betaFeedback

WorksPhysics and Astronomy

The Strong Sensitivity of the Characteristics of Binary Stochastic Neurons Employing Low Barrier Nanomagnets to Small Geometrical Variations

Abstract

Binary stochastic neurons (BSNs) are excellent activators for machine learning. An ideal platform for implementing them are low- or zero-energy-barrier nanomagnets (LBMs) possessing in-plane anisotropy (e.g. circular or slightly elliptical disks) whose fluctuating magnetization encodes a probabilistic (p-) bit. Here, we show that such a BSN's activation function, the pinning current (which pins the output to a particular binary state), and the response time - all exhibit strong sensitivity to very slight geometric variations in the LBM's cross section. A mere 1% change in the diameter of a circular nanomagnet in any arbitrary direction can alter the response time by a factor of ~4 at room temperature and a 10% change can alter the pinning current by a factor of ~2. All this causes large device-to-device variation which is detrimental to integration. We also show that the energy dissipation is lowered but the response time is increased by replacing a circular cross-section with a slightly elliptical one and then encoding the p-bit in the magnetization component along the major axis. Encoding the p-bit in the magnetization component along the minor axis has the opposite effect. The energy-delay-product, however, is relatively independent of whether the cross-section is a circle or an ellipse and which magnetization component encodes the p-bit in the case of the ellipse.

Introduction

Artificial intelligence (AI) platforms typically employ deep neural networks (DNNs). DNNs are also dominant in statistical machine learning which includes advanced computer vision (image detection, recognition and segmentation [1]), natural language processing [2], etc. There is now a strong desire to improve the energy and speed performances of these systems, leading to the interest in spiking neural networks (SNN). One class of SNNs can be activated by binary stochastic impulses [2, 3, 4] and can be implemented with binary stochastic neurons (BSN) that have two distinct output states (-1 and +1). A BSN will output either a -1 or a +1 (in the bipolar representation) with a probability determined by a specific function of the input that is provided to it. It is also an excellent representation for Ising machines used to solve combinatorial optimization problems [5].

Low (or zero) energy barrier nanomagnets (LBMs) possessing in-plane magnetic anisotropy are a natural (and popular) choice for BSNs and have been shown to be capable of acting as efficient hardware accelerators for machine learning [5, 6, 7]. The magnetization vectors of these LBMs, which are circular or nearly circular disks, fluctuate at room temperature owing to thermal perturbations and the fluctuating (stochastic) magnetization can produce, with suitable design, either a -1 or a +1 state [6, 7] whose probabilities are engineered by passing a spin polarized current through the LBM [6, 7]. The current, which is the input to the BSN, biases the output towards either -1 or +1 and the degree of bias can be tuned with the current. At any time step n+1n+1, the BSN will output a state given by the analytical expression

mi(n+1)=sgn⁡[tanh⁡(Ii(n))−ri](1)m_i(n+1)=\operatorname{sgn}\left[\tanh\left(I_i(n)\right)-r_i\right] \tag*{(1)}

where IiI_i is the dimensionless input spin current that biases the output either towards 0 or towards 1 (depending on its sign) and rir_i is a random number uniformly distributed between -1 and +1. Each BSN described by Equation (1) receives its input from a weighted sum of other BSNs obtained from a synapse Ii(n)=∑jWijmj(n)I_i(n)=\sum_j W_{ij}m_j(n). A wide variety of problems can be solved by properly designing or learning the weights WijW_{ij}, e.g. classification problems [8], constraint satisfaction problems [9], generation of cursive letters [10], etc.

There are three critical parameters for a BSN. The first is the “activation function” which represents the probability of outputting a given state (-1 or +1) as a function of the input. In the case of LBMs implementing BSNs, the input is a spin polarized current of magnitude I0I_0 and the activation function is the “⟨s⃗⋅m⃗⟩\langle\vec{s}\cdot\vec{m}\rangle versus I0I_0” relation where s⃗\vec{s} and m⃗\vec{m} are the unit vectors in the direction of the spin polarization of the current and the magnetization of the LBM, respectively [6, 7]. The angular bracket denotes time averaging. The quantity s⃗⋅m⃗\vec{s}\cdot\vec{m} is

This work was supported by the US National Science Foundation under grants CCF-2001255 and CCF-2006843.

Rahnuma Rahman and Supriyo Bandyopadhyay are with the Department of Electrical and Computer Engineering, Virginia Commonwealth University, Richmond, VA 23284, USA

Color versions of one or more of the figures in this article are available online at link text.

This is the accepted version of an article in IEEE Trans. Nanotechnol. © 2023 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. the component of the magnetization along the direction of the current’s spin polarization at any instant of time. A positive value will represent the bit 1 and a negative value the bit -1. If pp is the time averaged probability of outputting bit 1 for a given I0I_0, then p=[⟨s⃗⋅m⃗⟩+1]/2p = [\langle\vec{s} \cdot\vec{m}\rangle+ 1]/2. The second critical parameter is the “pinning current” which is the minimum value of the current for which p=1p = 1. The third critical parameter, which determines the response speed of BSNs implemented with LBMs, is the “correlation time” τc\tau_c which is the full-width-at-half maximum (FWHM) of the decay characteristic of the auto-correlation function of the magnetization fluctuations [6]. It is also called the “response time”. LBMs with in-plane anisotropy typically have a value of τc\tau_c that is about two orders of magnitude smaller than LBMs with perpendicular magnetic anisotropy [6] and are therefore favored because the BSN speed (i. e. the speed of hardware accelerators) will be higher if they are employed instead of ones with perpendicular anisotropy.

In this paper, we present four important results pertaining to BSNs implemented with LBMs possessing in-plane anisotropy. First, we show that the activation function (⟨s⃗⋅m⃗⟩(\langle\vec{s}\cdot\vec{m}\rangle vs I0)I_0) is quite sensitive to small changes in the shape of the nanomagnet’s cross section. This poses a serious problem since the activation function determines what the probability of producing a +1 output is for a given current. This probability must be controllable by the current and that controllability is lost if the activation function exhibits significant device to device variability due to the strong sensitivity to small shape variations.

Second, we show that the pinning current, which is the current required to pin the output to either +1 or -1 with 100% probability also depends on the cross-sectional shape.

Third, we show that the auto-correlation function of the magnetization fluctuation and its FWHM (i. e. τc\tau_c) can change significantly depending on whether the LBM is perfectly circular or slightly elliptical. We show this with the example of a perfectly circular LBM of diameter 100 nm and a very slightly elliptical LBM with major axis of 100 nm and minor axis of 99 nm. Their τc\tau_c-s differ by a factor of 3-4, even though their lateral dimensions in just one direction differ by a mere 1%. The circular nanomagnet has the longer (shorter) correlation time if s⃗\vec{s} is along the minor (major) axis of the elliptical nanomagnet. The quantity τc\tau_c is proportional to the “response time” of the BSN. The significance of the response time is that a BSN does not begin to produce an output of +1 with a probability pp (and -1 with a probability 1−p1-p) as soon as the bias current is set for pp. It takes a certain amount of time for that to happen, which is the response time. If this time varies significantly, then the circuits will become error-prone since one would not know exactly how long the output has to be sampled to obtain correct results.

Fourth, we show that “spin-inertia” [11], which causes short-lived mutation in magneto-dynamics and can sometimes have long-term effects [12], has no significant effect on the auto-correlation decay characteristics of BSNs implemented with LBMs and hence has no effect on τc\tau_c. It also has no perceptible effect on the activation function. That is reassuring since spin inertia cannot be easily controlled or engineered.

Theory

To study the dynamical characteristics of BSNs implemented with LBMs, such as a circular or slightly elliptical disk shown in Figure 1, we first simulate their magneto-dynamics in the presence of thermal noise by solving the stochastic Landau-Lifshitz-Gilbert (LLG) equation (sometimes referred to as the Landau-Lifshitz-Gilbert-Langevin equation) with the effect of spin-inertia included as a second order time derivative of the magnetization with a characteristic pre-factor τ\tau which is the time scale over which the angular momentum of magnetization relaxes [11]. It is also the time scale over which nutation takes place in the magneto-dynamics [11]. Additionally, we model the effect of a spin polarized dc current flowing normal to the cross-section of the nanomagnet with the spin polarized along either the xx-axis or the yy-axis (minor and major axes in the case of an elliptical nanomagnet) shown in Figure 1. This allows us to evaluate the activation function and the pinning current which will pin the magnetization along the direction of spin polarization which is equivalent to pinning it to either the state +1 or the state -1 if we assume that magnetization in the direction of spin polarization encodes +1 and in the opposite direction encodes -1. Pinning the magnetization to the minor axis (hard axis) with a current will be harder than pinning it to the major axis (easy axis) since the easy axis orientation corresponds to the potential energy minimum and the hard axis orientation corresponds to the potential energy maximum in an elliptical nanomagnet. Therefore, the pinning current will be larger in the former case.

Schematic of an elliptical nanomagnet, current, spin polarization, and coordinate axes

Figure 1. A nanomagnet shaped like an elliptical disk with the major axis along the yy-axis and minor axis along the xx-axis.

The stochastic LLG equation including the effect of spin inertia is [12]

dm⃗(t)dt=−∣γ∣[m⃗(t)×H⃗(t)−α∣γ∣(dm⃗(t)dt+τd2m⃗(t)dt2)]+am⃗(t)×[ξj⃗μBqMsΩ×m⃗(t)]+bξj⃗μBqMsΩ×m⃗(t)(2)\frac{d\vec{m}(t)}{dt}=-|\gamma|\left[\vec{m}(t)\times\vec{H}(t)-\frac{\alpha}{|\gamma|}\left(\frac{d\vec{m}(t)}{dt}+\tau\frac{d^2\vec{m}(t)}{dt^2}\right)\right]+a\vec{m}(t)\times\left[\frac{\xi\vec{j}\mu_B}{qM_s\Omega}\times\vec{m}(t)\right]+b\frac{\xi\vec{j}\mu_B}{qM_s\Omega}\times\vec{m}(t) \tag*{(2)}

where H⃗(t)=H⃗thermal(t)+H⃗demag(t)\vec{H}(t)=\vec{H}_{\mathrm{thermal}}(t)+\vec{H}_{\mathrm{demag}}(t)

H⃗thermal(t)=2αkTγ(1+α2)μ0MsΩΔt[G(0,1)x(t)x^+G(0,1)y(t)y^+G(0,1)z(t)z^]\vec{H}_{\mathrm{thermal}}(t)=\sqrt{\frac{2\alpha kT}{\gamma(1+\alpha^2)\mu_0 M_s \Omega\Delta t}}\left[G_{(0,1)}^x(t)\hat{x}+G_{(0,1)}^y(t)\hat{y}+G_{(0,1)}^z(t)\hat{z}\right]
H⃗demag(t)=−MsNd−xxmx(t)x^−MsNd−yymy(t)y^\vec{H}_{\mathrm{demag}}(t)=-M_sN_{d-xx}m_x(t)\hat{x}-M_sN_{d-yy}m_y(t)\hat{y}
−MsNd−zzmx(t)z^-M_sN_{d-zz}m_x(t)\hat{z}

and

I⃗s=I0x^(for spin polarized in the x-direction)\vec{I}_s=I_0\hat{x}\quad\text{(for spin polarized in the }x\text{-direction)}
I⃗s=I0y^(for spin polarized in the y-direction)\vec{I}_s=I_0\hat{y}\quad\text{(for spin polarized in the }y\text{-direction)}

The demagnetization factors are given by [15]

Nd−xx=π4(dl)[1+54(l−wl)+2116(l−wl)2]N_{d-xx}=\frac{\pi}{4}\left(\frac{d}{l}\right)\left[1+\frac{5}{4}\left(\frac{l-w}{l}\right)+\frac{21}{16}\left(\frac{l-w}{l}\right)^2\right]
Nd−yy=π4(dl)[1−14(l−ww)−316(l−wl)2]N_{d-yy}=\frac{\pi}{4}\left(\frac{d}{l}\right)\left[1-\frac{1}{4}\left(\frac{l-w}{w}\right)-\frac{3}{16}\left(\frac{l-w}{l}\right)^2\right]
Nd−zz=1−Nd−xx−Nd−yyN_{d-zz}=1-N_{d-xx}-N_{d-yy}

Here m⃗(t)\vec{m}(t) is the magnetization vector normalized to the saturation magnetization of the magnetic material MsM_s, and mi(t)m_i(t) is its ii-th scalar component. The quantity H⃗(t)=H⃗demag+H⃗thermal\vec{H}(t)=\vec{H}_{\mathrm{demag}}+\vec{H}_{\mathrm{thermal}} where the first term is due to any shape anisotropy of the LBM and the second is due to thermal noise and is random. The magnetization vector precesses around H⃗(t)\vec{H}(t) and hence fluctuates randomly.

In Equation (2), the quantity γ\gamma is the gyromagnetic precession constant (2.21×1052.21\times10^5 rad-m/A-s), α\alpha is the Gilbert damping factor in the nanomagnet associated with damping of the precession (a material constant), and τ\tau is the relaxation time of the angular momentum associated with spin inertia (the timescale over which nutation takes place [11]). The magnitude of the charge current injected into the nanomagnet is I0I_0 and the degree of spin polarization of the current is ξ\xi. The spin in the current is polarized along the intended pinning direction, which is either the major axis of the nanomagnet (y-direction) or the minor axis (x-direction) for the ellipse in this study.

The last two terms in Equation (2) represent the Slonczewski and field-like torques, respectively, with μB\mu_B being the Bohr magneton. The relative strengths of these two torques are given by the quantities aa and bb, and following ref. [13], we assume a=1.0a=1.0 and b=0.3b=0.3. The value of τ\tau can range from a few fs to ∼100\sim100 ps in ferromagnets [14].

The demagnetizing field H⃗demag(t)\vec{H}_{\mathrm{demag}}(t) depends on the demagnetization coefficients Nd−xxN_{d-xx}, Nd−yyN_{d-yy} and Nd−zzN_{d-zz}, which, in turn, depend on the nanomagnet dimensions ll (major axis), ww (minor axis) and dd (thickness) as shown in Figure 1. In the expression for the thermal noise field term H⃗thermal(t)\vec{H}_{\mathrm{thermal}}(t), Ω\Omega is the volume of the nanomagnet [Ω=(π/4)l×w×d]\left[\Omega=(\pi/4)l\times w\times d\right], Δt\Delta t is the time step used in the simulation, μ0\mu_0 is the permeability of free space and the quantities G(0,1)iG_{(0,1)}^i (i=x,y,zi=x,y,z) are three statistically independent Gaussians with zero mean and unit standard deviation [16]. All simulations assume room temperature (T=300T=300 K). We also assume the following material parameters: Ms=8×105M_s=8\times10^5 A/m; α=0.1\alpha=0.1. This corresponds to a material like cobalt or Terfenol-D. The time step chosen is Δt=0.1\Delta t=0.1 ps and it was found in numerous previous simulations (e.g. [12, 13] references therein) that making the time step any smaller does not make any perceptible change (1% or more) in the magnetodynamics results.

The vector equation (2) is decomposed into three coupled scalar equations. When the spin in the injected current is polarized along the xx-direction, the coupled equations are:

dmx(t)dt=−∣γ∣[my(t)Hz(t)−mz(t)Hy(t)]+α[my(t)dmz(t)dt−mz(t)dmy(t)dt]+α[my(t)τd2mz(t)dt2−mz(t)τd2my(t)dt2]+a(ξI0μBqMsΩ)[my2(t)+mz2(t)]\begin{aligned} \frac{dm_x(t)}{dt}&=-|\gamma|\left[m_y(t)H_z(t)-m_z(t)H_y(t)\right]\\ &\quad+\alpha\left[m_y(t)\frac{dm_z(t)}{dt}-m_z(t)\frac{dm_y(t)}{dt}\right]\\ &\quad+\alpha\left[m_y(t)\tau\frac{d^2m_z(t)}{dt^2}-m_z(t)\tau\frac{d^2m_y(t)}{dt^2}\right]\\ &\quad+a\left(\frac{\xi I_0\mu_B}{qM_s\Omega}\right)\left[m_y^2(t)+m_z^2(t)\right] \end{aligned}
dmy(t)dt=−∣γ∣[mz(t)Hx(t)−mx(t)Hz(t)]+α[mz(t)dmx(t)dt−mx(t)dmz(t)dt]+α[mz(t)τd2mx(t)dt2−mx(t)τd2mz(t)dt2]−b(ξI0μBqMsΩ)mz(t)−a(ξI0μBqMsΩ)mx(t)my(t)\begin{aligned} \frac{dm_y(t)}{dt}&=-|\gamma|\left[m_z(t)H_x(t)-m_x(t)H_z(t)\right]\\ &\quad+\alpha\left[m_z(t)\frac{dm_x(t)}{dt}-m_x(t)\frac{dm_z(t)}{dt}\right]\\ &\quad+\alpha\left[m_z(t)\tau\frac{d^2m_x(t)}{dt^2}-m_x(t)\tau\frac{d^2m_z(t)}{dt^2}\right]\\ &\quad-b\left(\frac{\xi I_0\mu_B}{qM_s\Omega}\right)m_z(t)-a\left(\frac{\xi I_0\mu_B}{qM_s\Omega}\right)m_x(t)m_y(t) \end{aligned}
dmz(t)dt=−∣γ∣[mx(t)Hy(t)−my(t)Hx(t)]+α[mx(t)dmy(t)dt−my(t)dmx(t)dt]+α[mx(t)τd2my(t)dt2−my(t)τd2mx(t)dt2]+b(ξI0μBqMsΩ)my(t)−a(ξI0μBqMsΩ)mz(t)mx(t)(3)\begin{aligned} \frac{dm_z(t)}{dt}&=-|\gamma|\left[m_x(t)H_y(t)-m_y(t)H_x(t)\right]\\ &\quad+\alpha\left[m_x(t)\frac{dm_y(t)}{dt}-m_y(t)\frac{dm_x(t)}{dt}\right]\\ &\quad+\alpha\left[m_x(t)\tau\frac{d^2m_y(t)}{dt^2}-m_y(t)\tau\frac{d^2m_x(t)}{dt^2}\right]\\ &\quad+b\left(\frac{\xi I_0\mu_B}{qM_s\Omega}\right)m_y(t)-a\left(\frac{\xi I_0\mu_B}{qM_s\Omega}\right)m_z(t)m_x(t) \tag*{(3)} \end{aligned}

where Hi(t)H_i(t) (i=x,y,zi=x,y,z) is the ii-th component of the effective magnetic field at time tt.

When the spin in the injected current is polarized in the y-direction, the coupled equations change to

dmx(t)dt=−∣γ∣[my(t)Hz(t)−mz(t)Hy(t)]+α[my(t)dmz(t)dt−mz(t)dmy(t)dt]+α[my(t)τdmz2(t)dt2−mz(t)τdmy2(t)dt2]+b(ξI0μBqMsΩ)mz(t)−a(ξI0μBqMsΩ)mx(t)my(t)\frac{d m_x(t)}{dt}=-|\gamma|\left[m_y(t)H_z(t)-m_z(t)H_y(t)\right] +\alpha\left[m_y(t)\frac{d m_z(t)}{dt}-m_z(t)\frac{d m_y(t)}{dt}\right] +\alpha\left[m_y(t)\tau\frac{d m_z^2(t)}{dt^2}-m_z(t)\tau\frac{d m_y^2(t)}{dt^2}\right] +b\left(\frac{\xi I_0\mu_B}{qM_s\Omega}\right)m_z(t)-a\left(\frac{\xi I_0\mu_B}{qM_s\Omega}\right)m_x(t)m_y(t)
dmy(t)dt=−∣γ∣[mz(t)Hx(t)−mx(t)Hz(t)]+α[mz(t)dmx(t)dt−mx(t)dmz(t)dt]+α[mz(t)τdmx2(t)dt2−mx(t)τdmz2(t)dt2]+a(ξI0μBqMsΩ)[mx2(t)+mz2(t)]\frac{d m_y(t)}{dt}=-|\gamma|\left[m_z(t)H_x(t)-m_x(t)H_z(t)\right] +\alpha\left[m_z(t)\frac{d m_x(t)}{dt}-m_x(t)\frac{d m_z(t)}{dt}\right] +\alpha\left[m_z(t)\tau\frac{d m_x^2(t)}{dt^2}-m_x(t)\tau\frac{d m_z^2(t)}{dt^2}\right] +a\left(\frac{\xi I_0\mu_B}{qM_s\Omega}\right)\left[m_x^2(t)+m_z^2(t)\right]
dmz(t)dt=−∣γ∣[mx(t)Hy(t)−my(t)Hx(t)]+α[mx(t)dmy(t)dt−my(t)dmx(t)dt]+α[mx(t)τdmy2(t)dt2−my(t)τdmx2(t)dt2]−b(ξI0μBqMsΩ)mx(t)−a(ξI0μBqMsΩ)mz(t)my(t)(4)\frac{d m_z(t)}{dt}=-|\gamma|\left[m_x(t)H_y(t)-m_y(t)H_x(t)\right] +\alpha\left[m_x(t)\frac{d m_y(t)}{dt}-m_y(t)\frac{d m_x(t)}{dt}\right] +\alpha\left[m_x(t)\tau\frac{d m_y^2(t)}{dt^2}-m_y(t)\tau\frac{d m_x^2(t)}{dt^2}\right] -b\left(\frac{\xi I_0\mu_B}{qM_s\Omega}\right)m_x(t)-a\left(\frac{\xi I_0\mu_B}{qM_s\Omega}\right)m_z(t)m_y(t) \tag*{(4)}

These three coupled equations [3(a) or 3(b)] are solved using MATLAB to find the magnetization components mx(t),my(t),mz(t)m_x(t),m_y(t),m_z(t) at any time tt.

Simulation Results

Pinning current and activation function of the BSN

Equation (2) contains a random term associated with the thermal noise field H⃗thermal\vec{H}_{\mathrm{thermal}} and hence the magnetization versus time trajectories [mi(t) vs t]; i=x,y,z\left[m_i(t)\ \mathrm{vs}\ t\right];\ i=x,y,z calculated from it will differ slightly from run to run. Consequently, all results reported here are ensemble averaged over 1000 runs.

In order to determine the activation function and pinning current of the BSN, we first calculate the quantity ⟨s⃗⋅m⃗⟩\langle\vec{s}\mathbin{\cdot}\vec{m}\rangle as a function of the current amplitude I0I_0 after steady state is reached. Here s⃗\vec{s} is the unit vector along the direction of the spin polarization in the injected current. When the spin-polarization is along the yy-direction, ⟨s⃗⋅m⃗⟩=⟨my⟩\langle\vec{s}\mathbin{\cdot}\vec{m}\rangle=\langle m_y\rangle. Similarly, when the spin-polarization is along the xx-direction, ⟨s⃗⋅m⃗⟩=⟨mx⟩\langle\vec{s}\mathbin{\cdot}\vec{m}\rangle=\langle m_x\rangle. In Figs. 2(a) and 2(b), we show plots of mx(t)m_x(t) vs tt and my(t)m_y(t) vs. tt for an elliptical nanomagnet (major axis = 100 nm, minor axis = 90 nm and thickness = 2 nm) for I0=2I_0=2 mA. The spin polarization in the injected current is in the +xx direction and +yy direction, respectively, for the two cases. The initial orientation of the magnetization is assumed to be always perpendicular to the axis of the current’s spin polarization. Note that the time to achieve pinning in the two cases are different with the latter being longer. This is because the major axis (y-axis) is the easy axis and the minor axis (x-axis) is the hard axis. Since it is easier to pin the magnetization along the easy axis and harder to pin it along the hard axis, the time to pin is shorter in the former case.

Magnetization component along minor axis

Figure 2a.

Magnetization component along major axis

Figure 2b.

Plots of magnetization components versus time for an elliptical nanomagnet

Figure 2. Plots of (a) mx(t)m_x(t) vs tt and (b) my(t)m_y(t) vs. tt for an elliptical nanomagnet (major axis = 100 nm, minor axis = 90 nm and thickness = 2 nm) for I0=2.0I_0=2.0 mA. The spin polarization in the injected current is in the +xx direction and +yy direction, respectively, for the two cases. The time to pin is the time taken to pin the magnetization in the orientation of the spin polarization. It has no relation to the response time or correlation time. The nutation time τ\tau was taken to be 10 ps.

The time it takes for mx(t),my(t),mz(t)m_x(t),m_y(t),m_z(t) to reach steady-state depends on the magnitude of I0I_0. Steady state is never reached if I0=0I_0=0. For the non-zero magnitudes of I0I_0 that we considered, steady state is always reached within 25 ns of turning on the current. Hence, we take the value of mx(t)m_x(t) or my(t)m_y(t) at the end of 25 ns for every value of I0I_0 that we consider and then average that over 1000 runs to calculate ⟨s⃗⋅m⃗⟩=⟨mx⟩\langle\vec{s}\cdot\vec{m}\rangle=\langle m_x\rangle and ⟨s⃗⋅m⃗⟩=⟨my⟩\langle\vec{s}\cdot\vec{m}\rangle=\langle m_y\rangle for that value of I0I_0. At steady-state, the system is ergodic and hence the time-average is equal to the ensemble-average.

In Figure 3, we show the results of ⟨s⃗⋅m⃗⟩\langle\vec{s}\cdot\vec{m}\rangle versus I0I_0 (i.e. the activation function) for three different values of the nutation duration τ\tau associated with spin inertia, and find no significant τ\tau-dependence. We consider three cases: (a) elliptical nanomagnet with the injected current spin-polarized along the major axis, (b) same nanomagnet with the injected current spin-polarized along the minor axis, and (c) perfectly circular nanomagnet. In all cases, the magnetization initially points perpendicular to the axis of spin polarization of the injected current (i.e., perpendicular to the intended direction of pinning). The positive and negative values of the current represent opposite (mutually antiparallel) spin polarizations. The curve has the expected tanh⁡(I0/Ic)\tanh(I_0/I_c) [Ic=I_c= constant] behavior [5, 6]. The two values of I0I_0 where ⟨s⃗⋅m⃗⟩\langle\vec{s}\cdot\vec{m}\rangle reaches the value ±1\pm1 are the two pinning currents for pinning the magnetization along the direction of spin polarization and making the probability of bit 1 either 0% or 100%. These results reveal that spin inertia does not affect the activation function or the pinning current perceptibly, which is good happenstance since spin inertia cannot be easily engineered.

Plots of the steady-state value of the magnetization as a function of current for elliptical and circular nanomagnets

Figure 3. Plot of the steady-state value of ⟨s⃗⋅m⃗⟩\langle\vec{s}\cdot\vec{m}\rangle as a function of current for different values of the parameter τ\tau representing the duration of nutation caused by spin inertia. (a) Elliptical nanomagnet (major axis = 100 nm, minor axis = 90 nm and thickness = 2 nm) with an in-plane shape anisotropy energy barrier of 4.6 kT at room temperature and the current is spin-polarized along the y-axis or major axis (easy axis), (b) same nanomagnet when the current is spin-polarized along the x-axis or minor axis (hard axis), and (c) circular nanomagnet of diameter 100 nm and thickness 2 nm with no in-plane shape anisotropy energy barrier. The initial orientation of the magnetization is always orthogonal to the direction of spin polarization of the current.

However, there is clearly some dependence on whether the nanomagnet is circular or elliptical and whether the spin polarization of the injected current is along the major axis or minor axis in the case of the elliptical nanomagnet. In fact, a 10% variation in a principal axis dimension (minor axis varying from 100 nm to 90 nm while the major axis remains invariant at 100 nm, i.e. circular to slightly elliptical) decreases the pinning current from 1.0 mA to 0.5 mA when the current is spin-polarized along the major (easy) axis and increases the pinning current from 1.0 mA to 2.0 mA when the current is spin-polarized along the minor (hard) axis. This happens because it is easier to pin the magnetization along the easy axis and harder to pin it along the hard axis, whereas the circular nanomagnet has no distinction between easy and hard axes. Therefore, we expect that the pinning current will be least in an elliptical nanomagnet with spin polarization along the major axis and most in the same nanomagnet with the spin polarization along the minor axis, with the circular nanomagnet being in between. That is exactly what we observe.

There are two lessons learned from this study. The first is that slight shape changes (going from circular to slightly elliptical) can change the activation function and the pinning current significantly. This sensitivity can lead to large device to device variation which is deleterious for large-scale integration. The second is that the minimum pinning current for pinning the probability of bit 1 to either 0% or 100% (and hence minimum energy dissipation) is obtained if we use slightly elliptical nanomagnets and encode the probabilistic p-bit in the component of the magnetization along the major axis. Thus, the optimum choice for minimum energy dissipation is a slightly elliptical nanomagnet with the p-bit encoded in the magnetization component along the major axis.

Response time of the BSN

Next, we calculate the response time of the BSN. For this purpose, we first assume that the p-bit is encoded in the magnetization component along the minor axis, which we call the xx-axis, and compute its auto-correlation function defined as Cx(t′)=∫0∞mx(t)mx(t+t′) dtC_x(t')=\int_{0}^{\infty}m_x(t)m_x(t+t')\,dt in the absence of any spin-polarized current, starting with two initial conditions: the magnetization initially pointing approximately along the +x+x-axis and along the +y+y-axis. These two initial conditions are, respectively, mx(0)=0.995;my(0)=0.095;mz(0)=0.031m_x(0)=0.995;m_y(0)=0.095;m_z(0)=0.031 and mx(0)=0.095;my(0)=0.995;mz(0)=0.031m_x(0)=0.095;m_y(0)=0.995;m_z(0)=0.031. Starting with either initial condition, we follow the time trajectory of the magnetization component mx(t)m_x(t) [obtained by solving Equation (2)] to obtain the auto-correlation function Cx(t′)C_x(t') versus t′t' as mx(t)m_x(t) fluctuates between -1 and +1 owing to thermal noise. Because the fluctuation history is slightly different from one run to another owing to the randomness of thermal noise, we average Cx(t′)C_x(t') versus t′t' over 1000 runs to obtain the final Cx(t′)C_x(t') versus t′t' plots.

The purpose of this calculation is to find the FWHM of the auto-correlation function, which is the correlation time τc\tau_c that determines the response time of BSNs. In this case, the BSN should output bit 1 and bit -1 with equal probability (50% each) and we wish to estimate the time that will elapse before the BSN begins to output bits with this probability. This is the response time if the p-bit is encoded in the magnetization component parallel to the minor axis in an elliptical nanomagnet. Our purpose was to study if slight changes in the nanomagnet shape (from perfectly circular to slightly elliptical) can have a significant effect on the correlation time τc\tau_c and hence the BSN response time.

We also obtain the auto-correlation function of the magnetization component along the yy-axis which would be the major axis of the elliptical nanomagnet. This is defined as Cy(t′)=∫0∞my(t)my(t+t′) dtC_y(t')=\int_{0}^{\infty}m_y(t)m_y(t+t')\,dt. This will yield the response time if the p-bit is encoded in the magnetization component parallel to the major axis in an elliptical nanomagnet.

In Figure 4, we plot the calculated auto-correlation functions as a function of the delay [i.e. Cx(t′)C_x(t') versus t′t' and Cy(t′)C_y(t') versus t′t'] for a slightly elliptical nanomagnet (major axis = 100 nm, minor axis = 99 nm) and a perfectly circular nanomagnet (diameter = 100 nm), both with the same thickness of 6 nm. The first has an in-plane shape anisotropy barrier of 4 kTkT at room temperature and the second has no such barrier. In the case of the slightly elliptical nanomagnet, we plot the auto-correlation functions for two cases: initial orientation of the magnetization is along the minor axis (xx-axis) and initial orientation is along the major axis (yy-axis).

Auto-correlation functions of magnetization fluctuations for slightly elliptical and circular nanomagnets

Figure 4. Auto-correlation function of the fluctuations in the magnetization component along the xx-axis (minor axis of the elliptical nanomagnet) versus delay shown for three different nutation times τ\tau for (a) a slightly elliptical (100 nm, 99 nm) nanomagnet with initial orientation of the magnetization along the major axis, (b) the same slightly elliptical nanomagnet with initial orientation along the minor axis, and (c) a perfectly circular nanomagnet (diameter = 100 nm). Auto-correlation function of the fluctuations in the magnetization component along the yy-axis (major axis) of the slightly elliptical nanomagnet versus delay with (d) initial orientation along the major axis and (e) initial orientation along the minor axis.

In Table I, we list the calculated full-width-at-half-maximum (FWHM) of the auto-correlation functions (or the correlation time τc\tau_c) for the five cases. These times determine the response times of the corresponding BSNs.

τ=0\tau= 0 psτ=10\tau= 10 psτ=100\tau= 100 ps
Circular LBM5.8 ns5.6 ns5.7 ns
mxm_x in slightly elliptical nanomagnet. Initial orientation along major axis1.5 ns1.5 ns1.7 ns
mxm_x in slightly elliptical nanomagnet. Initial orientation along minor axis1.6 ns1.6 ns1.7 ns
mym_y in slightly elliptical nanomagnet. Initial orientation along major axis20.25 ns21.07 ns20.24 ns
mym_y in slightly elliptical nanomagnet. Initial orientation along minor axis20.79 ns20.42 ns20.79 ns

Table I. TABLE I. THE CORRELATION TIME τc\tau_c IN CIRCULAR AND SLIGHTLY ELLIPTICAL LBMS FOR THREE DIFFERENT SPIN INERTIA RELAXATION TIMES τ\tau

From Table I, we find that spin inertia (i. e. the nutation time τ\tau) does not cause any significant change in the correlation time, but even a slight change in geometry (1% change in the diameter of the circular nanomagnet in any arbitrary direction to make it slightly elliptical)) can decrease the correlation time associated with fluctuation of the magnetization component along the minor axis by a factor of nearly 4 and increase the correlation time associated with fluctuation of the magnetization component along the major axis by a factor of nearly 4 as well. The opposite behaviors of the two magnetization components of course have to do with the fact that the major axis is the easy axis and the minor axis is the hard axis. The magnetization prefers to loiter around the easy axis and shun the hard axis. That is why the auto-correlation function of the fluctuation of the magnetization component along the easy axis decays much more slowly than that along the hard axis. What is surprising is that a very slight ellipticity (major axis = 100 nm and minor axis = 99 nm) can make so much difference. When probabilistic (p-) bits are encoded in the randomly fluctuating magnetization component of a slightly elliptical LBM [5, 6, 7], it makes a very significant difference as to whether the component along the major axis or the minor axis is chosen to encode the p-bit, since they have very different correlation times (and hence response times) that can differ by more than an order of magnitude.

Earlier, we posited that if we wish to reduce energy dissipation, then we should encode the p-bit in the magnetization component along the major axis of a slightly elliptical LBM. However, that also makes the response much slower. There is an unavoidable trade-off between energy dissipation and speed in most information processing systems and that trend holds for this system as well.

We also notice that the initial magnetization orientation (whether it is pointing along the major axis or the minor axis) does not make any significant difference in the case of the slightly elliptical nanomagnet. One might have expected to see some dependence on the initial orientation because the major axis (yy-axis) is the “easy axis” and hence corresponds to a stable direction while the minor axis (xx-axis) is the “hard axis” and corresponds to an unstable direction. Therefore, if the magnetization is initially along the easy axis, we will expect it to resist straying from it while if it is initially along the hard axis, it should quickly leave that state. We do not see this effect since the decay time of the auto-correlation function exceeds 1 ns and over that long duration, the magnetization loses any memory of the initial orientation. That is why we see no effect of the initial state on the auto-correlation function decay characteristic or the correlation time τc\tau_c. Because the decay is relatively slow and occurs over a time scale exceeding 1 ns, which is much longer than the nutation time τ\tau, we also do not see any effect of spin inertia. Sometimes spin inertia can have a very significant effect in slow magneto-dynamics that lasts over time scales orders of magnitude longer than τ\tau [12], but that does not happen here.

The most interesting point to note is that the correlation time in the circular nanomagnet differs by a factor of ∼4\sim4 from that in the slightly elliptical nanomagnet (either 4×4\times larger or 4×4\times smaller depending on which magnetization component – along the major axis or along the minor axis – is chosen to encode the p-bit). The difference between the two nanomagnets is that the circular nanomagnet has no in-plane shape anisotropy energy barrier, whereas in the slightly elliptical nanomagnet, the in-plane shape anisotropy energy barrier is small (4 kT) but nonzero. Surprisingly, this small difference has a strong effect on the correlation time and hence the response speed of the BSN.

Conclusion

Low barrier nanomagnets (LBM) with in-plane anisotropy are a natural fit for BSNs, but unfortunately it turns out that their activation function, response time and pinning current are extremely sensitive to small geometric variations in their cross-section, which poses a serious problem for large-scale integration [17]. One possible way to mitigate this may be to introduce redundancy, i.e. encode a logical p-bit in multiple physical p-bits in the spirit of well-known quantum error correction schemes for qubits, but only at the cost of increased resources. A simpler approach has also been suggested [18]. We note that device-to-device variability is known to be a serious problem in memristor based neural networks [19, 20] which has prompted research in high entropy materials to counter this issue in memristors [21].

The activation function of a BSN plays a central role in the execution of algorithms, while the autocorrelation time τc\tau_c determines how long it takes for a BSN based algorithm to produce a new output in response to a new input [22]. In Appendix II, we shown that device-to-device variation in the activation function of a restricted Boltzmann machine can cause significant variability in the root mean square error in training. If the autocorrelation time becomes uncertain and varies significantly from device to device, the execution of algorithms with BSNs is affected because we would not know how long to sample the BSN states to get the correct result. If we terminate the algorithm prematurely, then an incorrect result will be produced, as was demonstrated dramatically in ref. [23] which used stochastic nanomagnets to implement inverse AND logic. Finally, any uncertainty in the pinning current will introduce a mismatch in the drive currents, similar to the problem caused by threshold variability in MOSFET circuits, which can be debilitating in high-performance circuits [24]. In order to mitigate that problem, one would be forced to use excessive currents and hence suffer excessive energy dissipation.

In the past, we showed that extended structural defects in LBMs (e.g. random thickness variations) introduce a similar variability in the response times (τc\tau_c) of BSNs [25]. Here, we show that even slight variation in the lateral geometry can have a serious deleterious effect. In future, we will extend this to study the effect of edge roughness, which could be equally or more harmful.

We also show that the energy dissipation can be reduced by choosing slightly elliptical LBMs and encoding the p-bit in the magnetization component along the major axis, but this also increases the response time. Conversely, the response time can be reduced by encoding the p-bit in the magnetization component along the minor axis, but that increases the energy dissipation.

We can view the energy-delay product as being proportional to the product of the response time and the square of the pinning current for pinning the probability of either bit to 100%. For the cases considered here, the energy delay product turns out to be relatively independent of whether we use a circular or a slightly elliptical nanomagnet. In the elliptical case, the energy-delay product is also relatively independent of which magnetization component – the one along the major axis or the one along the minor axis – is chosen to encode the p-bit.

Appendix I

Ref. [6] derived approximate analytical expressions for the correlation time and pinning currents. These expressions are

τc=8ln⁡(2)1γMsΩ4π∣Hzdemag∣kT(5)\tau_c = \sqrt{8\ln(2)}\frac{1}{\gamma}\sqrt{\frac{M_s\Omega}{4\pi\lvert H_z^{\mathrm{demag}}\rvert kT}} \tag*{(5)}
Ip=2qℏ2π4π∣Hzdemag∣MsΩkT(6)I_p = \frac{2q}{\hbar}\sqrt{\frac{2}{\pi}}\sqrt{4\pi\lvert H_z^{\mathrm{demag}}\rvert M_s\Omega kT} \tag*{(6)}

where qq is the electronic charge, ℏ\hbar is the reduced Planck constant and HzdemagH_z^{\mathrm{demag}} is the normal-to-plane demagnetizing field given by Hzdemag=−MsNd−zzH_z^{\mathrm{demag}}=-M_sN_{d-zz}. The other quantities were previously defined. These expressions were probably derived for the perfectly circular nanomagnet since no account is made of any shape anisotropy. It is informative to compare the analytical results with the numerically calculated results.

In our case, the quantity HzdemagH_z^{\mathrm{demag}} is 7.25×1057.25 \times10^5 A/m for the circular nanomagnet. The volume of the perfectly circular nanomagnet that we considered is 47,124 nm3^3. Using these results in Equation (5), we find that the correlation time τc\tau_c is 333 ns in the perfectly circular nanomagnet, which is about two orders of magnitude larger than what we found and report in Table I.

The value of the pinning current calculated from Equation (6) is about 2.9 A, which is three orders of magnitude larger than what we find in the circular nanomagnet. The expressions in Equations (A1) and (A2) are therefore overly pessimistic and predict much slower response and much higher current (hence much higher energy dissipation) than what is actually the case. It is not surprising that the analytical expressions are at odds with the numerical simulations because stochastic dynamics is rarely amenable to precise analytical treatment.

Appendix II

To highlight the effect of variability in the activation function on training a neural network, we used an open source simulator for a restricted Boltzmann machine [26] where we redefined the sigmoid activation function as sigmoid⁡(z)=tanh⁡(ηz)\operatorname{sigmoid}(z)=\tanh(\eta z) and varied η\eta between 0.8 and 1.0 to simulate the variability in the activation function. The results for the root mean square error (RMSE) in training versus number of iterations are shown in Figure A2 for various values of η\eta. One can see that there is significant variability in this plot if the activation function is varied.

Root mean square error in training a restricted Boltzmann machine as a function of the number of iterations

Figure A2. Root mean square error in training a restricted Boltzmann machine of ref. [26] as a function of the number of iterations when the activation function sigmoid⁡(z)=tanh⁡(ηz)\operatorname{sigmoid}(z)=\tanh(\eta z) is varied by varying η\eta from 0.8 to 1.0 in steps of 0.025. This illustrates the effect of device-to-device variations of the activation function on training a restricted Boltzmann machine and how sensitive the training is to device-to-device variations of the activation function.

References

  1. [1]Y. LeCun, Y. Bengio and G. Hinton, “Deep learning”, Nature, vol. 521, 436, 2015.DOI
  2. [2]D. Roy, I. Chakraborty and K. Roy, “Scaling deep spiking neural networks with binary stochastic activations”, Proc. 2019 IEEE International Conference on Cognitive Computing (ICCC), 50, 2019.DOI
  3. [3]M. Pfeiffer and T. Pfeil, “Deep learning with spiking neurons: Opportunities and challenges”, Frontiers in Neuroscience, vol. 12, 774, 2018.DOI
  4. [4]S. D. Pyle, R. Zand, S. Sheikhfaal and R. F. Demara, “Subthreshold spintronic stochastic spiking neural networks with probabilistic Hebbian plasticity and homeostasis”, vol. 5, 43-51 (2019).DOI
  5. [5]O. Hassan, S. Datta and K. Y. Camsari, “Quantitative evaluation of hardware binary stochastic neurons,” Phys. Rev. Appl., vol. 15, 064046, 2021.DOI
  6. [6]R. Faria, K. Y. Camsari and S. Datta, “Low-barrier nanomagnets as p-bits for spin logic,” IEEE Magn. Lett., vol 8, 4105305, 2017.DOI
  7. [7]O. Hassan, R. Faria, K. Y. Camsari, J. Z. Sun and S. Datta, “Low-barrier magnet design for efficient hardware binary stochastic neurons,” IEEE Magn. Lett., vol. 10, 4502805, 2019. We adopted the acronym “LBM” from this reference.DOI
  8. [8]U Cilingiroglu, "A purely capacitive synaptic matrix for fixed-weight neural networks", IEEE Trans. Circuits Syst., vol. 38, pp. 210-217, 1991.
  9. [9]G. A. Fonseca Guerra and S. B. Furber, “Using stochastic spiking neural networks on SpiNNaker to solve constrain satisfaction problems”, Front. Neurosci., vol. 11, Article 714, 2017.
  10. [10]A. Mizrahi, T. Hirtzlin, A. Fukushima, H. Kubota, S. Yuasa, J. Grollier and D. Querlioz, “Neural like computing with populations of superparamagnetic basis functions”, Nat. Commun., vol. 9, 1533, 2018.DOI
  11. [11]J-E Wegrowe and M. C. Ciornei, “Magnetization dynamics, gyromagnetic relation and inertial effects,” Am. J. Phys., vol. 80, 607, 2012.DOI
  12. [12]R. Rahman and S. Bandyopadhyay, “An observable effect of spin inertia in slow magneto-dynamics: Increase of the switching error rates in nanoscale ferromagnets”, J. Phys: Condens. Matter, vol. 33, 355801, 2021.DOI
  13. [13]K. Roy, S. Bandyopadhyay and J. Atulasimha, “Metastable state in a shape-anisotropic single-domain nanomagnet subjected to spin-transfer torque”, Appl. Phys. Lett., vol. 101, 162405, 2012.DOI
  14. [14]K. Neeraj et al., “Inertial spin dynamics in ferromagnets,” Nature Phys., vol. 17, 245-250, 2020.DOI
  15. [15]S. Chikazumi, Physics of Magnetism. New York, NY, USA: Wiley, 1964.
  16. [16]W. F. Brown,“Thermal fluctuations of a single domain particle,” Phys. Rev., vol. 130, 1677, 1963.
  17. [17]J. Kaiser, W. A. Borders, K Y Camsari, S. Fukami, H. Ohno and S. Datta, “Hardware-aware in-situ learning based on stochastic magnetic tunnel junctions”, Phys. Rev. Appl., vol. 17, 014016, 2022.
  18. [18]R. Rahman and S. Bandyopadhyay, “Robustness of binary stochastic neurons implemented with low-barrier nanomagnets made of dilute magnetic semiconductors”, IEEE Magn. Lett., vol. 13, 4505104, 2022.DOI
  19. [19]S. Kim, H-D Kim and S-J Choi, “Impact of synaptic device-to-device variations on classification accuracy in a binarized neural network”, Sci. Rep.,vol. 9, 15237, 2019.DOI
  20. [20]D. Kuzum, S. Yu and H-S Philip Wong, “Synaptic electronics: materials, devices and applications”, Nanotechnology, vol. 24, 382001, 2013.DOI
  21. [21]M. Ahn, Y. Park, S. H. Lee, S. Chae, J. Lee, J. T. Heron, E. Kioupakis, W. D. Lu and J. D. Phillips, “Memristors based on (Zr, Hf, Nb, Ta, Mo, W) high-entropy oxides”, Adv. Elec. Mater., vol. 7, 2001258, 2021.DOI
  22. [22]D. Khilwani, V. Moghe, S. Lashkare, V. Saraswat, P. Kumbhare, M. S. Baghini, S. Jandhyala, S. Subramoney and U. Ganguly, “PrxCa1−xMnO3 based stochastic neuron for Boltzmann machine to solve “maximum cut” problem”, APL Materials, vol. 7, 091112, 2019.
  23. [23]Y. Lv, R. P. Bloom and J-P Wang, “Experimental demonstration of probabilistic spin logic by magnetic tunnel junctions,” IEEE Magn. Lett., vol. 10, 4510905, 2019.
  24. [24]Siva G. Narendra, “Effect of MOSFET threshold voltage variation on high-performance circuits”, Ph.D. Thesis. Dept. of Electrical Engineering and Computer Science, Massachusetts Institute of Technology, Cambridge, MA, USA. 2002.
  25. [25]M. A. Abeed and S. Bandyopadhyay, “Low energy barrier nanomagnet design for binary stochastic neurons: Design challenges for real nanomagnets with fabrication defects”. IEEE Magn. Lett., vol. 10, 4504405, 2019.DOI
  26. [26]T. Berghout, “Restricted Boltzmann Machine” MathWorks, https://www.mathworks.com/matlabcentral/fileexchange/71212-restricted-boltzmann-machine

Paper details

vlnc:W_hnt74ane
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 3

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents