France



Recent research on deep learning shows that Binary Neural Network (BNN) is able to significantly reduce computation and memory cost with small precision loss [1][2]. By binarizing inputs and weights to +1 or -1 in BNNs, multiply-and-accumulate (MAC) operations can be replaced by XNOR operation in memory. Nonvolatile memory for BNN such as resistive random access memory (RRAM) and magnetic random access memory (MRAM) was reported in [3]-[6]. In this paper, we use spin-toque transfer magnetic random access memory (STT-MRAM) to complete XNOR and single-step convolution in memory. The main contributions include: (1) A STT-MRAM based structure with four transistors and two magnetic tunnel junctions (MTJ) (4T-2M) is proposed for binary convolution and fully connection operations using current mode sense amplifier. (2) The impact of different tunnel magnetoresistance ratio (TMR) and different size of MTJ on accumulating in memory has been explored. (3) By quantizing sensing yield into model, the trade-off between energy and performance is investigated.&lt;br/&gt;As shown in Fig. 1 (a), a 4T-2M bit-cell structure represents one synaptic weight, ‘-1’ is represented by two MTJs as parallel (P) and anti-parallel state (AP). The opposite combination is used to represent ‘1’. Two encoded word lines (WLs) represent one input value. To complete the XNOR operation, an offset-canceling current-sampling sense amplifier (OCCS SA) [7] is used (see Fig. 1(b)) to compare the current difference between source line (SL) and SLB. Fig.1 (c) shows the structure of the proposed STT-MRAM block. For convolution operation, the row decoder gated by appropriate timing decodes the multiple WLs to input parallel, and output through path &amp;lt;1&amp;gt; as the result of convolution. For fully connected operation, the row decoder activates the WLs row by row, and output through path &amp;lt;2&amp;gt; with bit-counting as the result of one fully connected. We implement 9 bits XNOR and accumulation to perform 3 × 3 convolution. Fig. 1(d) lists all the combinations in the convolution. The worst case is four ‘-1’ and five ‘1’ accumulation or its opposite combination, which causes the minimum current difference between “5I&lt;sub&gt;AP&lt;/sub&gt;+4I&lt;sub&gt;P&lt;/sub&gt;” and “5I&lt;sub&gt;P&lt;/sub&gt;+4I&lt;sub&gt;AP&lt;/sub&gt;”. Fig. 1(e) illustrates the circuit realization of BNN. When one group of WLs is simultaneously activated, I&lt;sub&gt;SL&lt;/sub&gt; and I&lt;sub&gt;SLB&lt;/sub&gt; will be different depending on the resistance of MTJs on two SLs. A CSA is implemented to generate high level when weighted sum is positive, and low level when weighted sum is negative. The output of the CSA represents the output of the neuron.&lt;br/&gt;SPICE-level simulation is performed using a Verilog-A model of STT MTJ and an industrial 28-nm CMOS process. Fig.2 (a) shows the simulation results that the sensing yield in different TMR and MTJ critical dimension (CD) can up to 99.8%. The sensing yield is positively correlated with TMR and negatively correlated with MTJ CD. As Fig.2(b) indicated, smaller MTJ CD leads to higher sensing margin [8]. As shown in Fig. 2(c), a BNN model consisting of two convolutional stages and one fully connected layer as the classifier is designed. Validation results on MNIST handwritten digit dataset is shown in Fig. 2(d). Sensing yield is quantized into model when sensing processes is failed, the results of convolution will be flipped. Lower TMR and larger MTJ CD result in a lower sensing yield and a dramatic decrease in accuracy. When TMR is greater than 2 with 40nm MTJ or TMR is greater than 1.5 with MTJ smaller than 35nm, the accuracy could achieve &amp;gt;95% (&amp;lt;1.8% loss). Fig. 2(e) illustrates the energy saving of single-step convolution can achieve 51.2% compared with that of row by row with a slight decrease in accuracy.

**References:**

[1] M. Courbariaux, et al., arXiv: 1602.02830, 2016.&lt;br/&gt;[2] M. Rastegari, et al., arXiv: 1603.05279, 2016.&lt;br/&gt;[3] Z. Dong et al., IEEE Transactions on Electron Devices, vol. 66, no. 1, pp. 793-801(2019)&lt;br/&gt;[4] L. Chang et al., IEEE Transactions on Very Large Scale Integration (VLSI) Systems, vol. 27, no. 11, pp. 2668-2679(2019)&lt;br/&gt;[5] Y. Pan et al., IEEE Transactions on Magnetics, vol. 54, no. 11, pp. 1-5(2018)&lt;br/&gt;[6] Y. Zhang et al., IEEE Transactions on Electron Devices, vol. 67, no. 2, pp. 469-473, Feb. 2020.&lt;br/&gt;[7] T. Na et al., IEEE Journal of Solid-State Circuits, vol. 52, no. 2, pp. 496-504, Feb. 2017.&lt;br/&gt;[8] Zhao, W. S., et al. Microelectronics Reliability, vol. 52, no. 9–10, Sept. 2012.

**Images:**

![](https://assets.underline.io/uploads/markdown_image/1/image/946bef43ce91a98b12b19a0fcba53e6c.jpg)
Fig.1 (a) Cell configurations to represent one weight using two MTJs and four transistors. (b) Schematic of OCCS SA. (c) 4T2M based array to complete convolution and fully connection, &lt;1&gt; indicates path 1 and &lt;2&gt; indicates path 2. (d)The list of all the combinations in the convolution. (e) Schematic of the XNOR and accumulate operation as convolution in memory.

![](https://assets.underline.io/uploads/markdown_image/1/image/f8614fa534b410c9fbcbfa4ee258bd1b.jpg)
Fig.2 (a) The sensing yield in different TMR with different MTJ CDs. (b) The resistance distribution of different MTJ CDs. (c) Diagram of the BNN topology used for evaluation with MNIST. (d) The performance of binary neural network with different MTJ CDs and different TMR. (e) Trade-off between energy saving and accuracy loss.

INTERMAG 2021

 In-MRAM Processing Elements with Single-Step Convolution for Binary Neural Network

#### Welcome to the 67th Annual Conference on Magnetism and Magnetic Materials, MMM 2022!

The Conference was held from October 31 to November 4 and offers both in-person and on-demand content.

MMM 2022 is sponsored jointly by AIP Publishing and the IEEE Magnetics Society. Members of the international scientific and engineering communities interested in recent developments in fundamental and applied magnetism are invited to attend and contribute to the technical sessions. The technical program will include invited and contributed papers in oral and poster sessions, invited symposia, a tutorial, and several special sessions. This Conference provides an outstanding opportunity for worldwide participants to meet their colleagues and collaborators and discuss developments in all areas of magnetism research.

<iframe width="560" height="315" src="https://s3.amazonaws.com/pf-user-files-01/u-59356/uploads/2022-11-02/un13zp7/MMM%20Welcome%20Video%20v5.mp4" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture" allowfullscreen></iframe>

For tickets to the event, please register at the following link: https://magnetism.org/registration

Please register for the MMM 2022 conference to join the event.

MMM 2022

This Conference provides an outstanding opportunity for worldwide participants to meet their colleagues and collaborators and discuss developments in all areas of magnetism research.

Recent research on deep learning shows that Binary Neural Network (BNN) is able to significantly reduce computation and memory cost with small precision loss [1][2]. By binarizing inputs and weights to +1 or -1 in BNNs, multiply-and-accumulate (MAC) operations can be replaced by XNOR operation in memory. Nonvolatile memory for BNN such as resistive random access memory (RRAM) and magnetic random access memory (MRAM) was reported in [3]-[6]. In this paper, we use spin-toque transfer magnetic random access memory (STT-MRAM) to complete XNOR and single-step convolution in memory. The main contributions include: (1) A STT-MRAM based structure with four transistors and two magnetic tunnel junctions (MTJ) (4T-2M) is proposed for binary convolution and fully connection operations using current mode sense amplifier. (2) The impact of different tunnel magnetoresistance ratio (TMR) and different size of MTJ on accumulating in memory has been explored. (3) By quantizing sensing yield into model, the trade-off between energy and performance is investigated. As shown in Fig. 1 (a), a 4T-2M bit-cell structure represents one synaptic weight, ‘-1’ is represented by two MTJs as parallel (P) and anti-parallel state (AP). The opposite combination is used to represent ‘1’. Two encoded word lines (WLs) represent one input value. To complete the XNOR operation, an offset-canceling current-sampling sense amplifier (OCCS SA) [7] is used (see Fig. 1(b)) to compare the current difference between source line (SL) and SLB. Fig.1 (c) shows the structure of the proposed STT-MRAM block. For convolution operation, the row decoder gated by appropriate timing decodes the multiple WLs to input parallel, and output through path &lt;1&gt; as the result of convolution. For fully connected operation, the row decoder activates the WLs row by row, and output through path &lt;2&gt; with bit-counting as the result of one fully connected. We implement 9 bits XNOR and accumulation to perform 3 × 3 convolution. Fig. 1(d) lists all the combinations in the convolution. The worst case is four ‘-1’ and five ‘1’ accumulation or its opposite combination, which causes the minimum current difference between “5IAP+4IP” and “5IP+4IAP”. Fig. 1(e) illustrates the circuit realization of BNN. When one group of WLs is simultaneously activated, ISL and ISLB will be different depending on the resistance of MTJs on two SLs. A CSA is implemented to generate high level when weighted sum is positive, and low level when weighted sum is negative. The output of the CSA represents the output of the neuron. SPICE-level simulation is performed using a Verilog-A model of STT MTJ and an industrial 28-nm CMOS process. Fig.2 (a) shows the simulation results that the sensing yield in different TMR and MTJ critical dimension (CD) can up to 99.8%. The sensing yield is positively correlated with TMR and negatively correlated with MTJ CD. As Fig.2(b) indicated, smaller MTJ CD leads to higher sensing margin [8]. As shown in Fig. 2(c), a BNN model consisting of two convolutional stages and one fully connected layer as the classifier is designed. Validation results on MNIST handwritten digit dataset is shown in Fig. 2(d). Sensing yield is quantized into model when sensing processes is failed, the results of convolution will be flipped. Lower TMR and larger MTJ CD result in a lower sensing yield and a dramatic decrease in accuracy. When TMR is greater than 2 with 40nm MTJ or TMR is greater than 1.5 with MTJ smaller than 35nm, the accuracy could achieve &gt;95% (&lt;1.8% loss). Fig. 2(e) illustrates the energy saving of single-step convolution can achieve 51.2% compared with that of row by row with a slight decrease in accuracy.

**References:**

[1] M. Courbariaux, et al., arXiv: 1602.02830, 2016. [2] M. Rastegari, et al., arXiv: 1603.05279, 2016. [3] Z. Dong et al., IEEE Transactions on Electron Devices, vol. 66, no. 1, pp. 793-801(2019) [4] L. Chang et al., IEEE Transactions on Very Large Scale Integration (VLSI) Systems, vol. 27, no. 11, pp. 2668-2679(2019) [5] Y. Pan et al., IEEE Transactions on Magnetics, vol. 54, no. 11, pp. 1-5(2018) [6] Y. Zhang et al., IEEE Transactions on Electron Devices, vol. 67, no. 2, pp. 469-473, Feb. 2020. [7] T. Na et al., IEEE Journal of Solid-State Circuits, vol. 52, no. 2, pp. 496-504, Feb. 2017. [8] Zhao, W. S., et al. Microelectronics Reliability, vol. 52, no. 9–10, Sept. 2012.

**Images:**

![](https://assets.underline.io/uploads/markdown_image/1/image/946bef43ce91a98b12b19a0fcba53e6c.jpg)
Fig.1 (a) Cell configurations to represent one weight using two MTJs and four transistors. (b) Schematic of OCCS SA. (c) 4T2M based array to complete convolution and fully connection, <1> indicates path 1 and <2> indicates path 2. (d)The list of all the combinations in the convolution. (e) Schematic of the XNOR and accumulate operation as convolution in memory.

![](https://assets.underline.io/uploads/markdown_image/1/image/f8614fa534b410c9fbcbfa4ee258bd1b.jpg)
Fig.2 (a) The sensing yield in different TMR with different MTJ CDs. (b) The resistance distribution of different MTJ CDs. (c) Diagram of the BNN topology used for evaluation with MNIST. (d) The performance of binary neural network with different MTJ CDs and different TMR. (e) Trade-off between energy saving and accuracy loss.

technical paper

INTERMAG 2021 Welcome Video



	
<figure class="video_container">
 <iframe src="https://screencast-o-matic.com/player/crflqFVnUpZ?v=6&ff=1&title=0&controls=1" width=640 height=360 frameborder="0" allowfullscreen="true"> </iframe>
</figure>

Want to buy a ticket? 

Register for INTERMAG 2021 here

[![](https://assets.underline.io/uploads/markdown_image/1/image/17ce76e7960db633258024eef8d6da7f.png)](https://intermag.org/register)

Got your ticket, but still can’t access the event?

Contact us and our team will look into it.

Magnetic tunnel junctions offer significant and near-term improvements to new computer chips, most immediately from non-volatile memory applications [1]. These devices consist of two magnetic layers whose relative alignment – parallel or antiparallel – determine the device's two-terminal resistance. When the energy barrier between these two states is sufficiently small, thermal fluctuations can cause the device to exhibit random telegraph noise in its resistive state. In this case, the device is called a superparamagnetic tunnel junction [2]. Computing with random telegraph noise can offer favorable properties such as noise-tolerance, massive parallelism, and – for certain operations – low power and area footprints. The field of stochastic computing addresses questions of engineering such systems but has mostly been restricted to the use of low-quality pseudorandom number generators. The low periodicity and high cross-correlations between these devices have restricted traditional stochastic computing by forcing frequent regeneration of stochastic bitstreams. Crossing in and out of the stochastic domain can incur significant energy penalties [3]. In our work [4], we introduce a low-energy circuit for extracting programmable randomness from superparamagnetic tunnel junctions. The required probability is determined via a digital input signal, rather than an analog current, so ohmic losses are minimized (Figure 1). Using detailed analog circuit simulation, we find that our device requires only half the energy of the traditional all-CMOS solution, while also offering infinite-period truly random bitstreams. To demonstrate the value of truly random bitstreams, we designed a deep neural network modeled after LeNet5 [5], a famous convolutional neural network used for handwriting recognition. Traditional neural networks multiply neuron outputs by a synaptic weight, add these products together at the input of another neuron, and then apply a nonlinear operation to that input sum. In our implementation, the values passed between neurons are encoded in the expected value of the voltage on any given wire. To multiply two signals together, we need only use a Boolean AND gate to construct the joint distribution of the wires. Feeding many wires into a Boolean OR gate generates a polynomial function which looks like linear addition near the origin, but acquires high-order nonlinear corrections when the inputs are strong. However, the OR gate only functions in this way because the inputs can be regarded as statistically independent, which would be impossible with pseudorandom number generators absent an unacceptable increase in energy consumption. We implement LeNet5 using superparamagnetic tunnel junctions together with these simple logic gates and train a simulation of this deep neural network. We are able to achieve 97 % accuracy on handwritten digit recognition using only 150 nJ per inference. This energy efficiency represents a factor of 1.4 to 7.7 improvement over comparable proposals that do not use tunnel junctions. However, after systematic optimization, the best CMOS-only solution is able to achieve 99.1 % accuracy [6]. Questions about the best way to train OR-gate neurons for optimal performance is a question for future research. Nevertheless, our demonstration shows that taking a digital approach to utilizing SMTJ fluctuations can unlock extremely energy-efficient circuit designs with the potential for offering stochastic neural networks computing in near-term application specific integrated circuits.

**References:**

[1] Apalkov, Dieny, and Slaughter, Proc. IEEE 104, 1796 (2016) [2] Camsari, Faria, Sutton, and Datta, Phys. Rev. X 7, 031014 (2017) [3] Manohar, IEEE Comp. Arch. Lett. 14, 119 (2015) [4] Daniels, Madhavan, Talatchian, et al., Phys. Rev. App. 13, 034016 (2020) [5] LeCun, Bottou, Bengio, et al., Proc. IEEE 86, 2278 (1998) [6] Li, Ji, Ren, et al., IEEE Trans. Comp. Aid. Des. Int. Circ. Sys. 38, 1543 (2019)

**Images:**

![](https://assets.underline.io/uploads/markdown_image/1/image/a7dfe1770e892cca4d62d45c58f83e71.jpg)
Neurons in the brain have analog outputs in analog time; superparamagnetic tunnel junctions have digital states in analog time. To use digital circuitry to discretize the time variable, removing the need for steady state currents and the associated ohmic losses.

 Running Deep Neural Networks on Superparamagnetic Fluctuations

There is a growing pursuit to create neuromorphic devices that are able to perform machine learning tasks, both in an energy efficient manner and compatible with CMOS architecture. These approaches are based on taking the complex response of materials and mimicking aspects of a given machine learning model. Typically, what cannot be mimicked directly in a material, such as learning, is supplemented by other devices, circuitry, or external computers. These various approaches often face a number of bottlenecks, related to scaling, functionality independent of computers, and energy efficient operation. Ultimately constructing autonomous materials, namely individual materials that independently adapt their physical properties based on input data, requires deeper fundamental research linking physical processes to machine learning concepts. In this way, it is necessary to link quantum phenomena in materials to concepts like plasticity, memory, and learning. In this way, new types of quantum materials can be designed for autonomous recognition, as well as new functionalities can be explored, such as quantum machine learning. Energy-based models, like the Boltzmann machine, provide a natural link between physics and machine learning functionality via the Ising model and the physics of multi-well potentials and glasses (1, 2). I will discuss a new model system, based on the recently developed concept of orbital memory (3), which mimics the behavior of a Boltzmann machine directly in its stochastic dynamics and in one elemental material, down to the atomic-scale (4). Using scanning tunneling microscopy (STM) to fabricate nano-scale arrays of cobalt atoms on semiconducting black phosphorus, we observe stochastic noise described by the tunable multi-well landscapes seen in stochastic energy-based models. Utilizing a concept based on separation of time scales and the anisotropic behavior of black phosphorus (5), I will show that these multi-well landscapes are adaptive and exhibit inherently coupled neural and synaptic dynamics, at the atomic-scale. I will discuss perspectives of learning in these systems, and how such model systems provide fundamental insight into the future design of quantum-based brain-like devices. This work was done in collaboration with H.J. Kappen, Brian Kiraly, Elze J. Knol, and Werner M.J. van Weerdenburg. This work was funded by NWO-Vidi Manipulating the interplay between superconductivity and chiral magnetism at the single-atom level” with project number 680-47-534. This project has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement No 818399).

**References:**

 1. A. Kolmus, M. I. Katsnelson, A. A. Khajetoorians and H. J. Kappen, New J. Phys. 22 (2), 023038 (2020). 2. U. Kamber, A. Bergman, A. Eich, D. Iusan, M. Steinbrecher, N. Hauptmann, L. Nordström, M. I. Katsnelson, D. Wegner, O. Eriksson and A. A. Khajetoorians, Science 368 (6494), eaay6757 (2020). 3. B. Kiraly, A. N. Rudenko, W. M. J. van Weerdenburg, D. Wegner, M. I. Katsnelson and A. A. Khajetoorians, Nat. Commun. 9 (1), 3904 (2018). 4. B. Kiraly, E. J. Knol, H. J. Kappen and A. A. Khajetoorians, arXiv, 2005.01547 (2020). 5. B. Kiraly, E. J. Knol, K. Volckaert, D. Biswas, A. N. Rudenko, D. A. Prishchenko, V. G. Mazurenko, M. I. Katsnelson, P. Hofmann, D. Wegner and A. A. Khajetoorians, Phys. Rev. Lett. 123 (21), 216403 (2019).

**Images:**

 An atomic-scale Boltzmann machine capable of self-adaption

Ferromagnetic devices, where current pulses flowing through the heavy metal layer below the ferromagnetic layer move the ferromagnetic domain wall through spin-orbit torque, have been proposed and experimentally demonstrated as synapses for neuromorphic computing [1–3]. But since ferrimagnets are known to exhibit faster domain wall motion compared to ferromagnets, the potential for ferrimagnetic synapses also needs to be explored [4]. Here, we first carry out micromagnetic simulations of current-induced domain wall motion in a ferromagnetic device as well as a ferrimagnetic Co/Gd-bilayer-based device. The working principle of the ferrimagnetic synapse is just the same as the ferromagnetic synapse: current flowing through the underlying heavy metal layer moves the domain wall in the ferrimagnetic bi-layer just like it moves it in the ferromagnetic layer (Fig. 1a) [4]. Motion of the domain wall leads to a change in conductance and hence an update in the weight stored in the synapse [1, 2]. In our simulation, the spin Hall angle of the heavy metal (Pt), current flow through which generates the necessary spin-orbit toque for domain wall motion, is considered to be 0.13 [4]. For the ferrimagnetic system, dynamics of the magnetic moments of both the Co and Gd layers, exchange-coupled with each other (coupling strength = -0.9 mJ/m2), has been modeled in ‘mumax3’ [4]. The temperature-dependent variation of the Co and Gd magnetic moments is also considered. From our simulations (Fig. 1b), we observe that in the ferrimagnetic system, the domain wall velocity is highest at the angular-momentum-compensation temperature, T= 165 K, at which moment in Gd layer (mGd) = 0.9 times moment in Co layer (mCo). This is consistent with the experimental observation and corresponding one-dimensional-domain-wall calculations reported by Blasing et al [4]. Fig. 2a shows that across a wide range of current density through the heavy metal layer, domain wall velocity in the ferrimagnetic system at the angular-momentum-compensation temperature (165 K) is 2X–2.5X higher than that in the ferromagnetic system at room temperature. Next, we carry out system-level simulations with the device characteristic in Fig. 2a to compare the performance of the ferromagnetic and the ferrimagnetic devices as synapses in a crossbar-array-based neuromorphic system that mimics a Fully Connected Neural Network (FCNN) with a hidden layer. Synaptic weights are updated in both the ferromagnet-synapse and ferrimagnet-synapse crossbars using the thresholding algorithm, proposed by Kaushik et al. [5], to achieve on-chip learning in the crossbars on Iris data set of flowers and MNIST data set of handwritten digits. Using this thresholding algorithm, for Iris data set, we obtain train accuracy of 100% and test accuracy of 96%, and for MNIST, we obtain train accuracy of 98% and test accuracy of 92%. We obtain these same accuracy numbers for both ferromagnetic and ferrimagnetic synapse crossbars because we allow the same number of conductance states for both these types of synapses (50). According to this thresholding algorithm, for each synapse in the FCNN, the synaptic weight/ conductance is either increased or decreased by a fixed amount or kept the same at any particular iteration [5]. Thus training can be achieved only by applying current pulses of fixed magnitude and duration to the ferrimagnetic or ferromagnetic synapses — positive polarity of current for weight increase and negative polarity for weight decrease (Fig. 1a). A shorter duration pulse needs the domain wall to move with a higher velocity compared to a longer duration pulse to cause the same weight update. Hence, a shorter pulse needs to be of higher current magnitude and will consume more energy. Fig. 2b shows how energy consumed by a pulse increases with a decrease in duration of the pulse, both in the case of the ferrimagnetic system and the ferromagnetic system. We do observe that for the same duration of current pulse, energy consumed is 4X lower for the ferrimagnetic system than for the ferromagnetic system. Similarly, if the energy consumption per pulse is kept constant, pulses of time duration 4X lower can be used for the ferrimagnet compared to the ferromagnet. As a result, when we compare the energy consumption and time taken for on-chip learning in the two aforementioned crossbar systems, we observe that for the same net energy consumption in the system, the ferrimagnetic crossbar can be trained 4X faster than the ferromagnetic crossbar. Similarly, for the same speed across the two crossbars, the ferrimagnetic crossbar can be trained at 4X lower energy than the ferromagnetic crossbar. Thus, our study shows that ferrimagnetic-bilayer-based systems can pave the way for fast and energy-efficient neuromorphic computing in the future.

**References:**

[1] A. Sengupta and K. Roy, IEEE Trans. Biomed. Circuits Syst., vol. 10, no. 6 (2016) [2] D. Kaushik, U. Singh, U. Sahu, I. Sreedevi, and D. Bhowmik. AIP Adv. vol. 10, no. 025111 (2020) [3] R. Blasing, T. Ma, S-H. Yang et al. Nature Communications, vol. 9, no. 4984 (2018) [4] T. Jin, W. Gan, F. Tan et al, J. Phys. D: Appl. Phys., vol. 52, no. 445001 (2019) [5] D. Kaushik, J. Sharda, and D. Bhowmik. Nanotechnology vol. 31, no. 364004 (2020)

**Images:**

![](https://assets.underline.io/uploads/markdown_image/1/image/a5a1e18c735248823a90d465d55cd209.jpg)
Fig. 1 (a) Schematics of ferrimagnetic (Co/Gd-bilayer-based) and ferromagnetic (Co-based) synaptic devices. The domain wall is shown through rotating green arrows, which correspond to magnetic moments. (b) Domain-wall velocity vs current density flowing through the heavy metal layer (Pt), for different values of temperature, which correspond to different ratios of Gd’s magnetic moment (mGd) to Co’s magnetic moment (mCo).

![](https://assets.underline.io/uploads/markdown_image/1/image/eab3ad1532ecb07e4520b5a23f9f3af2.jpg)
Fig. 2. Comparison of (a) domain-wall velocity vs current density characteristic and (b) energy per synaptic-weight-update pulse vs time duration of such a pulse for ferrimagnetic synapses (at angular-momentum compensation temperature, when domain-wall velocity is highest) and ferromagnetic synapses (at room temperature).

 Ferrimagnetic Co-Gd-Bilayer-Based Fast and Energy-Efficient Synaptic Devices for Neuromorphic Computing

There are many potential applications for spintronics including in-memory, neuromorphic, and radiation hard computing. In- memory computing offers a solution to the memory wall bottleneck that hinders traditional architectures. Neuromorphic computing has the potential to revolutionize how data-intensive tasks, such as image recognition, are computed by emulating how a biological brain operates. Radiation hard computing is useful for certain situations, mainly space applications, where the Earth’s atmosphere does not protect sensitive CMOS hardware. All of these applications can be achieved using domain wall motion-based devices. The domain wall magnetic tunnel junction (DW-MTJ) device architecture is explored in this work as shown in figure 1. DW-MTJ devices operate by sending a current pulse along a ferromagnetic wire to propagate a domain wall from the input to the output of the device. As the domain wall passes under the MTJ, the magnetization state of the bottom layer of the CoFeB/MgO/CoFeB junction changes. By controlling where the domain wall is, the MTJ can have programmable resistance. We use current driven domain wall motion to electrically switch the devices. It has been shown that spin logic requires high tunneling magnetoresistance, low switching voltages, and high stability. These hurdles have been overcome in this work by utilizing perpendicular anisotropy, spin orbit torque switching, and optimized fabrication. Incorvia et al. in 2015 demonstrated that information could be encoded in a magnetic wire by moving a domain wall, but the TMR of those devices was 12% [1]. Here, we develop a process that maintains as much TMR as possible, achieving TMR &gt; 150%, as well as no reduction in resistance-area (RA) product after device fabrication. There were several improvements made to the previous fabrication process. The main change was to remove the reactive ion etch step to contact the MTJ. Incorvia et al. opened a via in polymethylmethacrylate (PMMA) and then used a CF4 reactive ion etch through a hydrogen silsesquioxane (HSQ) layer to expose the MTJ and wire contacts. In our new process, we replaced this step with a negative tone resist and encapsulation step to achieve the same result of an open via to the MTJ and domain wall track without exposing the sensitive layers to a reactive etch. This way the MTJ is never directly exposed: during all etches and encapsulation, the MgO is protected by negative resist. Furthermore, the encapsulation was performed at relatively low temperatures of 100 °C to prevent any additional loss of TMR, and silicon nitride was used instead of silicon oxide to reduce oxidation. These factors allowed our devices to maintain their high TMR post-processing. The TMR is shown in the field loops of four devices in figure 2. The average achieved was 164±17%, compared to the unpatterned film TMR of 168±6%. Before fabrication, the film had RA = 35±2 Ω-μm2 and the devices have RA = 31±3 Ω-µm2. This indicates that our fabrication process had little to no negative impact on the magnetic integrity of our material. These devices utilize electrical initialization of domain walls rather than an external magnetic field, so they can be fabricated as straight wires rather than curved. This allows for increased density if this technology becomes industrialized. Previous iterations of these devices were curved to allow an external magnetic field to nucleate a domain wall at an initial position. Here, we opted for an electrical approach by implementing an oersted field line near the left contact of each device. A short pulse is sent through this line which nucleates a domain wall in the ferromagnetic track [2]. This is one step closer towards all electric operation of the memory-in-logic devices. These oersted field lines increased stability and repeatability by nucleating the domain wall in the same position each time, rather than an unpredictable location via external magnetic field nucleation. Calculating the variability based on the switching voltage range over 10 cycles with both nucleation methods showed an improvement from external magnetic field (105%) variability to the oersted field line (7%) variability. We show in simulation that these improvements bring the variability of the technology within the regime for large circuit function such as a 32-bit full adder [3]. Partial funding from Sandia National Lab (SNL). SNL is managed and operated by NTESS under DOE NNSA contract DE-NA0003525.

**References:**

1J. A. Currivan-Incorvia, S. Siddiqui, S. Dutta, E. R. Evarts, C. A. Ross & M. A. Baldo. 2015 IEEE Int. Electron Devices Meet. 32–36 (2015) 2G. Zahnd, V. T. Pham, A. Marty, M. Jamet, C. Beigné, L. Notin, C. Vergnaud, F. Rortais, L. Vila & J. P. Attané. J. Magn. Magn. Mater. Vol. 406, p. 166–170 (2016) 3T. Patrick Xiao, M. J. Marinella, C. H. Bennett, X. Hu, B. Feinberg, R. Jacobs-Gedrim, S. Agarwal, J. S. Brunhaver, J. S. Friedman & J. A. C. Incorvia. IEEE J. Explor. Solid-State Comput. Devices Circuits., Vol. 5, p. 188-196 (2019)

**Images:**

![](https://assets.underline.io/uploads/markdown_image/1/image/bee6a10b4751acb8eb32a297ccf3365a.jpg)
Fig.1: Top-down SEM image of a DW-MTJ device. Patterned domain wall track (labeled DW track) and magnetic tunnel junction (labeled MTJ) identified. Device width (labeled w) is 450 nm. Widths down to 250 nm were patterned and tested.

![](https://assets.underline.io/uploads/markdown_image/1/image/694813823627de4f60ea1974a8edb598.jpg)
Fig. 2: Field loop for devices in this work. RMTJ is the measured resistance through the tunnel junction. Spin-orbit torque current switching was also performed and will be presented.

 Domain Wall-Magnetic Tunnel Junction Spin Orbit Torque Devices for In-Memory Computing

Stochastic behaviour has traditionally been a limiting factor in the development of nanomagnetic technology. However, it can also give rise to rich physical behaviours. For example, in artificial spin ice (ASI) arrays [1,2] stochastic effects give rise to emergent behaviour that can help us to understand phenomena in frustrated systems. In particular, this emergent behaviour is useful for a novel form of neuromorphic computing called ‘reservoir computing’ which is highly efficient for time domain processing of signals [3]. One such system that exhibits probabilistic, emergent behaviour with complexity is an array of interconnected nanowire rings of Ni80Fe20 [4-5]. Here we characterise the emergent behaviour of domain wall (DW) states in a nanoring array using X-ray imaging under the application of rotating fields of different amplitudes in order to better understand the array’s magnetic response and ascertain its suitability for reservoir computing. Fig. 1 (A) shows an SEM micrograph of a representative rectangular Ni80Fe20 ring array of which we have obtained the magnetic response by applying an in-plane rotating magnetic field. Magneto-optic Kerr effect (MOKE) measurements show a strongly nonlinear magnetisation variation with field amplitude (Fig. 1 (B)) - an essential response for reservoir computing. We have shown this system to be a promising candidate for reservoir computing using various macroscopic measurement techniques [5-6]. It is clear that these complex dynamics are due to the interplay of pinned and propagating DWs (Fig. 1 (B)) and different magnetic states in the array. We used X-ray Photo Emission Electron Microscopy (X-PEEM) to image the magnetic states across a ring array with rings of radius 2 microns and width of 500 nm with an Ni80Fe20 thickness of 10 nm and with a 2nm Al layer to prevent oxidation. X-PEEM images were obtained with an Elmitec SPELEEM-III microscope on the I06 beamline at the Diamond Light Source using XMCD at the Fe-L3 edge. Samples were mounted on cartridges with a quadrupole magnet, to provide an in-plane magnetic field with arbitrary direction, designed and built at the CIRCE XPEEM beamline at the ALBA Synchrotron. Samples were subject to an initialisation field pulse of 160 Oe followed by 30 rotations of an in-plane magnetic field of various strengths. A representative X-PEEM image of the array (Fig. 2 (A)) obtained with a rotating field strength of 30 Oe at a frequency of 8 Hz shows multiple magnetisation states. The allowed magnetic configurations are shown schematically in Fig. 2 (B). There are four types of ‘onion’ states (two DWs separated by 180°), two ‘vortex’ states (zero DWs) and eight ‘three-quarter’ states (two DWs separated by 90°) possible. A dedicated Python code was used to analyse the magnetisation states in the X-PEEM image (Fig. 2 (A)). The rings were mapped to circles using the Hough transform available in the OpenCV library [7] and the color of each quarter was recorded. Then the color scheme of an entire circle was assigned to a magnetic configuration. The population of various states at different amplitudes of the rotating field was extracted and is shown in Fig. 2 (C). We can see that ‘onion’ states (at lower fields) give way to ‘vortex’ and ‘three-quarter’ states and at higher fields ‘onion’ states dominate again. This is consistent with the previous MOKE measurements that were attributed to static onion states at low field magnitudes, due to DW pinning, and dynamic onion states at high fields (as shown in Fig. 1 (B)) when DWs readily overcome the pinning potentials presented by ring junctions. Note that this behaviour leads to the emergent dynamics which the ring assembly shows under the influence of a rotating magnetic field. The variation in the magnetic states population also indicates ‘fading memory’ as the state population at a given field is independent of the population at a previous field. This is another desirable property for a system to function as a reservoir. The normalised magnetisation response of the array can be extracted from the state population (Fig. 2 (D)) and in order to measure the array’s response to field strength. This is similar to that obtained from previous MOKE measurements but highlights the significance of dynamic local variation in magnetic structure. We have thus imaged the emergent DW dynamics in nanoring arrays using X-PEEM and have obtained their magnetic response to show their suitability for reservoir computing applications. The different complex magnetic states in the array, which give rise to their emergent magnetic behaviour, provide insights into the evolution of these states. We hope that this study will pave the way for further investigations into the suitability and implementation of these systems for advanced computing purposes.

**References:**

[1] Wang et al., Nature 439, 303 (2006) [2] Branford et al., Science 335, 1597 (2012) [3] Jensen et al., ALIFE 2018: The 2018 Conference on Artificial Life (July 2018) pp.15-22 [4] Negoita et al., J. Appl. Phys. 114, 013904 (2013) [5] R. Dawidek et al., Adv. Funct. Mater. 2021, 2008389 [6] I. Vidamour et al. presented at MMM-2020 [7] Bradski, Dr. Dobb’s Journal of Software Tools 2236121 (2000)

**Images:**

![](https://assets.underline.io/uploads/markdown_image/1/image/4c1d76f0e5846822be2b0bb1d2560a9a.jpg)
Fig. 1: (A) An SEM image of a fabricated ring array. (B) MOKE measurements of the array showing the ‘nonlinear response’ of its magnetization and the different dynamic states in the system.

![](https://assets.underline.io/uploads/markdown_image/1/image/0b16e317875081de61e6fa8d49103c5f.jpg)
Fig. 2: (A) A representative X-PEEM image of a ring array. (B) A schematic showing the different types of magnetic states possible in the array. The arrows show the magnetisation direction which the measurement was sensitive to. (C) The variation of the magnetisation states in the array with rotating field amplitude and (D) the normalised magnetisation derived from the magnetisation states.

 Imaging the emergent behaviour in nanoring assemblies for reservoir computing applications

Reservoir computing has been gaining traction as a bio-inspired, machine-learning approach for solving computationally hard problems, such as speech recognition and prediction of chaotic time series, that are becoming fundamental to modern technology [1]. By making use of a complex, fixed reservoir with a single output layer, the approach minimises expensive training steps. Magnetic systems are well positioned to fulfill the requirements of the reservoir, providing non-linearity, fading memory, and reproducible responses. In particular, magnetic systems subject to a temporal input sequence can perform the entire role of the reservoir, replacing complex networks of transistors whilst only requiring a single input and output [2]. This offers the potential for compact devices, with reduced energy costs for computation. Here, we propose strain-mediated, voltage controlled superparamagnetic ensembles as an ideal candidate for an ultra-low-energy reservoir. Single superparamagnetic nano-dots with uniaxial anisotropy exhibit stochastic behavior, switching between states with a characteristic timescale set by their energy barriers. Micromagnetic simulations (Fig. 1a) demonstrate this telegraphing behavior for CoFeB nano-dots, where the auto-correlation time has been verified to be below 5 ns and the strain-induced anisotropy can compete with the intrinsic anisotropy [3-6]. Borders et al. [7] have shown how tuning such telegraph noise in magnetic tunnel junctions can be used for integer factorization. Here, however, we make use of the collective behavior of extended ensembles of nano-dots such that the average magnetisation becomes predictable; thus, showing the reproducible response required for reservoir computing. Input is provided by addition of a strain-mediated anisotropy (Fig. 1b)—a technique that is predicted to achieve ultra-low-power consumption [8]. This acts to rotate the intrinsic anisotropy axis, biasing dwell time in one state over the other [9]. Whilst the internal system dynamics are driven by thermal noise, the average magnetisation obeys simple rules that, when subject to a temporally varying input, produce a complex non-linear response with fading memory (Fig. 1c). Using an analytical model of the response of the superparamagnetic ensemble [10], we simulate the physical reservoir. By inputting a temporal sequence and training a single layer of weights from the output, we can perform standard machine learning benchmarks with competitive performance over a range of input timescales. Fig. 2 presents the performance for spoken digit recognition on the TI-46 dataset and chaotic time series prediction on the NARMA10 task. In both cases the performance approaches that of competing reservoirs [1,2], despite the relative simplicity of the system. The intrinsic timescale of the system (τ) is controlled by the ratio of energy barrier (anisotropy, KV) to the thermal energy (KBT at 300 K). Fig. 2a demonstrates that by tuning the input rate and strength (relative to the intrinsic timescale and anisotropy respectively), optimal performance of 95 % recognition of the spoken digits 0-9 with five female speakers can be achieved. This is realised for τ ~ 50 ns (KV/KBT = 20) and τ ~ 1 ms (KV/KBT = 50). For the NARMA10 task (Fig. 2b), the highest performance is achieved across this entire timescale range by tuning the strength of a feedback term (i.e. the output is fed back into the input stream with a given weight). Robust performance on timescales ranging from hundreds of nanoseconds up to seconds would allow a physical realisation of such a reservoir to be tuned to provide computation in real time for a wide range of possible physical inputs: from decision-making in driverless cars (fast) to speech recognition (slow). The simplicity of the system and the ease of manufacture, coupled with the low energy consumption expected for such a, thermally driven, device makes it an ideal candidate for edge computing where high performance is needed at very low latency and power.

**References:**

[1] L. Appeltant et al., Nat Commun. 2, 468 (2011). [2] J. Torrejon et al., Nature. 547, 428–431 (2017). [3] A. Vansteenkiste et al., AIP Advances. 107133 (2014). [4] J. Leliaert et al., AIP Advances. 7, 125010 (2017). [5] S. Zhang et al., Scientific Reports. 4, 3727 (2014). [6] C. Safranski et al., arXiv:2010.14393[cond-mat] (2020). [7] W. A. Borders et al., Nature. 573, 390–393 (2019). [8] K. Roy, S. Bandyopadhyay and J. Atulasimha, Applied Physics Letters. 99 (2011). [9] Z. Guo et al., Appl. Phys. Lett. 112, 052904 (2018). [10] O. Hovorka et al., Appl. Phys. Lett. 97, 062504 (2010).

**Images:**

![](https://assets.underline.io/uploads/markdown_image/1/image/2ab92e894162fe72ba365c3a5186c89a.jpg)
Fig. 1: a. Telegraphing behaviour of a single superparamagnetic nano-dot. Micromagnetic simulation for a 40 nm diameter, 4 nm thick CoFeB cylinder with 8.2 KJ/m-3 uniaxial anisotropy at 300 K. b. Rotation of the intrinsic anisotropy axis in a nano-dot via strain anisotropy. The rotation biases the dwell time in one magnetic state (left/right pointing) versus the other. c. Analytical response of an ensemble of superparamagnetic nano-dots to a global input via strain. A temporal input sequence produces the non-linear response with fading memory required for reservoir computing.

![](https://assets.underline.io/uploads/markdown_image/1/image/7c8ac57f39d07c9f25ded428f97ed89f.png)
Fig 2: a. Performance on a spoken-digit recognition test using the TI-46 dataset. Performance is tuned by the input rate (relative to the intrinsic timescale) and strength. Optimal performance (95 % recognition accuracy) is achieved for intrinsic timescales of 50 ns (KV/KBT = 20, left) and 1 ms (KV/KBT = 50, right) and approaches that of competing reservoirs [2]. b. Performance on a chaotic time series prediction task (NARMA10). Performance is maintained across the entire timescale range by tuning a feedback parameter.

 Voltage-Controlled, Thermally Driven Superparamagnetic Ensembles for Tuneable Timescale Reservoir Computing

Stochastic computing has recently gain attention mainly due to its low cost, fast and energy efficient circuit level implementations. In addition, it is rather fault-tolerant and noise resilient. The success of stochastic computing is due to the fact that complex functions are extremely difficult to be implemented by conventional binary logic design while they can be effectively approximated by stochastic computing. Stochastic computing uses probabilistic representation of real-valued numbers. It represents and processes numbers thanks to bitstreams. Such a stream represents the number p if the proportion of digit 1 in the stream is equal to p [1]. One of the major requirements for accurate stochastic computation is to use non-repeating and uncorrelated random number generators to produce the bitstreams. In pure digital conventional circuits, such randomness arises from Linear Feedback Shift Register (LFSR) based Random Number Generator (RNG). They output a bitstream that represents the quantity 0.5, and is then processed to represent whatever number [1]. However, LFSRs are made of CMOS logic gates and flip-flops. The produced bitstream is totally predictable, requires an important amount of energy along with a high silicon area. Recently, nanoscale devices such as Magnetic Tunnel Junctions (MTJ) have been considered for bistream generators due to their inner randomness [2]. The so-designed circuits produce pure random bitstreams (the sequence is not deterministic), and they operate in a very energy efficient way. The MTJ is a structure in which a binary information is stored as the relative direction of the magnetization of a magnet with respect to a reference one. The behaviour of this magnetization depends in part on thermal fluctuations [2], which can be leveraged as a source of stochasticity. Two types of schemes have been proposed for an MTJ based bitstream generator. First, a MTJ based RNG [3-5]: a sequence of bits ‘0’ and ‘1’ with a proportion of 50% each is generated, then processed to produce the bitstream. Second, a programmable circuit tunes the MTJ in such way that it generates directly the required bitstream, without the need of any post-processing [6]. Its implementation implies less hardware cost. However, full circuit implementation of the second scheme has not been proposed so far. In this work, we designed a bitstream generator composed of a single MTJ, a writing circuit, a reading circuit and a feedback loop. The operating principle, schematically represented in Figure 1, is the following: a current passes through the writing path of the MTJ (that has been previously set in the logic state “0”) during a short period with a specific amplitude. This brings about a probability (denoted p) that the magnetization experiences a switching. After repeating the process N times, the resulting bitstream is composed of an average of Np bit seen as “1”. The obtained bitstream then represents the number p in its stochastic representation. The reading circuit, similar to the latch-type voltage-mode sense amplifier published in [7], designed and optimized to allow asynchronous operations for low power operation, has the advantage that the static currents required to measure the resistance of the MTJ last a very short time, allowing for important energy saving. The inner CMOS based latch decorrelates the output of the reading circuit to the MTJ so that the state of the read MTJ is stored and remains available after the reading process, while the MTJ can be used again. The feedback loop, connected to the output of the reading circuit extracts the number that is represented by determining the ratio of bits at “1” and “0”. It determines the gap between the number that is computed with respect to the one that is to be represented. In the case of a large gap, the writing current is modified accordingly. This current is increased if the there are too many “0”s, and reversely it is decreased when there are too many “1”s. The writing unit is connected in series to the writing path of the MTJ. It aims at imposing a current through the writing path that is determined by the feedback loop. It is made of several transistors in parallel, acting like independent current sources. This configuration allows the circuit to find the proper writing current required for a given bitstream. This also allows handling temperature dependence of the writing current. Indeed, with temperature variation the writing current may vary in time. In our proposed circuit, the real-time tracking of the optimal writing current by the proposed feedback loop handles this problem. In conclusion, we propose an innovative single MTJ based bitstream generator leveraging the unique stochastic properties of the magnetization of the storage layer. The critical issue of PVT variations and noise are solved by our system thanks to the feedback loop and to the versatility of the writing circuit that allow for the generation of a relevant writing current that can be generalized to any bitstream computation. This study opens the way to the design of fast and energy efficient stochastic computing based circuit.

**References:**

[1] A. Alaghi and J. P. Hayes, ACM Trans. Embed. Comput. Syst [2] W. Brown, Phys. Rev. 130 (1963) 1677-1685 [3] Akio Fukushima et al 2014 Appl. Phys. Express 7 083001 [4] F. Ouattara, A. Nejat and L. Torres, IEEE Access, vol. 7, pp. 59271-59277, 2019 [5] S. Oosawa, T. Konishi and N. Onizawa, IEEE 13th International New Circuits and Systems Conference (NEWCAS), Grenoble, 2015, pp. 1-4 [6] LA de Barros Naviner et al., 13th IEEE International New Circuits and Systems Conference (NEWCAS), 2015 [7] H. Zhang, W. Kang and, IEEE Access, vol. 6, pp. 64250–64260

**Images:**

![](https://assets.underline.io/uploads/markdown_image/1/image/6719b321db54b403c7f441a57764f6a8.jpg)
Fig.1 : Schematic of the bitstream generator

 Hardware Implementation of a Magnetic Tunnel Junction based Bitstream Generator for Stochastic Computing

The spintronic devices such as the domain wall and skyrmion based devices are being explored for their applications in neuromorphic computing [1-3], due to their non-volatile behaviour, stochastic switching and energy efficiency. The electric field control of these devices has attracted extensive attention and has further gained popularity both in memory and logic applications as it provides an efficient way to improve the data storage density [4-8]. Neuromorphic computing has been shown to be a superior alternative in terms of speed and energy, when compared to training on a conventional computing unit. A magnetic tunnel junction (MTJ), which constitutes the building block of a magnetic random-access memory (MRAM), is a competitive candidate for the next generation memory devices because it is non-volatile and exhibits high endurance, low power consumption, high operation speed and integration capability [8-9]. The present work discusses the proposal of a spintronic neuromorphic system with spin orbit torque driven domain wall motion-based neuron and synapse. The domain wall motion in the neuron is controlled by the electric field at the gate generated by voltage signals from pre-neurons as shown in Fig. 1(a). The application of voltage as input and output variable helps in reducing power consumption, as pinning site can be turned ON/OFF only when required. Further it provides better fanout as post-neuron output from first stage can act as infinite number of pre-neurons to next stage, which is going to play an important role in realization of system level neuromorphic architectures. For the implementation of the synaptic weight, we propose 3- terminal MTJ with stochastic domain wall motion in the free layer. The edge roughness results in intrinsic pinning of domain wall and this adds to the stochastic behaviour of domain wall motion in presence of spin transfer torque and/or spin orbit torque [10]. We have incorporated these intrinsic pinning effects by creating triangular notches on the sides of free layer as shown in Fig. 1(b). The complete neuromorphic circuit containing neurons, synapses and control circuitry implementation of spike time dependent plasticity is shown in Fig. 1(a). The sensing unit consists of a voltage comparator which provides high positive or zero voltage feedback pulse to the weight-updating unit. The weight updating unit takes this signal and the pre-neuron signal to decide the sign of the synaptic weight update pulse based upon spike timing correlation between the pre- and post-neuron. For modelling the micromagnetics and spin transport in both synapse and neuron, we developed a micromagnetic Non-Equilibrium Green’s Function (MuMag-NEGF) coupled model. In Fig. 2(a), we show that in the presence of the gate voltage the uniaxial anisotropy for a 20 nm-wide region is varied slightly which results in the pinning of domain wall at 40 nm from the centre. In the absence of gate voltage uniaxial anisotropy is uniform and the domain wall moves more than 100 nm right in 10 ns, where it is sensed by the MTJ as post neuron spike. Also, the reduction of domain wall velocity is shown in Fig. 2(b). The thermal effects result in the stochastic domain wall motion but stochasticity is tuned by the external current in the form of spin orbit torque. Fig. 2 (c) shows the stochastic magnetization switching in MTJ synapse and stochasticity is tuned in either direction with application of positive and negative current pulses. In the presence of more positive pulses, the MTJ gradually moves into low resistance state, while in the presence of more negative pulses it moves into high resistance state, as shown in Fig. 2(d). When a current pulse is applied, the probability of domain wall moving increases. The dynamic pattern recognition capabilities of the circuit are tested on 9x9 cross-bar architecture, consisting of 9 pre-neurons, 9 post-neurons and 81 synapses. In summary, here we have successfully designed and demonstrated the on-chip stochastic dynamic learning by adopting the spike time dependent plasticity as the learning algorithm. At the end of training phase, we demonstrated that all pattern synapses stochastically evolve into low resistance state (high weight) while the remaining background synapses are in high resistance state (low weight).

**References:**

[1]. Deming Zhang, Weisheng Zhao et. al, “All spin neural networks based on compound spintronic synapses and neurons,” IEEE Advances in Biomedical Circuits and Systems, vol. 10, no.4, August-2016. [2]. Otitoaleke Akinola, Jean Anne C Incorvia et.al, “Three terminal magnetic tunnel junction circuits showing spike timing-dependent plasticity,” J. Phys. D: Appl. Phys. 52 (2019) 49LT01 (11pp). [3]. Yangqi Huang, Weisheng Zhao et.al, “Magnetic-skyrmion based synaptic device,” IOP Nanotechnology, 28 08LT02, 2017 [4]. Jiang Nan, Weisheng Zhao et.al, “Efficient Magnetic Domain Nucleation and Domain Wall Motion with Voltage Control Magnetic Anisotropy Effect and Antiferromagnetic/Ferromagnetic Coupling,” IEEE TRANSACTIONS ON MAGNETICS, Vol. 55, No. 7, JULY 2019. [5]. Jialin Cai et.al, “Voltage-Controlled Spintronic Stochastic Neuron Based on a Magnetic Tunnel Junction,” Phys. Rev. Applied 11, 034015, 6 March 2019.[6]. M. Bauer, R. Alexis, G. Atwood, B. Baltar, A. Fazio, K. Frary, M. Hensel, M. Ishac, J. Javanifard, M. Landgraf, D. Leak, K. Loe, D. Mills, P. Ruby, R. Rozman, S. Sweha, S. Talreja, and K. Wojciechowski, in IEEE International Solid-State Circuits Conference Digest of Technical Papers (1995), p. 132. [7]. T. S. Jung, Y. J. Choi, K. D. Suh, B. H. Suh, J. K. Kim, Y. H. Lim, Y. N. Koh, J. W. Park, K. J. Lee, J. H. Park, K. T. Park, J. R. Kim, J. H. Yi, and H. K. Lim, IEEE J. Solid-State Circuits 31, 1575 (1996). [8]. X. H. Lou, Z. Gao, D. V. Dimitrov, and M. X. Tang, Appl. Phys. Lett. 93, 242502 (2008). [9]. N. Locatelli, V. Cros, and J. Grollier, Nat. Mater. 13, 11 (2014). [10]. F. N. Tan et.al, “High Velocity Domain wall propagation using voltage control magnetic anisotropy,” Scientific Reports, 2019) 9:7369.

**Images:**

![](https://assets.underline.io/uploads/markdown_image/1/image/11865266d897f1c54b8d59f2476fef1b.jpg)
Fig. 1: (a) Neuromorphic circuit implementation of spike time dependent plasticity-based pattern recognition. 
The circuit consists of voltage-controlled magnetic anisotropy post neuron and stochastic 3T-MTJ synapse with domain wall driven by SOT. (b) 3T-MTJ stochastic synapse schematic with small notches. (c) Cross-bar architecture implementation of the circuit.

![](https://assets.underline.io/uploads/markdown_image/1/image/568d7875203ba7028c8a7d9c8698bfa0.jpg)
Fig. 2: (a) Domain wall position in neuron for pinning ON and pinning OFF controlled by gate voltage. (b) Domain wall velocity for the pinning and non-pinning case. (c) Magnetization in the free layer MTJ synapse showing stochastic behavior with a positive current pulse. (d) MuMAG – NEGF coupled computed resistance evolution of the synapse.

 Domain Wall Motion based Neuromorphic Computing with Voltage Controlled Spin Neuron and Stochastic Magnetic Tunnel Junction Synapse

Physical reservoir computing using spin-torque oscillator (STO) has been attracting much attention after the demonstration of highly accurate human voice recognition [1]. To enhance the computational ability of STOs, there are two approaches: an array of STOs, and delayed feedback system. In the STO array, theoretical study was performed on 10-1000 STOs with interaction to obtain highly nonlinear dynamics, resulting in enhancement of computational ability that is competitive to quantum reservoir [4]. A similar highly nonlinear dynamics can be generated by using the feedback system, where outputted power from STO was reinjected in the form of electrical current or magnetic field to induce additional torque. By tuning the additional torque, complex magnetization dynamics such as amplitude modulation [5] and chaos [6,7] can be induced. These nonlinear dynamics are expected to enhance the computational ability even without interactions between STOs. Recent study showed that tunable nonlinear dynamics and delayed response by the feedback effect are adaptable for specific temporal pattern recognition task [7]. However, quantitative comparison of computational ability between the array system and the feedback system has not been investigated. In this work, we quantitatively evaluate the computational ability of STO with delayed-feedback circuit by performing a general task. The computational ability was evaluated by performing second order nonlinear autoregressive moving average (NARMA2) task [3,4]. The NARMA2 task aims to reproduce target outputs, defined by nonlinear transformation of the random inputs, from the amplitude of the STO. Figure 1 is a schematic illustration of vortex-type STO [2] with a delayed feedback circuit. To excite an auto-oscillation of the vortex core, perpendicular magnetic field of about 0.5 T and bias voltage of 350 mV were applied. In addition, we superposed random-pulse voltage with the width of 25 ns on the bias voltage. The random pulse was based on the uniform random number in the range of [-1,1] with the magnitude of 75 mV; therefore, the total input voltage varied from 275 to 425 mV. The outputted signal of the STO was amplified and converted to the RF magnetic field, which was applied to the STO with time delay of about 30 ns. The strength of the feedback signal was controlled by a tunable attenuator. Figure 2(a) shows power spectral density of the STO with delayed feedback as a function of attenuation. In the large attenuation region, the power spectra show a broad peak. With decreasing the attenuation, sharp multiple peaks appeared in the spectra below 30 dB. Such a split of spectrum is often observed when the feedback effect induces the amplitude modulation [5,6]. Figure 2(b) summarizes the dependence of normalized mean square error (NMSE) of NARMA2 task on the attenuation. It is clearly shown that NMSE can be minimized at an attenuation of 34 dB, below which the sharp multi peak structure appeared in the spectra. The observed minimum NMSE (6.4×10-6) is comparable to that theoretically obtained in the array of 16-coupled STOs [3]. This result indicates that single STO with delayed-feedback circuit can compete with the STO array. This work is supported by NEDO.

**References:**

[1] J. Torrejon et al., Nature 547, 428 (2017). [2] S. Tsunegi et al., Jpn. J. Appl. Phys. 57, 120307 (2018). [3] T. Kanao et al., Phys. Rev. Applied 12, 024052 (2019). [4] K. Fujii et al., Phys. Rev. Applied 8, 024030 (2017). [5] D. Kumar et al., Sci. Rep. 6, 30747 (2016). [6] T. Taniguchi et al., Phys. Rev. B 100, 174425 (2019). [7] J. Williame et al., Appl. Phys. Lett. 114, 232405 (2019). [8] M. Riou et al., Phys. Rev. Applied 12, 024049 (2019).

**Images:**

![](https://assets.underline.io/uploads/markdown_image/1/image/d49dc6b3239e7837d4a4db73ab12d4ba.jpg)
Schematic illustration of spin-torque oscillator (STO) with delayed-feedback circuit. Auto-oscillation is driven by random pulse voltage Vin(t). Power of the feedback is controlled by the rate of attenuator rAtt. Reservoir computing is performed using time-series amplitude Aout(t) obtained by the Hilbert transformation of output voltage Vout(t).

![](https://assets.underline.io/uploads/markdown_image/1/image/ae2ed6bb6522935b9ce2976e571c4ff4.jpg)
(a) Power spectral density (PSD) of STO with the delayed feedback circuit as a function of attenuation rate rAtt. (b) Normalized mean-squared error (NMSE) of second order nonlinear autoregressive moving average (NARMA2) task against attenuation rate. NMSE shows a minimum value (6.4×10-6) at rAtt = 34 dB.

 Enhanced Computational Ability of Spin Torque Oscillator with Delayed-Feedback Circuit for Physical Reservoir Computing

Classifying or predicting complex time-dependent signals (e.g. speech, financial data, the weather) is a challenging computational task. Reservoir computing (RC) is an efficient neuromorphic computing approach that is ideally suited to such tasks and is typically implemented in software using a recurrent neural network (RNN) with fixed synaptic weights (the reservoir) connected to a single, trainable readout layer. However, more efficient implementations of RC are possible if the software RNN reservoir is substituted with a physical system with the correct properties, such as non-linear response to input signals and inherent memory, leading to a readily deployable, hardware-based neuromorphic computing platform [1]. In this work, we propose a novel approach to RC where the dynamics of a single magnetic domain wall (DW) trapped between two defect sites in a nanostrip acts as a hardware-based reservoir. We demonstrate how such a device, with dimensions smaller than 1 µm, is capable of performing complex data analysis tasks, such as speech recognition. We have modelled a Ni nanowire with two anti-notches (shown in the inset of Fig. 1(a)) using both a simple 1D model[2] and micromagnetic simulations. Both models show the DW exhibit complex oscillatory dynamics similar to the Duffing oscillator, thus giving highly non-linear responses to applied magnetic fields (Fig. 1(a)). We exploit the DW dynamics for RC by using an applied field to inject time-multiplexed input signals into the reservoir and show how this approach allows the device to perform classification tasks. We have explored how the regime of applied fields affects the accuracy of sine and square wave classification, showing that the best recognition rate is obtained at the edge of a chaotic regime of oscillation (Fig. 1(b)). We have also demonstrated that the same approach can be used for more complex tasks, such as spoken digit recognition and handwritten digits recognition. Our work opens a new perspective for neuromorphic computing in nanomagnetic hardware.

**References:**

[1] M. Riou et al., Phys. Review Applied,2019 [2] A. Pivano et al., Phys Review B, 2016

**Images:**

![](https://assets.underline.io/uploads/markdown_image/1/image/947f0ad21dbe4e988fab9e7c037ab764.jpg)
Fig. 1: Simulations of Ni nanostrip at 500 MHz. Figure (a) shows a typical bifurcation diagram computed with 1D model at 500MHz and the inset of Fig. 1(a) illustrates the typical geometry of a Ni nano-strip with two anti-nothces separated by 350 nm. Figure (b) shows the recognition rate for the sine and square task computed with 1D model of a domain wall in a Ni nanostrip. The input field is oscillatory comprising two terms: H-the constant offset of the field and ΔH-the amplitude of the stimulus. The inset shows a sequence of inputs and its corresponding output calculated during the classification process using 1D model.

Downloads

Next from INTERMAG 2021

Running Deep Neural Networks on Superparamagnetic Fluctuations

Similar lecture

Time domain Computing for Boolean Logic using STT

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from INTERMAG 2021

Running Deep Neural Networks on Superparamagnetic Fluctuations

Similar lecture

Time domain Computing for Boolean Logic using STT

Downloads