulting magnetization 428 of the magnet 202 serves to define the output signal when such is generated.
  • the summation charge current 426 is applied for a fixed amount of time to charge the dielectric material 216d and cause the magnetization 428 of the magnet 202d to change based on the accumulated charge. As described above, the magnetization will remain even after the summation charge current 426 is removed because the magnet 202d is a ferromagnetic material.
  • a reading of the magnetization 428 may be accomplished (to define a final output for the artificial neuron 400) by applying a voltage between the supply electrode 226d and the ground electrode 228d of the output MESO device 408 to produce an output supply charge current that passes through the magnet 202d and the spin orbit effect stack 224d.
  • the applied voltage is denoted with a T (identified by reference number 430) to indicate that the voltage 430 serves as the threshold offset bias for the artificial neuron 400 (e.g., corresponding to the bias 107 of FIG. 1).
  • the output supply charge current produces a spin polarized current that is injected into the spin orbit effect stack 224d. Due to a majority of electrons having a spin
  • the spin orbit effect stack 224d will deflect a majority of the electrons either into or away from the output interconnect 210d, thereby producing a final output current for the artificial neuron 400.
  • the final output current corresponds to the output signal of the artificial neuron 400 that is proportional to the magnetization of the magnet 202d.
  • the final output signal may be stored and/or used as an input into another artificial neuron within a neural network.
  • FIG. 5 is a flowchart of an example method to manufacture the example artificial neuron 400 of FIG. 4.
  • the example method begins at block 502 by forming ground electrodes 228a-d for one or more input MESO devices 402, 404, 406 and an output MESO device 408 on a substrate (e.g., a semiconductor wafer).
  • the example method forms a spin orbit effect stack 224a-d for each of the MESO device 402, 404, 406, 408 on the corresponding ground electrode 228a-d.
  • the example method includes forming output interconnects 210a-d for each MESO device 402, 404, 406, 408 connected adjacent each corresponding spin orbit effect stack 224a-d.
  • the example method includes forming input interconnects 208a-d for each MESO device 402, 404, 406, 408.
  • the example method includes forming a magnetoelectric dielectric material 216a-d on the input interconnect 208a-d of each MESO device 402, 404, 406, 408.
  • the example method includes forming an elongate magnet 202a-d for each MESO device 402, 404, 406, 408 with a first end coupled to the corresponding magnetoelectric dielectric material 216a-d and a second end coupled to the corresponding spin orbit effect stack 224a-d.
  • the example method includes forming a conductive channel to connect the output interconnects 210a-c of the one or more input MESO device 402, 404, 406 to the input interconnect 208d of the output MESO device 408.
  • the example method includes forming supply electrodes on the magnet 202a-d of each MESO device 402, 404, 406, 408. Thereafter, the example method of FIG. 5 ends.
  • FIG. 6 is a block diagram of an example processor platform 600 capable of controlling one or more semiconductor fabrication machines to execute the method of FIG. 5 to manufacturing the example artificial neuron 400 of FIG. 4.
  • the processor platform 600 can be any type of computing device.
  • the processor platform 600 of the illustrated example includes a processor 612.
  • the processor 612 of the illustrated example is hardware.
  • the processor 612 can be implemented by one or more integrated circuits, logic circuits, microprocessors or controllers from any desired family or manufacturer.
  • the processor 612 of the illustrated example includes a local memory 613 (e.g., a cache).
  • the processor 612 of the illustrated example is in communication with a main memory including a volatile memory 614 and a non-volatile memory 616 via a bus 618.
  • the volatile memory 614 may be implemented by Synchronous Dynamic Random Access Memory (SDRAM), Dynamic Random Access Memory (DRAM), RAMBUS Dynamic Random Access Memory (RDRAM) and/or any other type of random access memory device.
  • the non-volatile memory 616 may be implemented by flash memory and/or any other desired type of memory device. Access to the main memory 614, 616 is controlled by a memory controller.
  • the processor platform 600 of the illustrated example also includes an interface circuit 620.
  • the interface circuit 620 may be implemented by any type of interface standard, such as an Ethemet interface, a universal serial bus (USB), and/or a PCI express interface.
  • one or more input devices 622 are connected to the interface circuit 620.
  • the input device(s) 622 permit(s) a user to enter data and commands into the processor 612.
  • the input device(s) can be implemented by, for example, an audio sensor, a microphone, a camera (still or video), a keyboard, a button, a mouse, a touchscreen, a track-pad, a trackball, isopoint and/or a voice recognition system.
  • One or more output devices 624 are also connected to the interface circuit 620 of the illustrated example.
  • the output devices 624 can be implemented, for example, by display devices (e.g., a light emitting diode (LED), an organic light emitting diode (OLED), a liquid crystal display, a cathode ray tube display (CRT), a touchscreen, a tactile output device, a light emitting diode (LED), a printer and/or speakers).
  • the interface circuit 620 of the illustrated example thus, typically includes a graphics driver card, a graphics driver chip or a graphics driver processor.
  • the interface circuit 620 of the illustrated example also includes a communication device such as a transmitter, a receiver, a transceiver, a modem and/or network interface card to facilitate exchange of data with external machines (e.g., computing devices of any kind) via a network 626 (e.g., an Ethemet connection, a digital subscriber line (DSL), a telephone line, coaxial cable, a cellular telephone system, etc.).
  • a network 626 e.g., an Ethemet connection, a digital subscriber line (DSL), a telephone line, coaxial cable, a cellular telephone system, etc.
  • the processor platform 600 of the illustrated example also includes one or more mass storage devices 628 for storing software and/or data. Examples of such mass storage devices 628 include floppy disk drives, hard drive disks, compact disk drives, Blu-ray disk drives, RAID systems, and digital versatile disk (DVD) drives.
  • Coded instructions 632 to implement the method of FIG. 5 may be stored in the mass storage device 628, in the volatile memory 614, in the non-volatile memory 616, and/or on a removable tangible computer readable storage medium such as a CD or DVD.
  • a non-transitory computer readable storage medium is expressly defined to include any type of computer readable storage device and/or storage disk and to exclude propagating signals and to exclude transmission media.
  • the above disclosed methods, apparatus and articles of manufacture enable the formation of magnetoelectric artificial neurons that may be implemented in a neural network to perform neuromorphic computing based on analog input signals and output signals. More particularly, the disclosed examples use the magnetization of a magnet as a weighting factor for an (electric current) input of an artificial neuron.
  • the magnetization i.e., the weighting factor
  • the magnetization may be switched and/or adjusted based on the magnetoelectric effect.
  • Using the magnetoelectric effect provides switching times that are much faster and require less energy than other spintronic-based artificial neurons and much fewer components than comparable CMOS-based artificial neurons.
  • the disclosed examples use the magnetization of an additional magnet to implement the transfer function for the artificial neuron to generate a final output signal.
  • the magnetization of the magnet at the output of the neuron is also switched and/or adjusted based on a magnetoelectric effect. More particularly, the output magnet is magnetized based on an accumulated charge associated with electrical charge currents proportional to the inputs of the neuron.
  • the use of electrical currents to transmit input signals to an output MESO device enable the production of artificial neurons that are larger (e.g., have more inputs and/or have components spaced farther apart) than what is possible using other known spintronic-based techniques for artificial neurons.
  • Example 1 is an artificial neuron for a neural network that includes a first magnetoelectric spin orbit (MESO) device including a first magnet.
  • the first MESO device converts an input current to a charge current having a proportional relationship with the input current.
  • the input current corresponds to an input signal of the artificial neuron.
  • the artificial neuron further includes a second MESO device including a second magnet.
  • a magnetization of the second magnet is to vary based on the charge current.
  • the magnetization of the second magnet to define an output signal of the artificial neuron.
  • Example 2 includes the subj ect matter of Example 1, wherein a magnetization of the first magnet defines the proportional relationship between the charge current and the input current.
  • Example 3 includes the subj ect matter of Example 2, wherein a magnetization of the first magnet serves as a weighting factor multiplier for the input signal of the artificial neuron.
  • Example 4 includes the subj ect matter of any one of Examples 1 -3, wherein the first MESO device includes a magnetoelectric dielectric material disposed between the first magnet and an electrical interconnect.
  • a magnet setting current at the electrical interconnect is to produce ferroelectric polarization within the magnetoelectric dielectric material.
  • the ferroelectric polarization is to alter the magnetization of the first magnet.
  • Example 5 includes the subj ect matter of Example 4, wherein the ferroelectric polarization is to alter the magnetization of the first magnet between a first value associated with a first direction of magnetization of the first magnet and a second value associated with a second direction of magnetization of the first magnet opposite the first direction.
  • Example 6 includes the subj ect matter of Example 5, wherein the ferroelectric polarization is to alter the magnetization of the first magnet to an intermediate value between the first value and the second value.
  • the intermediate value is based on a domain wall formed within the first magnet.
  • Example 7 includes the subject matter of Example 6, wherein the intermediate value is substantially continuously variable across a range based on a position of the domain wall within the first magnet.
  • Example 8 includes the subject matter of any one of Examples 1-7, wherein the first magnet is to convert the input current into a spin polarized current.
  • Example 9 includes the subject matter of Example 8, wherein the first MESO device includes a spin orbit effect stack electrically coupled between the first magnet and a conductive channel.
  • the spin orbit effect stack produces spin orbit coupling in response to the spin polarized current to generate the charge current along the conductive channel.
  • Example 10 includes the subject matter of any one of Examples 1-9, wherein the second MESO device is to provide an output charge current corresponding to the output signal of the artificial neuron.
  • the output charge current is proportional to a magnetization of the second magnet.
  • Example 11 includes the subject matter of any one of Examples 1-10, wherein the artificial neuron further includes a conductive channel electrically interconnecting the first MESO device to the second MESO device.
  • Example 12 includes the subject matter of Example 11, wherein the conductive channel is to carry the charge current from an output interconnect of the first MESO device to an input interconnect of the second MESO device.
  • Example 13 includes the subject matter of any one of Examples 11 or 12, wherein the second MESO device includes a magnetoelectric dielectric material coupled between the first magnet and the conductive channel. A summation charge current in the conductive channel to produce ferroelectric polarization within the magnetoelectric dielectric material. The summation charge current is based on the charge current. The magnetization of the second magnet is to vary based on changes in the ferroelectric polarization.
  • Example 14 includes the subject matter of Example 13, wherein the input current is a first input current, the input signal is a first input signal, and the current charge is a first current charge.
  • the artificial neuron further includes a third MESO device including a third magnet.
  • the third MESO device to receive a second charge current corresponding to a second input signal of the artificial neuron and to provide a second charge current proportional to the second charge current along the conductive channel.
  • the second charge current is to combine with the first charge current to form the summation charge current.
  • Example 15 includes the subject matter of any one of Examples 13 or 14, wherein variation in the magnetization of the second magnet serves as a transfer function for the artificial neuron.
  • Example 16 includes the subject matter of any one of Examples 13-15, wherein variation in the magnetization of the second magnet is based on switching a direction of magnetization of the second magnet.
  • Example 17 includes the subject matter of any one of Examples 13-16, wherein variation in the magnetization of the second magnet is based on a position of a domain wall within the second magnet.
  • Example 18 includes the subject matter of any one of Examples 1-17, wherein the first MESO device and the second MESO device have a structure that is substantially identical.
  • Example 19 is an artificial neuron for a neural network that includes an output magnetoelectric spin orbit (MESO) device to generate an output signal for the artificial neuron, the output MESO device including: an input interconnect, a magnet, and a magnetoelectric dielectric material disposed between the input interconnect and the magnet.
  • the artificial neuron further includes a conductive channel electrically coupling an input MESO device to the input interconnect.
  • MESO magnetoelectric spin orbit
  • Example 20 includes the subject matter of Example 19, wherein the input MESO device is to provide a charge current via the conductive channel to the output MESO device.
  • Example 21 includes the subject matter of Example 20, wherein a magnetization of the magnet is to vary based on a magnetoelectric effect produced by the magnetoelectric dielectric material when charged by the charge current.
  • Example 22 includes the subject matter of Example 21, wherein the output signal of the output MESO device is based on the magnetization of the magnet.
  • Example 23 includes the subject matter of any one of Examples 21 or 22, wherein variation in the magnetization of the magnet serves as a transfer function for the artificial neuron to generate the output signal.
  • Example 24 includes the subject matter of any one of Examples 21-23, wherein variation in the magnetization of the magnet is substantially continuously adjustable based on a position of a domain wall between two magnetic domains in the magnet.
  • Example 25 includes the subject matter of any one of Examples 21-24, wherein the output MESO device includes a spin orbit effect stack coupled to the magnet.
  • the output signal corresponds to an output charge current generated from spin orbit coupling of a spin polarized current in the spin orbit effect stack.
  • the spin polarized current is based on the
  • Example 26 includes the subject matter of any one of Examples 19-25, wherein the input MESO device includes: a second magnetoelectric dielectric material, a spin orbit effect stack, a second magnet coupled to the second magnetoelectric dielectric material and the spin orbit effect stack, and a supply electrode coupled to the second magnet.
  • the supply electrode is to receive an input current corresponding to an input signal of the artificial neuron.
  • the spin orbit effect stack is to produce a charge current based on input signal.
  • the conductive channel is to carry the charge current to the output MESO device.
  • Example 27 includes the subject matter of Example 26, wherein the charge current is proportional to a magnetization of the second magnet.
  • Example 28 includes the subject matter of Example 27, wherein the magnetization of the second magnet is defined based on a magnet setting current used to charge the second magnetoelectric dielectric material.
  • Example 29 includes the subject matter of any one of Examples 27 or 28, wherein the magnetization of the second magnet is adjustable between a first stable state corresponding to a first direction of magnetization and a second stable state corresponding to a second direction of magnetization.
  • Example 30 is a method to manufacture an artificial neuron for a neural network that includes forming a first magnetoelectric spin orbit (MESO) device on a substrate. The method further includes forming a second MESO device on the substrate. The method also includes forming a conductive channel to electrically interconnect the first MESO device and the second MESO device. The first MESO device is to provide a charge current to the second MESO device via the conductive channel. The charge current is proportional to an input current received by the first MESO device.
  • MEO magnetoelectric spin orbit
  • Example 31 includes the subject matter of Example 30, wherein forming the first MESO device involves the same processes used to form the second MESO device.
  • Example 32 includes the subject matter of any one of Examples 30 or 31, wherein the first and second MESO device are formed at the same time.
  • Example 33 includes the subject matter of any one of Examples 30-32, wherein the method further includes forming the second MESO device by: forming a magnetoelectric dielectric material on an interconnect electrically coupled to the conductive channel, and forming a magnet on the magnetoelectric dielectric material. The charge current is to adjust a magnetization of the magnet. The second MESO device is to produce an output signal based on the magnetization of the magnet.
  • Example 34 includes the subject matter of any one of Examples 30-33, wherein the input current corresponds to an input signal of the artificial neuron.
  • Example 35 is a method that includes converting, via a first magnetoelectric spin orbit (MESO) device, an input current to a charge current. The charge current corresponding to a proportion of the input current defined by a magnetization of a first magnet in the first MESO device. The method further includes transmitting, via a conductive channel, the charge current to a second MESO device. The method also includes magnetizing a second magnet in the second MESO device based on the charge current.
  • MEO magnetoelectric spin orbit
  • Example 36 includes the subject matter of Example 35, wherein the method further includes: receiving the input current at a supply electrode of the first MESO device, conducting the input current through the first magnet to produce a spin polarized current, and converting the spin polarized current to the charge current via a spin orbit effect stack of the first MESO device.
  • Example 37 includes the subject matter of any one of Examples 35 or 36, wherein the method further includes: receiving a magnet setting current at an input interconnect of the first MESO device, and adjusting the magnetization of the first magnet based on the magnet setting current.
  • Example 38 includes the subject matter of any one of Examples 35-37, wherein the method further includes: charging a magnetoelectric dielectric material in the second MESO device based on the charge current, and adjusting the magnetization of the second magnet based on the charge accumulated in the magnetoelectric dielectric material.
  • Example 39 includes the subject matter of Example 38, wherein the second MESO device includes a spin orbit effect stack coupled to the second magnet.
  • the method further includes applying a voltage across the second magnet and the spin orbit effect stack to generate a supply current passing through the second magnet and the spin orbit effect stack.
  • the method also includes generating an output current that is proportional to the supply current based on spin orbit coupling within the spin orbit effect stack.
  • the output current corresponds to an output signal of an artificial neuron of a neural network.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Biomedical Technology (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Microelectronics & Electronic Packaging (AREA)
  • Power Engineering (AREA)
  • Biophysics (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Neurology (AREA)
  • General Health & Medical Sciences (AREA)
  • Condensed Matter Physics & Semiconductors (AREA)
  • Artificial Intelligence (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Computer Hardware Design (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Ceramic Engineering (AREA)
  • Hall/Mr Elements (AREA)

Abstract

L'invention concerne des procédés et un appareil pour des neurones magnétoélectriques dans des réseaux neuronaux. Un exemple de neurone artificiel pour un réseau neuronal comprend un premier dispositif de spin-orbite magnétoélectrique (MESO) comprenant un premier aimant. Le premier dispositif MESO convertit un courant d'entrée en un courant de charge ayant une relation proportionnelle avec le courant d'entrée. Le courant d'entrée correspond à un signal d'entrée du neurone artificiel. Le neurone donné à titre d'exemple comprend en outre un second dispositif MESO comprenant un second aimant. Une magnétisation du second aimant est à varier sur la base du courant de charge. La magnétisation du second aimant définit un signal de sortie du neurone artificiel.
PCT/US2017/025413 2017-03-31 2017-03-31 Procédés et appareil pour neurones magnétoélectriques dans des réseaux neuronaux WO2018182694A1 (fr)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/US2017/025413 WO2018182694A1 (fr) 2017-03-31 2017-03-31 Procédés et appareil pour neurones magnétoélectriques dans des réseaux neuronaux

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/US2017/025413 WO2018182694A1 (fr) 2017-03-31 2017-03-31 Procédés et appareil pour neurones magnétoélectriques dans des réseaux neuronaux

Publications (1)

Publication Number Publication Date
WO2018182694A1 true WO2018182694A1 (fr) 2018-10-04

Family

ID=63677752

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/US2017/025413 WO2018182694A1 (fr) 2017-03-31 2017-03-31 Procédés et appareil pour neurones magnétoélectriques dans des réseaux neuronaux

Country Status (1)

Country Link
WO (1) WO2018182694A1 (fr)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP3632840A1 (fr) * 2018-10-05 2020-04-08 IMEC vzw Agencement destiné à être utilisé dans un multiplicateur de vecteur de matrice
WO2020136267A1 (fr) * 2018-12-28 2020-07-02 Thales Dispositif électronique, porte numérique, composant analogique et procédé de génération d'une tension
CN113161478A (zh) * 2021-03-08 2021-07-23 湖北大学 基于人工智能的自旋电子声音识别器件及制备方法与应用
WO2023248050A1 (fr) * 2022-06-24 2023-12-28 International Business Machines Corporation Structures magnétiques pour manipulation résonante de spin

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20120109866A1 (en) * 2010-10-29 2012-05-03 International Business Machines Corporation Compact cognitive synaptic computing circuits
US20130117209A1 (en) * 2011-11-09 2013-05-09 Qualcomm Incorporated Method and apparatus for using memory in probabilistic manner to store synaptic weights of neural network
US20150269478A1 (en) * 2014-01-29 2015-09-24 Purdue Research Foundation Compact implementation of neuron and synapse with spin switches
WO2016105436A1 (fr) * 2014-12-26 2016-06-30 Intel Corporation Logique spin-orbite avec interconnexions de charge et nœuds magnétoélectriques
WO2017044095A1 (fr) * 2015-09-09 2017-03-16 Intel Corporation Logique de spin à électrodes à effet hall de spin et interconnexions de charge

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20120109866A1 (en) * 2010-10-29 2012-05-03 International Business Machines Corporation Compact cognitive synaptic computing circuits
US20130117209A1 (en) * 2011-11-09 2013-05-09 Qualcomm Incorporated Method and apparatus for using memory in probabilistic manner to store synaptic weights of neural network
US20150269478A1 (en) * 2014-01-29 2015-09-24 Purdue Research Foundation Compact implementation of neuron and synapse with spin switches
WO2016105436A1 (fr) * 2014-12-26 2016-06-30 Intel Corporation Logique spin-orbite avec interconnexions de charge et nœuds magnétoélectriques
WO2017044095A1 (fr) * 2015-09-09 2017-03-16 Intel Corporation Logique de spin à électrodes à effet hall de spin et interconnexions de charge

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP3632840A1 (fr) * 2018-10-05 2020-04-08 IMEC vzw Agencement destiné à être utilisé dans un multiplicateur de vecteur de matrice
US11599138B2 (en) 2018-10-05 2023-03-07 Imec Vzw System and method for applying a magnonic matrix-vector-multiplier arrangement
WO2020136267A1 (fr) * 2018-12-28 2020-07-02 Thales Dispositif électronique, porte numérique, composant analogique et procédé de génération d'une tension
FR3091412A1 (fr) * 2018-12-28 2020-07-03 Thales Dispositif électronique, porte numérique, composant analogique et procédé de génération d’une tension
US11417453B2 (en) 2018-12-28 2022-08-16 Thales Electronic device, digital port, analog component, and method for generating a voltage
CN113161478A (zh) * 2021-03-08 2021-07-23 湖北大学 基于人工智能的自旋电子声音识别器件及制备方法与应用
CN113161478B (zh) * 2021-03-08 2022-08-16 湖北大学 基于人工智能的自旋电子声音识别器件及制备方法与应用
WO2023248050A1 (fr) * 2022-06-24 2023-12-28 International Business Machines Corporation Structures magnétiques pour manipulation résonante de spin

Similar Documents

Publication Publication Date Title
Kim et al. Silicon synaptic transistor for hardware-based spiking neural network and neuromorphic system
WO2018182694A1 (fr) Procédés et appareil pour neurones magnétoélectriques dans des réseaux neuronaux
US11411172B2 (en) Magnetoelectric spin orbit logic based full adder
US9864950B2 (en) Compact implementation of neuron and synapse with spin switches
Manipatruni et al. Spin-orbit logic with magnetoelectric nodes: A scalable charge mediated nonvolatile spintronic logic
CN107112413B (zh) 具有充电互连和磁电节点的自旋轨道逻辑单元
US9379712B2 (en) High speed precessionally switched magnetic logic
US10777247B1 (en) Spin-based storage element
Kim et al. 3-bit multilevel operation with accurate programming scheme in tio x/al2o3 memristor crossbar array for quantized neuromorphic system
EP3123537A1 (fr) Dispositifs logiques à paroi de domaine magnétique et interconnexion
Kim et al. Double MgO-based perpendicular magnetic tunnel junction for artificial neuron
CN110770737A (zh) 包括神经网络的半导体装置
Leonard et al. Stochastic domain wall-magnetic tunnel junction artificial neurons for noise-resilient spiking neural networks
Roy Separating read and write units in multiferroic devices
Lone et al. Magnetic tunnel junction based implementation of spike time dependent plasticity learning for pattern recognition
CN111596891B (zh) 真随机数发生装置及发生方法
Kumar et al. Ultra-low power domain wall device for spin-based neuromorphic computing
US20220350572A1 (en) Random number generation unit and computing system
Nasab et al. High performance and low power spintronic binarized neural network hardware accelerator
KR102321838B1 (ko) 인시츄 퇴적된 자기 스택들로부터 스핀 로직 디바이스들을 제조하는 방법
CN109937483A (zh) 垂直磁电自旋轨道逻辑
WO2017155510A1 (fr) Spin et interconnexions de charge ayant des effets rashba
Lee et al. A thermodynamic core using voltage-controlled spin–orbit-torque magnetic tunnel junctions
Soni et al. Energy efficient ternary computation unit using differential spin Hall effect MRAM
Sivasubramani et al. Area efficient in-plane nanomagnetic multiplier and convolution architecture design

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 17903822

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 17903822

Country of ref document: EP

Kind code of ref document: A1