[go: up one dir, main page]

WO2024197775A1 - 自旋电子器件、电路单元、计算网络、算子以及神经网络 - Google Patents

自旋电子器件、电路单元、计算网络、算子以及神经网络 Download PDF

Info

Publication number
WO2024197775A1
WO2024197775A1 PCT/CN2023/085355 CN2023085355W WO2024197775A1 WO 2024197775 A1 WO2024197775 A1 WO 2024197775A1 CN 2023085355 W CN2023085355 W CN 2023085355W WO 2024197775 A1 WO2024197775 A1 WO 2024197775A1
Authority
WO
WIPO (PCT)
Prior art keywords
layer
domain wall
random
magnetic
electronic device
Prior art date
Application number
PCT/CN2023/085355
Other languages
English (en)
French (fr)
Inventor
邢国忠
王迪
汤瑞丰
刘明
Original Assignee
中国科学院微电子研究所
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 中国科学院微电子研究所 filed Critical 中国科学院微电子研究所
Priority to PCT/CN2023/085355 priority Critical patent/WO2024197775A1/zh
Publication of WO2024197775A1 publication Critical patent/WO2024197775A1/zh

Links

Classifications

    • 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

Definitions

  • the present disclosure relates to the field of magnetic domain wall technology, and more specifically, to a random spin electronic device, a read-write circuit unit, a random computing network, an operator of random computing, and a random pulse neural network.
  • Stochastic computing and neuromorphic computing based on spintronic devices provide low-power, high-performance solutions for edge computing in the context of high-density chip integration and intensive computing. At the same time, without losing computing accuracy, it can significantly improve computing stability and noise resistance, providing a more reliable solution for edge computing.
  • the core of random computing is the random device that can generate random pulse sequences.
  • Spintronic devices have extremely high durability and stability, and can generate random pulse sequences with fixed probabilities under high working intensity, with great application potential.
  • the current random devices based on spintronics are mainly based on low-barrier magnetic tunnel junction (MTJ) devices, which flip probabilistically under the disturbance of thermal noise, but the modulation of probability usually requires multiple methods such as STT current and external magnetic field, which is not conducive to the life of the device and integrated applications, and does not consider the intrinsic randomness of the magnetic domain wall (DW).
  • MTJ magnetic tunnel junction
  • the embodiments of the present disclosure provide a random spin electronic device, a read-write circuit unit, a random computing network, a random computing operator, and a random pulse neural network.
  • One aspect of an embodiment of the present disclosure provides a random spin electronic device, including:
  • a heater configured to flow a write pulse signal
  • an antiferromagnetic layer formed on the top surface of the isolation layer
  • An RKKY coupling layer is formed on the top surface of the ferromagnetic layer
  • a magnetic tunnel junction is formed on one side of the top surface of the RKKY coupling layer, wherein when a write pulse signal is input, a magnetic domain wall in a magnetic domain wall free layer of the magnetic tunnel junction moves.
  • the magnetic tunnel junction includes:
  • a magnetic domain wall free layer is formed on the top surface of the RKKY coupling layer, wherein when a control pulse signal is input, the magnetic domain wall in the magnetic domain wall free layer moves;
  • a barrier layer is formed on one side of the top surface of the magnetic domain wall free layer
  • a reference layer is formed on the top surface of the barrier layer, wherein the magnetic domain wall free layer, the ferromagnetic layer and the reference layer all have perpendicular magnetic anisotropy.
  • the random spin electronic device further includes:
  • Two antiferromagnetic pinning layers are respectively formed on both sides of the magnetic domain wall free layer, wherein the barrier layer is located on both sides of the two antiferromagnetic pinning layers.
  • the materials of the domain wall free layer, the reference layer and the ferromagnetic layer include at least one of the following: Co, CoFeB, Co, Pt, CoFeAl, Co, Pd, CoFe, the thickness of the reference layer is greater than the thickness of the magnetic domain wall free layer, or the material coercivity of the reference layer is greater than the material coercivity of the magnetic domain wall free layer, or the reference layer is pinned by an antiferromagnetic or synthetic antiferromagnetic pinning layer; the material of the barrier layer includes at least one of the following: MgO, Al 2 O 3 ;
  • the material of the antiferromagnetic layer includes at least one of the following: IrMn, FeMn, NiMn, CoMn, PtMn, Mn 2 Au, NiO, MnO;
  • the material of the RKKY coupling layer includes at least one of the following: Ru, Ta, W, Pt, Cu, V, Cr, Rh, Nd, Mo, Re;
  • the material of the isolation layer includes at least one of the following: SiO 2 , SiN;
  • the material of the heater includes a derivative of a metal oxide, nitride, carbide or boride, the heater has a resistivity that meets a preset resistance threshold, and is configured to generate sufficient Joule heat;
  • the Joule heat generated by the heater is configured to modulate the RKKY action strength of the magnetic domain wall free layer, the RKKY coupling layer and the ferromagnetic layer.
  • the magnetic domain wall in the domain wall free layer is driven to move in a direction away from the barrier layer under the action of the net magnetic field of the RKKY effective field and the built-in magnetic field;
  • the random spin electronic device based on the amplitude, number and pulse width of the control pulse signal, has at least one of the characteristics of accumulation-discharge, accumulation-leakage-discharge and random through the formed RKKY effective field and built-in magnetic field.
  • a read/write circuit unit including:
  • a random spin electronic device wherein the magnetic tunnel junction in the random spin electronic device is configured to input a read current, and a power supply voltage is connected to an end of the heater away from the magnetic tunnel junction to make the current flow into the heater;
  • a source and drain of a write transistor one end of which is connected to the other end of the heater, and an input end of the source and drain of the write transistor is configured to input a control signal
  • a current limiting resistor one end of which is connected to the other end of the source and drain of the write transistor
  • a comparator wherein a first end of the comparator is adjacent to the magnetic tunnel junction, and a second end of the comparator is configured to output a pulse signal.
  • Another aspect of the disclosed embodiments provides a random computing network, including:
  • a plurality of read-write circuit units A plurality of read-write circuit units
  • a digital-to-analog converter configured to convert an input digital signal into an analog signal and input a control signal to the input terminal of the source and drain of the write transistor in each of the read-write circuit units;
  • a microcontroller or processor wherein the microcontroller or processor is configured to perform partial derivative processing on the multiple pulse signals output by each of the above-mentioned read-write circuit units based on a cost function, and output the above-mentioned input signal corresponding to the above-mentioned read-write circuit unit, so that the above-mentioned digital-to-analog converter generates a control signal acting on the above-mentioned read-write circuit unit according to the above-mentioned input signal.
  • Another aspect of the embodiments of the present disclosure provides an operator for random computing, including:
  • a plurality of read-write circuit units wherein the plurality of read-write circuit units output pulse signals with different probabilities when different control signals are input;
  • a plurality of operation units wherein the operation units are configured to perform an XOR gate logic operation on any two of the pulse signals and output an intermediate signal, wherein different operation units process different pulse signals;
  • the data selector is configured to perform summing, averaging and weighting processing on the plurality of intermediate signals mentioned above based on a preset probability, and output a target signal.
  • a random pulse neural network comprising:
  • a plurality of synapse arrays wherein the plurality of synapse arrays store weights with different resistance states, and output a control signal when an initial signal is input to the synapse array;
  • a plurality of read-write circuit units wherein an output terminal of the synapse array is connected to at least an input terminal of a source and drain of a write transistor in one of the read-write circuit units;
  • the read/write circuit unit is configured to output a pulse signal according to the control signal, and the pulse signal is configured to adjust the weight in the synapse array corresponding to the read/write circuit unit.
  • the spin electronic device, circuit unit, computing network, operator and neural network of the present disclosure can realize that the Joule heat generated by the heater when there is or is no current flowing in the full electric field weakens or restores the RKKY effect, so that the net magnetic field polarity of the RKKY effective field and the built-in magnetic field changes, thereby driving the magnetic domain wall to move and emit random pulse signals, thereby simulating the function of human brain neurons, and modulating different write pulse signals by electrical means to make the magnetic domain wall move randomly, so that the random spin electronic device can Able to generate a variety of different random characteristics.
  • FIG1 schematically shows a schematic structural diagram of a random spin electronic device according to an embodiment of the present disclosure
  • FIG2 schematically shows a schematic diagram of an output signal of a random spin electronic device according to an embodiment of the present disclosure
  • FIG3 schematically shows a schematic diagram of current modulation probability characteristics of a random spin electronic device according to an embodiment of the present disclosure
  • FIG4 schematically shows a schematic diagram of magnetic field modulation probability characteristics of a random spin electronic device according to an embodiment of the present disclosure
  • FIG5 schematically shows a physical image diagram of a random spin electronic device according to an embodiment of the present disclosure
  • FIG6 schematically shows a schematic diagram of a read/write circuit unit according to an embodiment of the present disclosure
  • FIG7 schematically shows a schematic diagram of a random computing network according to an embodiment of the present disclosure
  • FIG8 schematically shows a schematic diagram of an operator of random calculation according to an embodiment of the present disclosure.
  • FIG9 schematically shows a schematic diagram of a random pulse neural network according to an embodiment of the present disclosure.
  • FIG1 schematically shows a structural diagram of a random spin electronic device 100 according to an embodiment of the present disclosure.
  • the random spin electronic device 100 comprises:
  • a heater 110 configured to flow a write pulse signal
  • an antiferromagnetic layer 130 formed on a top surface of the isolation layer 120;
  • a ferromagnetic layer 140 is formed on the top surface of the antiferromagnetic layer 130, wherein the antiferromagnetic layer 130 is configured to fix the magnetic moment of the ferromagnetic layer 140 so that the ferromagnetic layer 140 does not generate a magnetic domain wall;
  • RKKY Rivestman-Kittel-Kasuya-Yosida Interaction
  • the magnetic tunnel junction 160 is formed on one side of the top surface of the RKKY coupling layer 150 , wherein when a write pulse signal is input, the magnetic domain wall in the magnetic domain wall free layer 161 of the magnetic tunnel junction 160 moves.
  • the shape of the heater 110 can be specifically set according to actual needs, for example, it can be set to a combination of a strip and two rectangles as shown in the figure.
  • the random spin electronic device can achieve the weakening or recovery of the RKKY effect by the Joule heat generated by the heater with or without current flowing under the full electric field condition, so that the net magnetic field polarity of the RKKY effective field and the built-in magnetic field changes, thereby driving the movement of the magnetic domain wall and emitting random pulse signals, thereby simulating the functions of human brain neurons.
  • different write pulse signals are modulated by electrical and other means to cause the magnetic domain wall to move randomly, so that the random spin electronic device can produce a variety of different random characteristics.
  • the magnetic tunnel junction 160 includes:
  • a magnetic domain wall free layer 161 is formed on the top surface of the RKKY coupling layer 150 , wherein when a control pulse signal is input, the magnetic domain wall in the magnetic domain wall free layer 161 moves;
  • the barrier layer 162 is formed on one side of the top surface of the magnetic domain wall free layer 161;
  • the reference layer 163 is formed on the top surface of the barrier layer 162 , wherein the magnetic domain wall free layer 161 , the ferromagnetic layer 140 and the reference layer 163 all have perpendicular magnetic anisotropy.
  • the magnetic tunnel junction 160 further includes a tunneling layer formed on the magnetic domain wall free layer 161.
  • the magnetic domain wall free layer 161 forms a threshold region in an area opposite to the tunneling layer.
  • the material of the tunneling layer includes at least one of the following: MgO, Al2O3 , etc. , and its thickness is 0.5 to 4 nm.
  • the random spin electronic device 100 further includes:
  • Two antiferromagnetic pinning layers 170 are formed on both sides of the magnetic domain wall free layer 161 , respectively, wherein the barrier layer 162 is located on both sides of the two antiferromagnetic pinning layers 170 .
  • the material of the antiferromagnetic pinning layer 170 includes at least one of the following: IrMn, FeMn, NiMn, CoMn, PtMn, Mn2Au, NiO, MnO or the material used for the magnetic domain wall free layer 161, etc.
  • the materials of the domain wall free layer, the reference layer 163 and the ferromagnetic layer 140 include at least one of the following: Co, CoFeB, Co, Pt, CoFeAl, Co, Pd, CoFe, the thickness of the reference layer 163 is greater than the thickness of the magnetic domain wall free layer 161, and the thickness of the magnetic domain wall free layer 161 and the thickness of the reference layer 163 are both 0.8-2 nm; or the material coercivity of the reference layer 163 is greater than the material coercivity of the magnetic domain wall free layer 161, and the shapes of the reference layer 163 and the tunneling layer may include circular, elliptical, rectangular, etc.; the reference layer is pinned by a pinning layer such as an antiferromagnetic or synthetic antiferromagnetic pinning layer; the material of the barrier layer includes at least one of the following: MgO, Al 2 O 3 .
  • the material of the antiferromagnetic layer 130 includes at least one of the following: IrMn, FeMn, NiMn, CoMn, PtMn, Mn 2 Au, NiO, MnO;
  • the material of the RKKY coupling layer 150 includes at least one of the following: Ru, Ta, W, Pt, Cu, V, Cr, Rh, Nd, Mo, Re;
  • the material of the isolation layer 120 includes at least one of the following: SiO 2 , SiN;
  • the material of the heater 110 includes a derivative of a metal oxide, a nitride, a carbide or a boride, and the heater 110 has a resistivity satisfying a preset resistance threshold and is configured to generate sufficient Joule heat;
  • the Joule heat generated by the heater 110 is configured to modulate the RKKY action strength of the magnetic domain wall free layer 161 , the RKKY coupling layer 150 , and the ferromagnetic layer 140 .
  • the magnetic domain wall in the domain wall free layer is driven to move in a direction away from the barrier layer 162 under the action of the net magnetic field of the RKKY effective field and the built-in magnetic field;
  • the generated Joule heat weakens the RKKY effect, causing the net magnetic field polarity of the RKKY effective field and the built-in magnetic field to change, driving the magnetic domain wall to move toward the barrier layer 162 .
  • the ferromagnetic layer 140, the RKKY coupling layer, and the magnetic domain wall free layer 161 are coupled through the RKKY exchange action of the RKKY coupling layer 150.
  • an antiferromagnetic coupling effect can be achieved.
  • the ferromagnetic and antiferromagnetic coupling changes with the increase in the thickness of the RKKY coupling layer 150 and presents an oscillatory change, and the oscillation period is about 1nm.
  • the RKKY coupling layer 150 is made of tungsten W
  • the thickness of W is less than 0.43nm
  • ferromagnetic coupling is achieved
  • the thickness of W is greater than 0.43nm and less than 0.75nm, antiferromagnetic coupling is achieved.
  • the RKKY coupling layer can make the ferromagnetic layer 140 and the magnetic domain wall free layer 161 produce RKKY exchange effect;
  • the antiferromagnetic pinning layer 170 can be made of antiferromagnetic materials such as IrMn and PtMn;
  • the isolation layer 120 is configured to isolate the electrical connection between the heater 110 and the random spin electronic device 100; the heater 110 has a large resistivity and can generate sufficient Joule heat.
  • the function of a random neuron can be realized by using the random spin electronic device 100 through the following steps:
  • the pinning and injection of the magnetic domain wall are achieved, wherein the pinning areas on both sides of the magnetic domain wall free layer 161 can be achieved by increasing the local thickness or width of the magnetic domain wall free layer 161.
  • the magnetic domain wall free layer 161, the RKKY coupling layer 150, and the ferromagnetic layer 140 form a synthetic antiferromagnetic structure, so that the magnetic domain wall is subjected to a constant bias.
  • the Joule heat generated by the heater 110 competes with the built-in magnetic field to drive the magnetic domain wall to reciprocate, and the movement exhibits randomness due to the natural randomness.
  • the magnetic domain wall is driven to move back and forth randomly.
  • the magnetic domain wall moves to the threshold region, that is, the magnetization direction of the corresponding local free layer below the reference layer 163 is reversed, the magnetic tunnel junction 160 combines with the peripheral circuit to output a spike signal (pulse signal).
  • FIG2 schematically shows a schematic diagram of the output signal of the random spin electronic device 100 according to an embodiment of the present disclosure.
  • the read anomalous Hall resistance R xy of the random spin electronic device 100 is different in size, reflecting that the position of the magnetic domain wall in the magnetic domain wall free layer 161 is different each time it moves, and it has a high degree of randomness.
  • HRKKY is the strength of the RKKY effective field
  • H external is the strength of the built-in magnetic field.
  • FIG. 3 schematically shows a schematic diagram of the current modulation probability characteristic of the random spin electronic device 100 according to an embodiment of the present disclosure.
  • the discharge probability of the random spin electronic device 100 presents a Sigmoid distribution, so the output probability of the random spin electronic device 100 can be adjusted by adjusting the current amplitude.
  • FIG. 4 schematically shows a schematic diagram of the magnetic field modulation probability characteristics of the random spin electronic device 100 according to an embodiment of the present disclosure.
  • the discharge probability of the random spin electronic device 100 under different currents presents different Sigmoid distributions, so the output probability of the random spin electronic device 100 can be adjusted by adjusting the current amplitude, wherein the built-in magnetic field can be achieved by depositing a magnetic free layer, designing interlayer stray fields, external magnetic fields, etc.
  • FIG5 schematically shows a physical image diagram of a random spin electronic device 100 according to an embodiment of the present disclosure. Due to the fact that the thin film deposition and random spin electronic device 100 processing process inevitably introduces spatially distributed pinning points and high and low Potential barriers or potential wells of varying widths will result in a certain pinning time when the domain wall moves. The time for the domain wall to be depinned is random, resulting in a high degree of randomness in the movement of the domain wall to the MTJ threshold, which can produce an output probability distribution as shown in Figure 5.
  • the random spin electronic device 100 by controlling the amplitude, number and pulse width of the pulse signal, the random spin electronic device 100 has at least one characteristic of accumulation-discharge, accumulation-leakage-discharge and random through the formed RKKY effective field and built-in magnetic field.
  • FIG6 schematically shows a schematic diagram of a read/write circuit unit 200 according to an embodiment of the present disclosure.
  • the read/write circuit unit 200 includes:
  • a random spin electronic device 100 wherein a magnetic tunnel junction 160 in the random spin electronic device 100 is configured to input a read current, and an end of the heater 110 away from the magnetic tunnel junction 160 is configured to be connected to a power supply voltage V dd so that current flows into the heater;
  • a write transistor source-drain 210 one end of which is connected to the other end of the heater 110, and an input end of the write transistor source-drain 210 is configured to input a control signal Vin ;
  • a current limiting resistor 220 one end of which is connected to the other end of the source and drain of the write transistor 210, and the other end of the current limiting resistor 220 is grounded;
  • the comparator 230 has a first terminal adjacent to the magnetic tunnel junction 160 , and a second terminal of the comparator 230 is configured to output a pulse signal.
  • FIG6 shows a read/write circuit unit 200 constructed based on a random spin electronic device 100.
  • the input control signal Vin can control the switch state and current magnitude of the write transistor source/drain 210, so that the current pulse flowing through the random spin electronic device 100 is controllable, driving the magnetic domain wall to move, and a small read current Iread flows through the MTJ structure and can output a pulse sequence through the comparator 230.
  • a neuron characteristic of accumulation-leakage-discharge with a certain degree of randomness can be achieved.
  • FIG. 7 schematically shows a schematic diagram of a random computing network 300 according to an embodiment of the present disclosure.
  • a random computing network 300 includes:
  • a plurality of read/write circuit units 200 A plurality of read/write circuit units 200;
  • a digital-to-analog converter 310 which is configured to convert the input digital data into an analog signal and input a control signal to the input terminal of the write transistor source and drain 210 in each read-write circuit unit 200 through the output terminal;
  • the microcontroller or processor 320 is configured to perform partial derivative processing on the multiple pulse signals output by each read-write circuit unit 200 based on a cost function, and output an input signal corresponding to the read-write circuit unit 200, so that the digital-to-analog converter 310 generates a control signal acting on the read-write circuit unit 200 according to the input signal.
  • FIG7 shows the network structure of a random computing network 300 constructed based on a random spin electronic device 100.
  • Each random neuron constructed by a read/write circuit unit 200 is assumed to be an odd number.
  • the partial derivative of the cost function for each random neuron is used to obtain the input drive of each random neuron.
  • Read and write control and calculation can be implemented by a microcontroller or processor 320 such as an FPGA, an MCU or a PC.
  • FIG8 schematically shows a schematic diagram of an operator 400 for random computing according to an embodiment of the present disclosure.
  • the operator 400 of random calculation includes:
  • a plurality of read/write circuit units 200 wherein the plurality of read/write circuit units 200 output pulse signals with different probabilities when different control signals are input;
  • a plurality of operation units 410 wherein the operation units 410 are configured to perform logic operations such as XOR gates on any two pulse signals and output intermediate signals, wherein different operation units 410 process different pulse signals;
  • FIG8 shows a Roberts operator 400 for random computing based on a random spin electronic device 100.
  • Each pixel value of the image is converted into a control signal Vin , and each random neuron generates a pulse sequence with different probabilities according to Vin .
  • the operation unit 410 and the data selector 420 are used to perform the logical operation shown in FIG8 on each pulse sequence, that is, the probability is operated, which conforms to the calculation paradigm of the Roberts operator 400 and can be used for edge detection of the image.
  • more complex Prewitt, Sobel, Canny operators, etc. can also be constructed according to the method of FIG8.
  • FIG9 schematically shows a schematic diagram of a random pulse neural network 500 according to an embodiment of the present disclosure.
  • a random pulse neural network 500 includes:
  • a plurality of synapse arrays 510 wherein the plurality of synapse arrays 510 store weights of different resistance states, and output a control signal when an initial signal is input to the synapse array 510;
  • a plurality of read-write circuit units 200, an output end of a synapse array 510 is connected to an input end of a write transistor source-drain 210 in at least one read-write circuit unit 200;
  • the read/write circuit unit 200 is configured to output a pulse signal according to the control signal, and the pulse signal is configured to adjust the weight in the synapse array 510 corresponding to the read/write circuit unit 200 .
  • FIG9 shows a random pulse neural network 500 constructed based on a random spin electronic device 100.
  • the input initial signal flows through the synaptic array 510, and matrix-vector multiplication is performed according to Kirchhoff's law, and then the current pulse flows through the read-write circuit unit 200 to drive the movement of the magnetic domain wall.
  • the random neuron will release an electrical signal to the next level. Due to the randomness of the magnetic domain wall movement, the discharge of the random neuron also has a certain randomness. The introduced randomness can improve the stability of the random pulse neural network 500.

Landscapes

  • Hall/Mr Elements (AREA)

Abstract

本公开提供了一种自旋电子器件、电路单元、计算网络、算子以及神经网络。该随机自旋电子器件包括：加热器，加热器被配置成写入控制脉冲信号；隔离层，形成于加热器的顶面；反铁磁层，形成于隔离层的顶面；铁磁层，形成于反铁磁层的顶面，其中，反铁磁层被配置成固定铁磁层的磁矩，以使得铁磁层不产生磁畴壁；RKKY耦合层，形成于铁磁层的顶面；磁隧道结，形成于RKKY耦合层的顶面一侧，其中，在写入脉冲信号输入的情况下，磁隧道结的磁畴壁自由层中的磁畴壁发生运动。

Description

自旋电子器件、电路单元、计算网络、算子以及神经网络 技术领域
本公开涉及磁畴壁技术领域，更具体地，涉及一种随机自旋电子器件、读写电路单元、随机计算网络、随机计算的算子以及随机脉冲神经网络。
背景技术
基于自旋电子器件的随机计算、神经形态计算是在芯片高密度集成、计算密集型背景下，为边缘计算提供了低功耗、高性能的解决方案。同时，可以在不损失计算精度的前提下，显著提升计算的稳定性与抗噪声特性，为边缘计算提供了更可靠的方案。
随机计算的核心的可以产生随机脉冲序列的随机器件，自旋电子器件具有极高的耐久性和稳定性，可以在高工作强度下产生固定概率的随机脉冲序列，应用潜力巨大。然而，目前基于自旋电子的随机器件主要是基于低势垒的磁隧道结(magnetic tunnel junction，MTJ)器件，在热噪声的扰动下概率性的翻转，但概率的调制通常需要STT电流、外部磁场等多种方式实现，对器件寿命及集成应用不利，同时没有考虑磁畴壁(DW)的本征随机性。
发明内容
有鉴于此，本公开实施例提供了一种随机自旋电子器件、读写电路单元、随机计算网络、随机计算的算子以及随机脉冲神经网络。
本公开实施例的一个方面提供了一种随机自旋电子器件，包括：
加热器，上述加热器被配置成流过写入脉冲信号；
隔离层，形成于上述加热器的顶面；
反铁磁层，形成于上述隔离层的顶面；
铁磁层，形成于上述反铁磁层的顶面，其中，上述反铁磁层被配置成固定上述铁磁层的磁矩，以使得上述铁磁层不产生磁畴壁；
RKKY耦合层，形成于上述铁磁层的顶面；
磁隧道结，形成于上述RKKY耦合层的顶面一侧，其中，在写入脉冲信号输入的情况下，上述磁隧道结的磁畴壁自由层中的磁畴壁发生运动。
根据本公开的实施例，上述磁隧道结包括：
磁畴壁自由层，形成于上述RKKY耦合层的顶面，其中，在输入控制脉冲信号的情况下，上述磁畴壁自由层内的上述磁畴壁发生运动；
势垒层，形成于上述磁畴壁自由层的顶面一侧；
参考层，形成于上述势垒层的顶面，其中，上述磁畴壁自由层、上述铁磁层和上述参考层均具有垂直磁各向异性。
根据本公开的实施例，随机自旋电子器件还包括：
两个反铁磁钉扎层，分别形成于上述磁畴壁自由层的两侧，其中，上述势垒层位于两个上述反铁磁钉扎层的两侧。
根据本公开的实施例，上述畴壁自由层、上述参考层和上述铁磁层的材料包括以下至少之一：Co、CoFeB、Co、Pt、CoFeAl、Co、Pd、CoFe，上述参考层的厚度大于上述磁畴壁自由层的厚度，或上述参考层的材料矫顽力大于上述磁畴壁自由层的材料矫顽力，或，上述参考层被反铁磁或合成反铁磁的钉扎层所钉扎；上述势垒层的材料包括以下至少之一：MgO、Al2O3；
上述反铁磁层的材料包括以下至少之一：IrMn、FeMn、NiMn、CoMn、PtMn、Mn2Au、NiO、MnO；
上述RKKY耦合层的材料包括以下至少之一：Ru、Ta、W、Pt、Cu、V、Cr、Rh、Nd、Mo、Re；
上述隔离层的材料包括以下至少之一：SiO2、SiN；
上述加热器的材料包括金属的氧化物、氮化物、碳化物或硼化物的衍生物，上述加热器具有满足预设电阻阈值的电阻率，被构造成产生充足的焦耳热；
其中，上述加热器产生的焦耳热用于被配置成调制磁畴壁自由层、RKKY耦合层和上述铁磁层的RKKY作用强度。
根据本公开的实施例，在没有电流流过随机自旋电子器件的情况下，上述畴壁自由层中的磁畴壁在RKKY有效场和内建磁场的净磁场作用下，驱动磁畴壁向远离上述势垒层的方向运动；
当电流流过加热器时，产生的焦耳热削弱RKKY作用，使RKKY有效场和内建磁场的净磁场极性改变，驱动磁畴壁向上述势垒层的方向运动。
根据本公开的实施例，根据控制脉冲信号的幅值、数量、脉宽，上述随机自旋电子器件通过形成的RKKY有效场和内建磁场而具备积累-放电、积累-泄露-放电、随机中的至少一种特性。
本公开实施例的另一个方面提供了一种读写电路单元，包括：
随机自旋电子器件，其中，上述随机自旋电子器件中的磁隧道结被配置成输入读电流，加热器远离上述磁隧道结的一端接入电源电压，使电流流入加热器；
写晶体管源漏极，一端与上述加热器的另一端连接，上述写晶体管源漏极的输入端被配置成输入控制信号；
限流电阻，一端与上述写晶体管源漏极的另一端连接；
比较器，上述比较器的第一端与上述磁隧道结邻接，上述比较器的第二端被配置成输出脉冲信号。
本公开实施例的另一个方面提供了一种随机计算网络，包括：
多个读写电路单元；
数模转换器，上述数模转换器被配置成对输入的输入数字信号转换为模拟信号，并向每个上述读写电路单元中的写晶体管源漏极的输入端输入控制信号；
微控制器或处理器，上述微控制器或处理器被配置成基于成本函数对每个上述读写电路单元输出的多个脉冲信号进行偏导处理，输出对应于上述读写电路单元的上述输入信号，以使得上述数模转换器根据上述输入信号生成作用于上述读写电路单元的控制信号。
本公开实施例的另一个方面提供了一种随机计算的算子，包括：
多个读写电路单元，其中，多个上述读写电路单元在输入不同的控制信号的情况，输出不同概率的脉冲信号；
多个运算单元，上述运算单元被配置成对任意两个上述脉冲信号进行异或门的逻辑运算，输出中间信号，其中，不同上述运算单元处理不同的脉冲信号；
数据选择器，被配置成基于预设概率对多个上述中间信号进行求和平均及加权处理，输出目标信号。
本公开实施例的另一个方面提供了一种随机脉冲神经网络，包括：
多个突触阵列，其中，多个上述突触阵列中存储有阻态不同的权重，在上述突触阵列输入初始信号的情况下，输出控制信号；
多个读写电路单元，一个上述突触阵列的输出端至少与一个上述读写电路单元中的写晶体管源漏极的输入端链接；
其中，上述读写电路单元被配置成根据上述控制信号，输出脉冲信号，上述脉冲信号被配置成调整与上述读写电路单元对应的上述突触阵列中的权重。
根据本公开的实施例，本公开的自旋电子器件、电路单元、计算网络、算子以及神经网络，由于随机自旋电子器件可以实现在全电场情况下，加热器在有无电流流过时产生的焦耳热对RKKY作用的削弱或恢复，使得RKKY有效场和内建磁场的净磁场极性发生改变，从而驱动磁畴壁运动并放出随机脉冲信号，进而能够模拟人脑神经元的功能，同时通过电学等方式调制不同的写入脉冲信号使得磁畴壁发生随机性的运动，从而使得随机自旋电子器件能 够产生多种不同的随机特性。
附图说明
通过以下参照附图对本公开实施例的描述，本公开的上述以及其他目的、特征和优点将更为清楚，在附图中：
图1示意性示出了根据本公开实施例的随机自旋电子器件的结构示意图；
图2示意性示出了根据本公开实施例的随机自旋电子器件的输出信号示意图；
图3示意性示出了根据本公开实施例的随机自旋电子器件的电流调制概率特性示意图；
图4示意性示出了根据本公开实施例的随机自旋电子器件的磁场调制概率特性示意图；
图5示意性示出了根据本公开实施例的随机自旋电子器件的物理图像示意图；
图6示意性示出了根据本公开实施例的读写电路单元的示意图；
图7示意性示出了根据本公开实施例的随机计算网络的示意图；
图8示意性示出了根据本公开实施例的随机计算的算子的示意图；以及
图9示意性示出了根据本公开实施例的随机脉冲神经网络的示意图。
具体实施方式
以下，将参照附图来描述本公开的实施例。但是应该理解，这些描述只是示例性的，而并非要限制本公开的范围。在下面的详细描述中，为便于解释，阐述了许多具体的细节以提供对本公开实施例的全面理解。然而，明显地，一个或多个实施例在没有这些具体细节的情况下也可以被实施。此外，在以下说明中，省略了对公知结构和技术的描述，以避免不必要地混淆本公开的概念。
在此使用的术语仅仅是为了描述具体实施例，而并非意在限制本公开。在此使用的术语“包括”、“包含”等表明了所述特征、步骤、操作和/或部件的存在，但是并不排除存在或添加一个或多个其他特征、步骤、操作或部件。
在此使用的所有术语(包括技术和科学术语)具有本领域技术人员通常所理解的含义，除非另外定义。应注意，这里使用的术语应解释为具有与本说明书的上下文相一致的含义，而不应以理想化或过于刻板的方式来解释。
在使用类似于“A、B和C等中至少一个”这样的表述的情况下，一般来说应该按照本领域技术人员通常理解该表述的含义来予以解释(例如，“具有A、B和C中至少一个的系统”应包括但不限于单独具有A、单独具有B、单独具有C、具有A和B、具有A和C、具有B和C、和/或具有A、B、C的系统等)。
图1示意性示出了根据本公开实施例的随机自旋电子器件100的结构示意图。
如图1所示，随机自旋电子器件100包括：
加热器110，加热器110被配置成流过写入脉冲信号；
隔离层120，形成于加热器110的顶面；
反铁磁层130，形成于隔离层120的顶面；
铁磁层140，形成于反铁磁层130的顶面，其中，反铁磁层130被配置成固定铁磁层140的磁矩，以使得铁磁层140不产生磁畴壁；
RKKY(Ruderman-Kittel-Kasuya-Yosida Interaction)耦合层150，形成于铁磁层140的顶面；
磁隧道结160，形成于RKKY耦合层150的顶面一侧，其中，在写入脉冲信号输入的情况下，磁隧道结160的磁畴壁自由层161中的磁畴壁发生运动。
根据本公开的实施例，加热器110的形状可以根据实际需求具体设置，例如可以设置为如图所示的条形与两个矩形的组合形状。
根据本公开的实施例，随机自旋电子器件可以实现在全电场情况下，加热器在有无电流流过时产生的焦耳热对RKKY作用的削弱或恢复，使得RKKY有效场和内建磁场的净磁场极性发生改变，从而驱动磁畴壁运动并放出随机脉冲信号，进而能够模拟人脑神经元的功能，同时通过电学等方式调制不同的写入脉冲信号使得磁畴壁发生随机性的运动，从而使得随机自旋电子器件能够产生多种不同的随机特性。
根据本公开的实施例，磁隧道结160包括：
磁畴壁自由层161，形成于RKKY耦合层150的顶面，其中，在输入控制脉冲信号的情况下，磁畴壁自由层161内的磁畴壁发生运动；
势垒层162，形成于磁畴壁自由层161的顶面一侧；
参考层163，形成于势垒层162的顶面，其中，磁畴壁自由层161、铁磁层140和参考层163均具有垂直磁各向异性。
根据本公开的实施例，磁隧道结160还包括形成于磁畴壁自由层161上的隧穿层，磁畴壁自由层161正对隧穿层的区域形成阈值区域，隧穿层的材料包括以下至少之一：MgO、Al2O3等，其厚度为0.5～4nm。
根据本公开的实施例，随机自旋电子器件100还包括：
两个反铁磁钉扎层170，分别形成于磁畴壁自由层161的两侧，其中，势垒层162位于两个反铁磁钉扎层170的两侧。
根据本公开的实施例，反铁磁钉扎层170的材料包括以下至少之一：IrMn、FeMn、NiMn、 CoMn、PtMn、Mn2Au、NiO、MnO或磁畴壁自由层161所应用的材料等。
根据本公开的实施例，畴壁自由层、参考层163和铁磁层140的材料包括以下至少之一：Co、CoFeB、Co、Pt、CoFeAl、Co、Pd、CoFe，参考层163的厚度大于磁畴壁自由层161的厚度，且磁畴壁自由层161的厚度参考层163的厚度均为0.8～2nm；或参考层163的材料矫顽力大于磁畴壁自由层161的材料矫顽力，参考层163及隧穿层的形状可以包括圆形、椭圆形、矩形等；所述参考层被反铁磁或合成反铁磁等钉扎层所钉扎；所述势垒层的材料包括以下至少之一：MgO、Al2O3。
反铁磁层130的材料包括以下至少之一：IrMn、FeMn、NiMn、CoMn、PtMn、Mn2Au、NiO、MnO；
RKKY耦合层150的材料包括以下至少之一：Ru、Ta、W、Pt、Cu、V、Cr、Rh、Nd、Mo、Re；
隔离层120的材料包括以下至少之一：SiO2、SiN；
加热器110的材料包括金属的氧化物、氮化物、碳化物或硼化物的衍生物，加热器110具有满足预设电阻阈值的电阻率，被构造成产生充足的焦耳热；
其中，加热器110产生的焦耳热被配置成调制磁畴壁自由层161、RKKY耦合层150和铁磁层140的RKKY作用强度。
根据本公开的实施例，在没有电流流过随机自旋电子器件100的情况下，畴壁自由层中的磁畴壁在RKKY有效场和内建磁场的净磁场作用下，驱动磁畴壁向远离势垒层162的方向运动；
当电流流过加热器110时，产生的焦耳热削弱RKKY作用，使RKKY有效场和内建磁场的净磁场极性改变，驱动磁畴壁向势垒层162的方向运动。
根据本公开的实施例，根据本公开的实施例，铁磁层140、RKKY耦合层、磁畴壁自由层161通过RKKY耦合层150的RKKY交换作用耦合起来，具体的根据RKKY耦合层150的厚度的变化，可以实现反铁磁耦合作用，通常铁磁和反铁磁耦合随着RKKY耦合层150厚度增加而且呈现振荡性的变化，振荡周期约为1nm，例如RKKY耦合层150采用钨W制成的情况下，W的厚度小于0.43nm时，实现铁磁耦合，当W的厚度大于0.43nm，小于0.75nm时，实现反铁磁耦合。
本实施例以反铁磁耦合作用为例。RKKY耦合层可以使铁磁层140和磁畴壁自由层161产生RKKY交换作用；反铁磁钉扎层170可以采用IrMn、PtMn等反铁磁材料构成；隔离层120被配置成隔离加热器110和随机自旋电子器件100的电学连接；加热器110具有较大的电阻率，可以产生充足的焦耳热。当没有电流流过加热器110时，磁畴壁自由层161中的磁 畴壁在RKKY有效场和内建场的净有效场的驱动下，向左运动；当电流流过加热器110时，RKKY作用被削弱，导致RKKY有效场和内建场的净有效场极性发生改变，驱动磁畴壁向右运动；撤掉电流后，焦耳热会耗散，RKKY作用恢复，驱动磁畴壁向左运动；经过上述操作，磁畴壁会运动到磁隧道结160(magnetic tunnel junction,MTJ)阈值处，随机自旋电子器件100输出一个脉冲信号。
根据本公开的实施例，利用随机自旋电子器件100通过如下步骤可实现随机神经元的功能：
首先，通过在磁畴壁自由层161两侧沉积磁化方向相反的反铁磁体或保留/沉积较厚的局部自由层，实现磁畴壁的钉扎与注入，其中，磁畴壁自由层161两侧的钉扎区域可以通过增加磁畴壁自由层161局部厚度或宽度等方式实现。
第二，通过磁畴壁自由层161、RKKY耦合层150、铁磁层140三者构成合成反铁磁结构，使磁畴壁受到恒定的偏置作用。
第三，利用加热器110产生的焦耳热与内建的磁场相互竞争，驱动磁畴壁往复运动，由于天然的随机性，运动表现出随机性。
最后，根据输入的控制脉冲信号的幅值、脉宽及数目，驱动磁畴壁随机往复运动，当磁畴壁运动到阈值区域，即参考层163下方对应局部自由层的磁化方向发生反转时，磁隧道结160结合外围电路输出一个尖峰信号(脉冲信号)。
图2示意性示出了根据本公开实施例的随机自旋电子器件100的输出信号示意图。在100个1s脉宽2.73mA的电脉冲情况下，由图2可知，读取到的随机自旋电子器件100的反常霍尔电阻Rxy大小不一，反映出磁畴壁自由层161中的磁畴壁每次运动的位置不同，具有高度的随机性。其中，HRKKY为RKKY有效场的强度，Hexternal为内建磁场的强度。
图3示意性示出了根据本公开实施例的随机自旋电子器件100的电流调制概率特性示意图。随着电流脉冲幅值在2.7mA到2.8mA范围变化时，随机自旋电子器件100的放电概率呈现Sigmoid型分布，因此可以通过调整电流幅值，调整随机自旋电子器件100的输出概率。
图4示意性示出了根据本公开实施例的随机自旋电子器件100的磁场调制概率特性示意图。随着RKKY有效场和内建磁场的净有效场在-75Oe到-60Oe范围变化时，不同电流下的随机自旋电子器件100的放电概率均呈现不同的Sigmoid型分布，因此可以通过调整电流幅值，调整随机自旋电子器件100的输出概率，其中，内建磁场可以通过沉积磁性自由层、设计层间杂散场、外部磁场等方式实现。
图5示意性示出了根据本公开实施例的随机自旋电子器件100的物理图像示意图。由于薄膜沉积及随机自旋电子器件100加工过程中不可避免的会引入空间分布的钉扎点，及高低 宽窄不一的势垒或势阱，在磁畴壁运动时会导致一定的钉扎时间，磁畴壁脱钉扎的时间具有随机性，导致磁畴壁运动到MTJ阈值处具有高度随机性，从而可以产生如图5的输出概率分布。
根据本公开的实施例，根据控制脉冲信号的幅值、数量、脉宽，随机自旋电子器件100通过形成的RKKY有效场和内建磁场而具备积累-放电、积累-泄露-放电、随机中的至少一种特性。
图6示意性示出了根据本公开实施例的读写电路单元200的示意图。
根据本公开的实施例，如图6所示，读写电路单元200包括：
随机自旋电子器件100，其中，随机自旋电子器件100中的磁隧道结160被配置成输入读电流，加热器110远离磁隧道结160的一端被配置成接入电源电压Vdd，使电流流入加热器；
写晶体管源漏极210，一端与加热器110的另一端连接，写晶体管源漏极210的输入端被配置成输入控制信号Vin；
限流电阻220，一端与写晶体管源漏极210的另一端连接，限流电阻220的另一端接地；
比较器230，比较器230的第一端与磁隧道结160邻接，比较器230的第二端被配置成输出脉冲信号。
根据本公开的实施例，图6所示为基于随机自旋电子器件100构建的读写电路单元200。输入控制信号Vin可以控制写晶体管源漏极210的开关状态及电流大小，从而使流过随机自旋电子器件100的电流脉冲可控，驱动磁畴壁运动，一个小的读取电流Iread流过MTJ结构通过比较器230可以输出脉冲序列。除此之外，通过调控Vin的幅值、脉宽，可以实现具有一定随机性的积累-泄露-放电的神经元特性。
图7示意性示出了根据本公开实施例的随机计算网络300的示意图。
根据本公开的实施例，如图7所示，随机计算网络300包括：
多个读写电路单元200；
数模转换器310，数模转换器310被配置成对输入的数字转换为模拟信号，并通过输出端向每个读写电路单元200中的写晶体管源漏极210的输入端输入控制信号；
微控制器或处理器320，微控制器或处理器320被配置成基于成本函数对每个读写电路单元200输出的多个脉冲信号进行偏导处理，输出对应于读写电路单元200的输入信号，以使得数模转换器310根据输入信号生成作用于读写电路单元200的控制信号。
根据本公开的实施例，图7所示为基于随机自旋电子器件100构建的随机计算网络300的网络结构。每个由读写电路单元200构建的随机神经元被假设成一个奇数，为使成本函数最小化，用成本函数对每个随机神经元的偏导获得每个随机神经元的输入驱动，输入输出的 读写控制与计算可通过FPGA、MCU或PC等微控制器或处理器320实现。
图8示意性示出了根据本公开实施例的随机计算的算子400的示意图。
根据本公开的实施例，如图8所示，随机计算的算子400包括：
多个读写电路单元200，其中，多个读写电路单元200在输入不同的控制信号的情况，输出不同概率的脉冲信号；
多个运算单元410，运算单元410被配置成对任意两个脉冲信号进行异或门等逻辑运算，输出中间信号，其中，不同运算单元410处理不同的脉冲信号；
数据选择器420，被配置成基于预设概率(例如图8所示的P＝0.5)对多个中间信号进行求和平均、加权等处理，输出目标信号。
根据本公开的实施例，图8所示为基于随机自旋电子器件100构建的用于随机计算的Roberts算子400。图片的每个像素值转化为控制信号Vin，每个随机神经元根据Vin产生不同概率的脉冲序列，利用运算单元410和数据选择器420对每段脉冲序列进行图8所示的逻辑运算，即对概率进行运算，符合Roberts算子400的计算范式，可以用来进行图片的边缘检测。除此之外，还可以根据图8的方式构建更复杂的Prewitt、Sobel、Canny算子等。
图9示意性示出了根据本公开实施例的随机脉冲神经网络500的示意图。
根据本公开的实施例，如图9所示，随机脉冲神经网络500包括：
多个突触阵列510，其中，多个突触阵列510中存储有阻态不同的权重，在突触阵列510输入初始信号的情况下，输出控制信号；
多个读写电路单元200，一个突触阵列510的输出端至少与一个读写电路单元200中的写晶体管源漏极210的输入端链接；
其中，读写电路单元200被配置成根据控制信号，输出脉冲信号，脉冲信号被配置成调整与读写电路单元200对应的突触阵列510中的权重。
根据本公开的实施例，图9所示为基于随机自旋电子器件100构建的随机脉冲神经网络500。输入的初始信号流过突触阵列510，根据基尔霍夫定律进行矩阵-向量乘法，进而电流脉冲流过读写电路单元200以驱动磁畴壁运动，当畴壁运动到MTJ阈值处，随机神经元会向下一级放出一个电信号，由于磁畴壁运动的随机性，随机神经元放电也具有一定随机性，引入的随机性可以提升随机脉冲神经网络500的稳定性。
以上对本公开的实施例进行了描述。但是，这些实施例仅仅是为了说明的目的，而并非为了限制本公开的范围。尽管在以上分别描述了各实施例，但是这并不意味着各个实施例中的措施不能有利地结合使用。本公开的范围由所附权利要求及其等同物限定。不脱离本公开的范围，本领域技术人员可以做出多种替代和修改，这些替代和修改都应落在本公开的范围 之内。

Claims (10)

  1. 一种随机自旋电子器件，包括：
    加热器，所述加热器被配置成流过写入脉冲信号；
    隔离层，形成于所述加热器的顶面；
    反铁磁层，形成于所述隔离层的顶面；
    铁磁层，形成于所述反铁磁层的顶面，其中，所述反铁磁层被配置成固定所述铁磁层的磁矩，以使得所述铁磁层不产生磁畴壁；
    RKKY耦合层，形成于所述铁磁层的顶面；
    磁隧道结，形成于所述RKKY耦合层的顶面一侧，其中，在写入脉冲信号输入的情况下，所述磁隧道结的磁畴壁自由层中的磁畴壁发生运动。
  2. 根据权利要求1所述的随机自旋电子器件，其中，所述磁隧道结包括：
    所述磁畴壁自由层，形成于所述RKKY耦合层的顶面，其中，在输入控制脉冲信号的情况下，所述磁畴壁自由层内的所述磁畴壁发生运动；
    势垒层，形成于所述磁畴壁自由层的顶面一侧；
    参考层，形成于所述势垒层的顶面，其中，所述磁畴壁自由层、所述铁磁层和所述参考层均具有垂直磁各向异性。
  3. 根据权利要求2所述的随机自旋电子器件，还包括：
    两个反铁磁钉扎层，分别形成于所述磁畴壁自由层的两侧，其中，所述势垒层位于两个所述反铁磁钉扎层的两侧。
  4. 根据权利要求2所述的随机自旋电子器件，其中，所述畴壁自由层、所述参考层和所述铁磁层的材料包括以下至少之一：Co、CoFeB、Co、Pt、CoFeAl、Co、Pd、CoFe，所述参考层的厚度大于所述磁畴壁自由层的厚度，或所述参考层的材料矫顽力大于所述磁畴壁自由层的材料矫顽力，或，所述参考层被反铁磁或合成反铁磁的钉扎层所钉扎；所述势垒层的材料包括以下至少之一：MgO、Al2O3；
    所述反铁磁层的材料包括以下至少之一：IrMn、FeMn、NiMn、CoMn、PtMn、Mn2Au、NiO、MnO；
    所述RKKY耦合层的材料包括以下至少之一：Ru、Ta、W、Pt、Cu、V、Cr、Rh、Nd、Mo、Re；
    所述隔离层的材料包括以下至少之一：SiO2、SiN；
    所述加热器的材料包括金属的氧化物、氮化物、碳化物或硼化物的衍生物，所述加热器 具有满足预设电阻阈值的电阻率，被构造成产生充足的焦耳热；
    其中，所述加热器产生的焦耳热被配置成调制磁畴壁自由层、RKKY耦合层和所述铁磁层的RKKY作用强度。
  5. 根据权利要求2所述的随机自旋电子器件，其中，在没有电流流过随机自旋电子器件的情况下，所述畴壁自由层中的磁畴壁在RKKY有效场和内建磁场的净磁场作用下，驱动磁畴壁向远离所述势垒层的方向运动；
    当电流流过加热器时，产生的焦耳热削弱RKKY作用，使RKKY有效场和内建磁场的净磁场极性改变，驱动磁畴壁向所述势垒层的方向运动。
  6. 根据权利要求1所述的随机自旋电子器件，其中，根据控制脉冲信号的幅值、数量、脉宽，所述随机自旋电子器件通过形成的RKKY有效场和内建磁场而具备积累-放电、积累-泄露-放电、随机中的至少一种特性。
  7. 一种读写电路单元，包括：
    如权利要求1～6中任一项所述的随机自旋电子器件，其中，所述随机自旋电子器件中的磁隧道结被配置成输入读电流，加热器远离所述磁隧道结的一端接入电源电压，使电流流入加热器；
    写晶体管源漏极，一端与所述加热器的另一端连接，所述写晶体管源漏极的输入端被配置成输入控制信号；
    限流电阻，一端与所述写晶体管源漏极的另一端连接；
    比较器，所述比较器的第一端与所述磁隧道结邻接，所述比较器的第二端被配置成输出脉冲信号。
  8. 一种随机计算网络，包括：
    多个如权利要求7所述的读写电路单元；
    数模转换器，所述数模转换器被配置成对输入的输入数字信号转换为模拟信号，并向每个所述读写电路单元中的写晶体管源漏极的输入端输入控制信号；
    微控制器或处理器，所述微控制器或处理器被配置成基于成本函数对每个所述读写电路单元输出的多个脉冲信号进行偏导处理，输出对应于所述读写电路单元的所述输入信号，以使得所述数模转换器根据所述输入信号生成作用于所述读写电路单元的控制信号。
  9. 一种随机计算的算子，包括：
    多个如权利要求7所述的读写电路单元，其中，多个所述读写电路单元在输入不同的控制信号的情况，输出不同概率的脉冲信号；
    多个运算单元，所述运算单元被配置成对任意两个所述脉冲信号进行异或门的逻辑运算， 输出中间信号，其中，不同所述运算单元处理不同的脉冲信号；
    数据选择器，被配置成基于预设概率对多个所述中间信号进行求和平均及加权处理，输出目标信号。
  10. 一种随机脉冲神经网络，包括：
    多个突触阵列，其中，多个所述突触阵列中存储有阻态不同的权重，在所述突触阵列输入初始信号的情况下，输出控制信号；
    多个如权利要求7所述的读写电路单元，一个所述突触阵列的输出端至少与一个所述读写电路单元中的写晶体管源漏极的输入端链接；
    其中，所述读写电路单元被配置成根据所述控制信号，输出脉冲信号，所述脉冲信号被配置成调整与所述读写电路单元对应的所述突触阵列中的权重。
PCT/CN2023/085355 2023-03-31 2023-03-31 自旋电子器件、电路单元、计算网络、算子以及神经网络 WO2024197775A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/CN2023/085355 WO2024197775A1 (zh) 2023-03-31 2023-03-31 自旋电子器件、电路单元、计算网络、算子以及神经网络

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2023/085355 WO2024197775A1 (zh) 2023-03-31 2023-03-31 自旋电子器件、电路单元、计算网络、算子以及神经网络

Publications (1)

Publication Number Publication Date
WO2024197775A1 true WO2024197775A1 (zh) 2024-10-03

Family

ID=92903024

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2023/085355 WO2024197775A1 (zh) 2023-03-31 2023-03-31 自旋电子器件、电路单元、计算网络、算子以及神经网络

Country Status (1)

Country Link
WO (1) WO2024197775A1 (zh)

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101465407A (zh) * 2007-12-21 2009-06-24 原子能委员会 热辅助磁写入存储器
CN108780779A (zh) * 2016-06-10 2018-11-09 Tdk株式会社 交换偏置利用型磁化反转元件、交换偏置利用型磁阻效应元件、交换偏置利用型磁存储器、非易失性逻辑电路及磁神经元元件
CN113168858A (zh) * 2019-06-24 2021-07-23 西部数据技术公司 热辅助垂直自旋转移矩mram存储器单元
WO2022062427A1 (zh) * 2021-05-17 2022-03-31 中国科学院微电子研究所 全电控自旋电子神经元器件、神经元电路和神经网络
WO2022083193A1 (zh) * 2021-07-21 2022-04-28 中国科学院微电子研究所 一种基于自旋轨道矩的神经元器件
CN114649468A (zh) * 2022-03-03 2022-06-21 中国科学院微电子研究所 基于离子门调控的可重构神经元器件及其制备方法
CN115188881A (zh) * 2022-04-12 2022-10-14 中国科学院微电子研究所 一种基于磁隧道结的神经元器件及神经网络装置

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101465407A (zh) * 2007-12-21 2009-06-24 原子能委员会 热辅助磁写入存储器
CN108780779A (zh) * 2016-06-10 2018-11-09 Tdk株式会社 交换偏置利用型磁化反转元件、交换偏置利用型磁阻效应元件、交换偏置利用型磁存储器、非易失性逻辑电路及磁神经元元件
CN113168858A (zh) * 2019-06-24 2021-07-23 西部数据技术公司 热辅助垂直自旋转移矩mram存储器单元
WO2022062427A1 (zh) * 2021-05-17 2022-03-31 中国科学院微电子研究所 全电控自旋电子神经元器件、神经元电路和神经网络
WO2022083193A1 (zh) * 2021-07-21 2022-04-28 中国科学院微电子研究所 一种基于自旋轨道矩的神经元器件
CN114649468A (zh) * 2022-03-03 2022-06-21 中国科学院微电子研究所 基于离子门调控的可重构神经元器件及其制备方法
CN115188881A (zh) * 2022-04-12 2022-10-14 中国科学院微电子研究所 一种基于磁隧道结的神经元器件及神经网络装置

Similar Documents

Publication Publication Date Title
Siddiqui et al. Magnetic domain wall based synaptic and activation function generator for neuromorphic accelerators
JP4625936B2 (ja) 乱数発生器
CN109952662B (zh) 自旋器件及其操作方法和制造方法
US9864950B2 (en) Compact implementation of neuron and synapse with spin switches
Zand et al. Low-energy deep belief networks using intrinsic sigmoidal spintronic-based probabilistic neurons
DE69421106T2 (de) Neuronales Netzwerk mit räumlich verteilter Arbeitsweise
CN101803192B (zh) 可重新配置的磁逻辑电路结构及用于制造和操作这种逻辑设备的方法
US8203871B2 (en) Reconfigurable magnetic logic device using spin torque
JP4583443B2 (ja) スピントロニクス応用のための磁気電気電界効果トランジスタ
US12100437B2 (en) Magnetic tunnel junction comprising an inhomogeneous granular free layer and associated spintronic devices
Ismael Salinas et al. Spintronic materials and devices towards an artificial neural network: accomplishments and the last mile
Sengupta et al. Spintronic devices for ultra-low power neuromorphic computation (Special session paper)
Roy et al. Beyond charge-based computation: Boolean and non-Boolean computing with spin torque devices
US8704547B2 (en) Method and system for providing spin transfer based logic devices
CN112864314B (zh) 磁电阻器件以及改变其阻态的方法、突触学习模块
WO2024197775A1 (zh) 自旋电子器件、电路单元、计算网络、算子以及神经网络
CN116322276A (zh) 自旋电子器件、电路单元、计算网络、算子以及神经网络
US20230394291A1 (en) Neuron device based on spin orbit torque
WO2022062427A1 (zh) 全电控自旋电子神经元器件、神经元电路和神经网络
CN113326928B (zh) 全电控自旋电子神经元器件、神经元电路和神经网络
Roy et al. Computing with spin-transfer-torque devices: Prospects and perspectives
Roy et al. Exploring Boolean and non-Boolean computing with spin torque devices
Bhuyan A modern review of the spintronic technology: fundamentals, materials, devices, circuits, Challenges, and Current Research Trends
EP2688072B1 (en) Spintronic circuit and method of operation therefore
EP4524965A1 (en) Tunable probabilistic mram elements

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: 23929367

Country of ref document: EP

Kind code of ref document: A1