Tunable Random Telegraph Noise in Stable Perpendicular Magnetic Tunnel Junctions for Unconventional Computing

Technical Summary: Tunable Random Telegraph Noise in Stable Perpendicular Magnetic Tunnel Junctions for Unconventional Computing

Problem Statement
The development of hardware for unconventional computing architectures, particularly those addressing the memory bottlenecks of von Neumann systems and supporting probabilistic or neuromorphic models, requires new physical building blocks. While perpendicular magnetic tunnel junctions (pMTJs) are standard for non-volatile memory (STT-MRAM) and have been explored for in-memory computing, their intrinsic stochastic behavior is typically limited to two regimes: deterministic switching for memory or thermally driven superparamagnetic switching (sMTJs) for randomness. The latter relies on low energy barriers ($\Delta \lesssim 20$), which compromises thermal stability and non-volatility. There is a need for a device that combines the thermal stability of standard memory cells with the tunable stochasticity required for probabilistic computing, without sacrificing the ability to function as a deterministic switch.

Methodology
The authors propose and demonstrate an "actuated-stochastic magnetic tunnel junction" (A-sMTJ) concept. Unlike conventional sMTJs that rely on spontaneous thermal fluctuations, the A-sMTJ utilizes pMTJs with thermally stable free layers ($\Delta \gtrsim 20$) where spontaneous switching is suppressed. Stochastic behavior is induced externally by applying a periodic stream of nanosecond-duration voltage pulses of alternating polarity (AC pulses).

The experimental setup involves 40 nm diameter pMTJs with effective energy barriers of $\Delta \approx 26$ (for AP$\to$P) and $\Delta \approx 51$ (for P$\to$AP). The device operation follows a specific pulse sequence:

1. Stochastic Write Pulse S1: A positive voltage pulse promoting a transition from state "0" (Parallel) to "1" (Antiparallel) with probability $\Gamma_{01}$.
2. Stochastic Write Pulse S2: A negative voltage pulse promoting a transition from "1" to "0" with probability $\Gamma_{10}$.
3. Read Pulse: A low-amplitude pulse to detect the state without perturbing it.

The switching probabilities are measured by applying these sequences at 104 MHz and recording state transitions over $10^8$ cycles. The system is modeled as a sequence of independent events where the steady-state probability of reading a "1" is derived from the ratio of switching probabilities.

Key Contributions and Results

• Demonstration of Tunable Stochasticity: The paper establishes that stable pMTJs can be actuated to exhibit random telegraph noise (RTN) with fully controllable fluctuation rates and probability biases.
• Wide Tunability Range: The fluctuation rate is tunable over more than two orders of magnitude. Average state dwell times ($\tau_0$ and $\tau_1$) were observed to vary from 29 ns to greater than 2.3 $\mu$s by adjusting the amplitude of the driving pulses.
• Poissonian Statistics: The temporal dynamics of the device were shown to be consistent with a Poisson process. The dwell-time histograms exhibit exponential decay, and the autocorrelation of the output signal matches the theoretical exponential decay of an ideal Poisson process for time delays $\tau < 100$ ns.
• Probability Control: The steady-state probability of the output state ($p$) is controllable via the pulse amplitudes. Experimental data aligns with the theoretical model $p = \Gamma_{01} / (\Gamma_{01} + \Gamma_{10} - \Gamma_{01}\Gamma_{10})$. The authors note that even when $\Gamma_{01} = \Gamma_{10}$, $p$ does not equal 0.5 due to the specific timing of the readout relative to the pulse sequence.
• Secondary Correlations: While the overall behavior is predominantly Poissonian, a secondary autocorrelation peak was observed at $\tau \approx 144$ ns, suggesting subtle temporal correlations or memory effects between switching events, a phenomenon also noted in conventional sMTJs.

Significance and Claims
The authors claim that this work establishes A-sMTJs as a versatile platform for integrating deterministic, stochastic, and in-memory functionality on a single chip. By utilizing thermally stable layers, the device offers enhanced resilience to external perturbations (temperature, geometry, material variations) compared to low-barrier sMTJs, while preserving non-volatility.

The paper positions the A-sMTJ as a key component for:

• Probabilistic and Neuromorphic Computing: Providing a hardware element that can generate tunable random bits for probabilistic algorithms.
• True Random Number Generation (TRNG): Offering a source of randomness with controlled statistical properties.
• Hybrid Architectures: Enabling a single device to serve dual roles as a memory element (deterministic switch) and a stochastic element, facilitating the exploration of multiple computing paradigms within a unified physical framework.

The authors maintain a modest tone regarding energy efficiency, acknowledging that the current switching energy (0.8 pJ) is higher than that of thermally activated sMTJs (2 fJ), but suggest this can be reduced through optimization strategies central to STT-MRAM development. The work does not claim to solve specific combinatorial problems directly but suggests that arrays of such devices could be leveraged for large-scale probabilistic and neuromorphic systems.