RECURRENT NEURON IMPLEMENTATION BASED ON MAGNETO-ELECTRIC SPIN ORBIT LOGIC

• US 20200134419A1
• Filed: 10/30/2018
• Published: 04/30/2020
• Est. Priority Date: 10/30/2018
• Status: Active Grant

• Current Assignee: Intel Corporation
• Original Assignee: Intel Corporation
• Inventors: Manipatruni, Sasikanth, Nikonov, Dmitri, Young, Ian
• Granted Patent: US 11,410,021 B2
• Time in Patent Office: Days
• CPC Class Codes: G06N 3/044 Recurrent networks, e.g. Hopfield networks
• CPC Class Codes: G06N 3/063 using electronic means
• CPC Class Codes: H01F 10/329 Spin-exchange coupled multilayers wherein the magnetisation of the free layer is switched by a spin-polarised current, e.g. spin torque effect
• CPC Class Codes: H01F 41/32 for applying conductive, insulating or magnetic material on a magnetic film , specially adapted for a thin magnetic film
• CPC Class Codes: H03K 19/18 using galvano-magnetic devices, e.g. Hall-effect devices
• CPC Class Codes: H10B 61/00 Magnetic memory devices, e.g. magnetoresistive RAM [MRAM] devices
• CPC Class Codes: H10N 50/01 Manufacture or treatment
• CPC Class Codes: H10N 50/20 Spin-polarised current-controlled devices magnetoresistive devices H10N50/10
• CPC Class Codes: H10N 50/80 Constructional details
• CPC Class Codes: H10N 50/85 Magnetic active materials
• CPC Class Codes: H10N 59/00 Integrated devices, or assemblies of multiple devices, comprising at least one galvanomagnetic or Hall-effect element covered by groups H10N50/00 - H10N52/00 MRAM devices H10B61/00

Abstract

Techniques are provided for implementing a recurrent neuron (RN) using magneto-electric spin orbit (MESO) logic. An RN implementing the techniques according to an embodiment includes a first MESO device to apply a threshold function to an input signal provided at a magnetization port of the MESO device, and scale the result by a first weighting factor supplied at an input port of the MESO device to generate an RN output signal. The RN further includes a second MESO device to receive the RN output signal at a magnetization port of the second MESO device and generate a scaled previous RN state value. The scaled previous state value is a scaled and time delayed version of the RN output signal based on a second weighting factor. The RN input signal is a summation of the scaled previous state value of the RN with weighted synaptic input signals provided to the RN.

Claims

1. A recurrent neuron comprising:

a first magneto-electric spin orbit (MESO) device to apply a threshold function to an input signal supplied through a magnetization port of the first MESO device, and to scale the thresholded input signal by a first weighting factor supplied through an input port of the first MESO device to generate an output signal at an output port of the first MESO device; and

a second MESO device to receive the output signal at a magnetization port of the second MESO device and to generate a scaled previous state value at an output port of the second MESO device, wherein the scaled previous state value is a scaled and time delayed version of the output signal, the scaling based on a second weighting factor supplied through an input port of the second MESO device;

wherein the input signal is a summation of the scaled previous state value with one or more weighted synaptic input signals provided to the magnetization port of the first MESO device.

2. The recurrent neuron of claim 1, wherein the first and second MESO devices comprise a charge-to-spin conversion node (CS node), a spin-to-charge conversion node (SC node), and a magnet coupling the CS node to the SC node, the magnetization port coupled to the CS node, the input port coupled to a supply electrode of the SC node, and the output port coupled to a spin orbit effect stack of the SC node.

3. The recurrent neuron of claim 2, wherein the CS node is to control a magnetization of the magnet based on a magnet setting current supplied to the magnetization port, and the magnet comprises a ferromagnetic material configured to retain the magnetization in a nonvolatile state.

4. The recurrent neuron of claim 3, wherein the SC node is to generate an output current at the spin orbit effect stack, delivered through the output port, based on the magnetization of the magnet and based on an input current applied to the supply electrode of the input port.

5. The recurrent neuron of claim 1, wherein the threshold function is a hyperbolic tangent function associated with a saturation curve of conversion from current to magnetization.

6. The recurrent neuron of claim 1, wherein the time delay of the scaled previous state value is based on a timing of application of the second weighting factor supplied through the input port of the second MESO device relative to a timing of application of the output signal at the magnetization port of the second MESO device.

7. The recurrent neuron of claim 1, wherein the weighted synaptic input signal is generated by a third MESO device configured to apply a third weighting factor to a synaptic input signal, the third weighting factor supplied through a magnetization port of the third MESO device, and the synaptic input signal supplied through an input port of the third MESO device.

8. An integrated circuit recurrent neural network (RNN) comprising two or more of the recurrent neurons of claim 1.

9. The integrated circuit RNN of claim 8, further comprising a variable resistor to couple an output port of a first of the recurrent neurons to a synaptic input port associated with a second of the recurrent neurons.

10. A processor comprising the integrated circuit RNN of claim 8.

11. A method of forming a recurrent neuron, the method comprising:

forming a first magneto-electric spin orbit (MESO) device on a substrate, the MESO device comprising a charge-to-spin conversion node (CS node), a spin-to-charge conversion node (SC node), a magnet coupling the CS node to the SC node, a magnetization port coupled to the CS node, an input port coupled to a supply electrode of the SC node, and an output port coupled to a spin orbit effect stack of the SC node;

forming a second MESO device on the substrate, the magnetization port of the second MESO device coupled to the output port of the first MESO device; and

forming a summing junction on the substrate, the summing junction coupled to the output port of the second MESO device and to the magnetization port of the first MESO device.

12. The method of claim 11, wherein the first MESO device is to apply a threshold function to an input signal supplied through the magnetization port of the first MESO device, and to scale the thresholded input signal by a first weighting factor supplied through the input port of the first MESO device to generate an output signal at the output port of the first MESO device.

13. The method of claim 12, wherein the second MESO device is to receive the output signal at the magnetization port of the second MESO device and to generate a scaled previous state value at the output port of the second MESO device, wherein the scaled previous state value is a scaled and time delayed version of the output signal, the scaling based on a second weighting factor supplied through the input port of the second MESO device.

14. The method of claim 13, wherein the summing junction is to sum the scaled previous state value with one or more weighted synaptic input signals provided to the magnetization port of the first MESO device.

15. The method of claim 14, further comprising forming a third MESO device on the substrate, to generate the weighted synaptic input signal.

16. The method of claim 15, wherein the third MESO device is to apply a third weighting factor to a synaptic input signal, the third weighting factor supplied through the magnetization port of the third MESO device, and the synaptic input signal supplied through the input port of the third MESO device.

17. The method of claim 11, wherein the CS node is to control a magnetization of the magnet based on a magnet setting current supplied to the magnetization port, and the magnet comprises a ferromagnetic material configured to retain the magnetization in a nonvolatile state.

18. The method of claim 17, wherein the SC node is to generate an output current at the spin orbit effect stack, delivered through the output port, based on the magnetization of the magnet and based on an input current applied to the supply electrode of the input port.

19. The method of claim 11, wherein the substrate is a semiconductor substrate.

20. The method of claim 11, wherein the substrate is a semiconductor wafer.

21. The method of claim 11, wherein the substrate is a printed circuit board.