Embodiments of the invention relate generally to random noise generation, and more specifically to systems that generate random noise using nanoscale magnetic devices.
The generation of truly random noise is of interest for a number of applications, such as random number generation for encryption, sampling or modeling, and more recently for neuromorphic computing using spiking schemes.
Magnetic tunnel junctions (MTJs) are nanoscale (i.e., in the range 10 s to 100 s of nanometers) magnetic devices that be integrated with ICs and CMOS circuits and have been proposed as a hardware-based approach to generating random noise. MTJs are used in a wide range of applications including magnetic sensing, hard disk drives, and magnetic random access memory (MRAM). A MTJ comprises two thin ferromagnetic (FM) layers separated by thin insulator layer, which is typically an oxide like MgO or a nitride. One of the FM layers is made so that the direction of its moment is difficult to change or is held “fixed”, usually by coupling it to an antiferromagnet. The other FM layer is “free” to have the direction of its moment changed. This allows for well-defined states due to the magnetoresistance (MR) effect that allows detection of the orientation of the free layer's moment with respect to the fixed layer's moment based on the device resistance. While the orientation of the moment of the free layer can be changed using a magnetic field, more recently electric currents have been used to manipulate the orientation of the magnetic moment of the MTJ's free layer through an effect known as spin transfer torque (STT).
STT involves the interaction of a spin-polarized current (i.e., a current that has some large fraction of electrons with spins in the exact same direction) with a FM layer. Any one electron in the current will either pass through the FM layer or reflect off it. Quantum mechanics dictates that the probability is high that these interactions cause the electron spin to be oriented preferentially parallel or antiparallel to the direction of the FM layer's moment for transmitted and reflected electrons, respectively. However, the angular momentum lost by the rotation of the electron's spin must be conserved and therefore becomes a torque acting upon the moment of the FM layer. The torque from a single spin is small, but for a spin-polarized current on the order of only a mA there are approximately 1015 electrons interacting with the FM layer per second. So if the current is above a critical current, the net torque on the moment of the FM layer can be sufficient to induce the moment into an oscillatory mode, for example to form a spin-torque oscillator (STO). Even higher currents can flip the orientation of the moment of the FM layer. Thus in a STT-MRAM device the free layers in an array of MTJ memory cells can be switched between parallel and antiparallel orientations, thereby resulting in the individual MTJ cells representing either ones or zeros.
MTJs have been proposed for random number generation. U.S. Pat. No. 8,495,118 B2; U.S. Pat. No. 9,110,746 B2 and U.S. Pat. No. 10,078,496 B2 describe random number generators based on reading the state of a MTJ, i.e., detecting the orientation of the free layer moment, in response to thermal fluctuations or pulses of current greater than the critical current.
MTJs have also been proposed for generating random noise for use in neuromorphic computing. In neuromorphic computing, neurons in the brain can be mimicked by circuit elements that produce electrical spikes that can then be used as basis sets in training the system to perform certain tasks. Mizrahi et. al, “Neural-like computing with populations of superparamagnetic basis functions”, Nature Communications 9, 1533 (2018), has discussed the use of superparamagnetic MTJs (MTJs where thermal fluctuations destabilize the state of the MTJ) as spiking elements by taking advantage of the random noise that naturally occurs in these devices.
For MTJs that generate random noise the transition time from one state to the other (e.g., parallel to antiparallel moment orientation) is generally in the mid-millisecond to mid-microsecond range. Thus the transition rates are in the high kHz range. This relatively low frequency would severely slow certain random number applications and computation time for a neuromorphic computing application, as well as increase the power consumption of the system.
Embodiments of this invention use a nanoscale magnetic device with two free layers separated by a non-magnetic spacer layer that operates as a chaotic oscillator excited using the STT effect. The signal from the device is driven through a clipping circuit which cuts off the amplitude of the signal such that spikes are converted into square pulses with tens or hundreds of picosecond widths. The output of the clipping circuit can be passed to further processing circuitry for generating random numbers or random spikes.
For a complete understanding of the nature and advantages of the invention, reference should be made to the following detailed description taken together with the accompanying figures.
Embodiments of the system for generating random noise according to this invention use a nanoscale magnetic device that, unlike a MTJ, has two free layers separated by a non-magnetic spacer layer. The current perpendicularly through the free layers is high enough above the critical current so that the STT effect causes repeated flopping of the directions of the moments of each layer from one direction to an antiparallel direction. The reversal of the moment directions is a random process so the magnetic device can be operated as a chaotic oscillator.
A nanoscale magnetic device with two free layers separated by an antiparallel coupling (APC) layer has been proposed as a STO magnetic field sensor, as described in U.S. Pat. No. 8,164,861 B2, which is assigned to the same assignee as this application.
Micro-magnetic computer simulations were performed at zero temperature (no thermal effects on the device) to examine the moment dynamics that could be excited within a dual free layer structure with high current. Even at zero temperature, nondeterministic moment reversal for both free layers occurred for given simulation parameters, so the addition of thermal noise would only be expected to improve the random nature of the moment dynamics. The simulations modeled two FM layers (assumed to be CoFe with a saturation moment, Ms, of 1680 emu/cm2) patterned into 40 nm circles with thicknesses of 4 nm and 2 nm, respectively and a spacer layer separation between them of 2 nm. The Gilbert damping for each layer was 0.014 and the spin polarization was assumed to be 60%. Simulations were run with interlayer coupling energy varied between 0 and ±1×10−3 J/m2.
In
While in the preferred embodiment the non-magnetic spacer layer provides no interlayer coupling, spacer layers with small interlayer coupling may be used to modify the lifetimes of the two oscillator states. When using an APC spacer layer, the response was sinusoidal for low currents (−1 mA). However, with an APC spacer layer with −1×10−4 J/m2 coupling strength, the response was irregular at −2 mA. Also, for parallel coupling spacer layers with coupling strengths+5×10−5 J/m2 to +5×10−4 J/m2 and higher currents (−1.5 mA or greater) the response was irregular. Thus, APC or parallel coupling spacer layers will still result in the desired irregular moment dynamics if the current is high enough. Changing both current and applied field can provide some tuning of the dynamics to affect the ratio of transition rates between the parallel and antiparallel states of the moments. The current required to induce the chaotic oscillation of the free layer moments depends on several factors, including the saturation magnetization, FM layer volume, anisotropy field and Gilbert damping, and can be determined experimentally or by micro-magnetic computer simulation. If an external magnetic field is to be applied during operation of the device, it is applied in a direction parallel to the orientation of the moments when the device is in its quiescent state (i.e., no current from the current source). The magnetic device can be located adjacent a permanent magnet or electromagnet that applies the external magnetic field. An applied magnetic field during operation is optional; it is not a requirement as the modeling shows similar irregular dynamics at zero applied field.
While the FM materials in the modeling had a Ms of 1680 emu/cm2, magnetic materials with much smaller Ms can be used to lower the current required to achieve the irregular dynamics of the device. While the modeling here considered FMs with in-plane magnetic moments, the device can have FMs with their moments perpendicular to the plane using a phenomenon known as perpendicular magnetic anisotropy.
The magnetic device for random noise generation according to the above-described embodiments can be made small, e.g. in the 10-20 nm nanoscale range, and is compatible with CMOS circuits. The device has transition rates at frequencies up to 20-30 GHz, which are substantially higher than previous transition rate frequencies seen for randomly generated pulse trains. The device can be operated at relatively low power, and includes the ability to tune the properties of the noise.
While a magnetic device that can randomly transition between two states for generating random noise has been described above, the subsequent response can be shaped into a signal more like square pulses. That can be achieved using some type of clipping circuit using nonlinear elements, such as diodes or transistors, that takes the signal from the magnetic device and cuts the signal off at certain voltage levels. One example of a clipping circuit is shown in
While the invention has been particularly shown and described with reference to the preferred embodiments, it will be understood by those skilled in the art that various changes in form and detail may be made without departing from the spirit and scope of the invention.
Accordingly, the disclosed invention is to be considered merely as illustrative and limited in scope only as specified in the appended claims.
Number | Name | Date | Kind |
---|---|---|---|
7589600 | Dimitrov | Sep 2009 | B2 |
8164861 | Braganca et al. | Apr 2012 | B2 |
8495118 | Wang et al. | Jul 2013 | B2 |
8675309 | Braganca et al. | Mar 2014 | B2 |
9110746 | Zhu et al. | Aug 2015 | B2 |
9183858 | Okawa | Nov 2015 | B2 |
9459835 | Braganca et al. | Oct 2016 | B2 |
9577653 | Buhrman | Feb 2017 | B2 |
10078496 | Gupta et al. | Sep 2018 | B2 |
20060114619 | Sbiaa | Jun 2006 | A1 |
20100131578 | Fukushima et al. | May 2010 | A1 |
20100174766 | Weeks et al. | Jul 2010 | A1 |
20110141629 | Braganca et al. | Jun 2011 | A1 |
20120026784 | Kanai et al. | Feb 2012 | A1 |
20140067890 | Zhu et al. | Mar 2014 | A1 |
20140108478 | Lee et al. | Apr 2014 | A1 |
20180254773 | Yoshinari et al. | Sep 2018 | A1 |
20190250887 | Liu et al. | Aug 2019 | A1 |
Number | Date | Country |
---|---|---|
2011013901 | Jul 2009 | JP |
Entry |
---|
Mizrahi et al., “Neural-like computing with populations of superparamagnetic basis functions”, Nature Communications | (2018) 9:1533 | DOI: 10.1038/s41467-018-03963-w. |
Zhu, “Spin Transfer Induced Noise in CPP Read Heads”, IEEE Transactions on Magnetics, vol. 40, No. 1, Jan. 2004. |
Platt, Charles, “Really, Really Random Number Generator”, Makezine, May 1, 2015, 25 pages, downloaded from https://makezine.com/projects/really-really-random-number-generator/. |
Fukushima, Akio et al., “Spin dice: A scalable truly random number generator based on spintronics”, Applied Physics Express, Jul. 25, 2014, 5 pages, vol. 7, No. 8, Japan Society of Applied Physics. |
International Search Report and Written Opinion for counterpart PCT/US2019/063448, dated May 5, 2020, 12 pages, ILPO. |