🔬 condensed matter

Superparamagnetic and Stochastic-Write Magnetic Tunnel Junctions for High-Speed True Random Number Generation in Advanced Computing

This paper reviews two magnetic tunnel junction (MTJ) approaches for high-speed, low-power true random number generation—passive-read superparamagnetic MTJs and stochastic-write MTJs—highlighting their distinct performance characteristics, integration benefits with advanced CMOS and STT-MRAM, and specific suitability for probabilistic computing versus edge cryptographic applications.

Original authors: Jonathan Z. Sun, Christopher Safranski, Siyuranga Koswata, Pouya Hashemi, Andrew D. Kent

Published 2026-01-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jonathan Z. Sun, Christopher Safranski, Siyuranga Koswata, Pouya Hashemi, Andrew D. Kent

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your computer needs a constant supply of truly random numbers to do things like secure your bank account, simulate complex weather patterns, or run advanced AI. Usually, computers generate these numbers using math formulas (pseudorandomness), which are fast but not truly unpredictable. To get real randomness, they need a physical source of chaos.

This paper reviews two new ways to build these "chaos generators" using tiny magnetic switches called Magnetic Tunnel Junctions (MTJs). Think of an MTJ as a microscopic door that can be either open or closed, representing a 0 or a 1. The researchers are showing how to make these doors flip randomly on their own or be nudged randomly to create a stream of unpredictable bits.

Here are the two main approaches they discuss, explained simply:

1. The "Spinning Coin" Approach (Superparamagnetic MTJs or sMTJs)

How it works:
Imagine a coin sitting on a table. If you shake the table hard enough, the coin will flip back and forth between heads and tails purely because of the vibration. It doesn't need anyone to push it; the heat energy in the room (thermal fluctuations) is enough to make it jitter.

In this approach, the researchers make the magnetic "coin" (the free layer of the MTJ) so small and light that the natural heat of the room makes it flip between "up" and "down" states constantly.

The Catch: To make it flip fast enough to be useful, the magnetic "coin" has to be very small. However, if it's too small, it becomes sensitive to tiny imperfections in the manufacturing process.
The Problem: The paper notes that these devices often have a hidden "tilt" in the table (called in-plane anisotropy). This tilt is caused by stress in the materials during manufacturing. If the table is tilted, the coin doesn't flip fairly; it might prefer heads over tails. The researchers found that this tilt varies from device to device, making it hard to get a perfectly fair coin across a whole chip.
Best Use: Because this method just "listens" to the noise (passive read) and doesn't need to push the coin, it is incredibly fast (up to 1 billion flips per second) and uses very little power. It's perfect for high-speed tasks like probabilistic computing, where you need massive amounts of random data right next to the processor.

2. The "Gentle Nudge" Approach (Stochastic-Write MTJs or SW-MTJs)

How it works:
Now, imagine a coin that is heavy and stable. It won't flip on its own. Instead, you have to give it a specific push to flip it. If you push it too hard, it always flips. If you push it just right—halfway between "too hard" and "too soft"—it will flip only 50% of the time.

In this approach, the magnetic switch is stable (it stays put until told otherwise). The computer sends a very specific, short electrical pulse to try and flip it. By carefully tuning the strength of this pulse, the researchers make it so the switch flips randomly about half the time.

The Advantage: This uses the exact same technology already used in modern computer memory (STT-MRAM). It's like taking a standard memory cell and just changing how you talk to it. This makes it very easy to build into existing computer chips without needing new factories.
The Trade-off: Because you have to actively push the switch every time, it is slower and uses more power than the "Spinning Coin" method. It's also sensitive to temperature changes; if the room gets too hot or cold, the "nudge" might become too strong or too weak, messing up the 50/50 chance.
Best Use: This is great for "edge" devices (like smart sensors or microcontrollers) that need a reliable source of randomness for security (cryptography) but don't need the extreme speed of the first method.

The Big Picture Comparison

Feature	The "Spinning Coin" (sMTJ)	The "Gentle Nudge" (SW-MTJ)
Mechanism	Lets heat shake the magnet until it flips.	Pushes the magnet with a precise electrical pulse.
Speed	Very Fast (Up to 1 Gbps).	Moderate (Around 0.1 Gbps).
Power	Ultra-low (Just reads the state).	Higher (Needs to write/push the state).
Compatibility	Needs special materials to avoid "tilt."	Uses standard memory manufacturing.
Main Challenge	Fixing the "tilt" so every coin is fair.	Keeping the "nudge" consistent over time and temperature.

Why Does This Matter?

The paper concludes that both methods are promising "True Random Number Generators" (TRNGs). They are much smaller and more energy-efficient than current methods that rely on big processors to generate random numbers.

sMTJs are the speed demons, ideal for future computers that need to process huge amounts of random data instantly.
SW-MTJs are the reliable workhorses, perfect for securing everyday devices because they fit easily into current technology.

The researchers emphasize that for these to become standard in our devices, engineers need to solve specific material science problems: making the "coins" perfectly flat (removing the tilt) for the first method, and making the "nudge" perfectly stable for the second. Once those hurdles are cleared, we could see these tiny magnetic switches powering the next generation of secure and efficient computing.

Technical Summary: Superparamagnetic and Stochastic-Write Magnetic Tunnel Junctions for High-Speed True Random Number Generation

Problem Statement
Advanced computing applications, including probabilistic computing, statistical modeling, and cryptography, require high-throughput, low-power, and compact sources of true randomness. While processor-based pseudo-random number generators (PRNGs) exist, they are often energy-inefficient and lack the physical entropy required for hardware security. Existing physical entropy sources (e.g., thermal noise in CMOS, chaotic circuits, or single-photon detectors) often face trade-offs between bit rate, integration density, and energy efficiency. The paper addresses the need for compact, CMOS-integrated entropy sources that can operate at gigabit-per-second speeds with minimal power consumption and footprint, specifically leveraging Magnetic Tunnel Junction (MTJ) technology already established for Spin-Transfer Torque Magnetic Random Access Memory (STT-MRAM).

Methodology and Approaches
The paper reviews two distinct MTJ-based approaches for generating True Random Numbers (TRNGs), both utilizing modified STT-MRAM structures but operating in different regimes:

Passive-Read Superparamagnetic MTJs (sMTJs):
- Mechanism: These devices utilize a free layer with a magnetic anisotropy energy barrier comparable to or lower than thermal energy ( $k_B T$ ). Thermal fluctuations drive the magnetic moment to fluctuate between states, generating entropy without active write currents.
- Subtypes:
  - Uniaxial sMTJs: Employ a low-barrier uniaxial anisotropy where the moment telegraph-switches between parallel and antiparallel states.
  - Easy-Plane sMTJs (EP-sMTJs): Employ strong in-plane easy-plane anisotropy, confining thermal fluctuations to the film plane. This produces an analog white-noise resistance signal that can be digitized.
- Operation: The device is passively read to extract entropy, minimizing power consumption.
Stochastic-Write MTJs (SW-MTJs):
- Mechanism: These use magnetically stable free layers (similar to standard STT-MRAM) but are operated with reduced write voltage or shorter pulse durations. This intentionally induces probabilistic switching (Write Error Rates $\approx 0.5$ ) rather than deterministic switching.
- Operation: An active "strobe" write pulse is applied, and the resulting state (0 or 1) is read. The randomness arises from the stochastic nature of the spin-transfer torque switching at finite temperatures.

Key Results and Performance Metrics
The authors present experimental demonstrations and theoretical analyses for both approaches, validated against NIST SP800 test suites:

sMTJ Performance:
- Speed: Single EP-sMTJ devices have demonstrated bit rates of 0.5–1 Gb/s. Combining two independent streams via XOR operations yields rates $>1$ Gb/s.
- Quality: Bit streams pass the NIST SP800-22r1a test suite.
- Efficiency: Estimated energy efficiency is $\sim 0.03$ pJ/bit, with cell footprints as small as $\sim 0.01$ $\mu m^2$ (excluding circuit overhead).
- Challenges: A significant technical hurdle is the presence of unintended, uncontrolled in-plane uniaxial anisotropy caused by magnetostriction coupling to residual stress fields in the Back-End-of-Line (BEOL) process. This variability affects device-to-device consistency. Additionally, the autocorrelation time ( $\tau_{ACF}$ ) in uniaxial designs is sensitive to temperature and magnetic volume.
SW-MTJ Performance:
- Speed: Achieves bit rates of $\gtrsim 0.1$ Gb/s per device.
- Quality: Bit streams pass NIST tests, often requiring a single XOR operation on two independent streams to remove bias.
- Efficiency: Estimated energy efficiency is $\sim 0.08$ pJ/bit.
- Stability: Operating in the "ballistic limit" (nanosecond pulses) reduces sensitivity to temperature variations. However, long-term temporal drift and device-to-device variability in switching probabilities remain challenges for robust deployment.
Comparative Analysis:
- Both approaches offer orders-of-magnitude better energy efficiency and smaller footprints compared to CPU/GPU-based generators and other physical entropy sources (e.g., chaotic CMOS or SPADs).
- sMTJs are best suited for applications requiring the highest data rates near logic processors (e.g., probabilistic computing) due to their passive, read-only nature which avoids CMOS current scaling limits.
- SW-MTJs are attractive for edge-oriented microcontrollers and cryptographic functions where co-integration with existing STT-MRAM arrays is prioritized over maximum bit rate.

Significance and Claims
The paper claims that MTJ-based entropy sources represent a promising foundation for next-generation secure and energy-efficient computing systems.

Integration: sMTJs offer compatibility with the most advanced CMOS nodes due to their lack of write-current requirements, while SW-MTJs leverage standard STT-MRAM process flows, enabling co-integration with embedded memory on the same chip.
Scalability: The authors emphasize that realizing the full potential of these technologies requires addressing specific material and process challenges:
- For sMTJs, reducing magnetostriction-induced in-plane anisotropy by an order of magnitude is critical to ensure uniform fluctuation characteristics across devices.
- For SW-MTJs, improving temporal stability and minimizing device-to-device variability are essential for reliable large-scale deployment.
System-Level Impact: The paper notes that while generating uniform random bits is the first step, future research must also address efficient methods for sampling and controlling the statistical distribution of these bits (e.g., generating Gaussian distributions) to fully realize efficiency gains in probabilistic computing workloads, as digital mapping steps currently dominate the computational cost.

In summary, the paper positions these two MTJ approaches as viable, high-performance alternatives to current TRNG technologies, provided that specific materials engineering and process control challenges regarding anisotropy and variability are resolved.

1. The "Spinning Coin" Approach (Superparamagnetic MTJs or sMTJs)

2. The "Gentle Nudge" Approach (Stochastic-Write MTJs or SW-MTJs)

The Big Picture Comparison

Why Does This Matter?

Technical Summary: Superparamagnetic and Stochastic-Write Magnetic Tunnel Junctions for High-Speed True Random Number Generation

More like this