Stochastic memristive devices for computing and neuromorphic applications

Siddharth Gaba, Patrick Sheridan, Jiantao Zhou, Shinhyun Choi, Wei Lü

Abstract

Nanoscale resistive switching devices, commonly known as memristors, have emerged as promising candidates for non-volatile memory, logic, and neuromorphic systems. However, a significant hurdle in their development is the inherent stochasticity—large variations in switching behavior across space and time. This research demonstrates that in metal-filament-based memristive devices, switching can be treated as a fully stochastic process. While individual switching events are random, their distribution and probability can be precisely modeled and controlled. By embracing rather than suppressing this randomness, these devices can serve as fundamental building blocks for error-tolerant computing schemes, such as stochastic computing, and provide the multi-level functionality required for neuromorphic applications. This study verifies these concepts through the demonstration of memristor-based stochastic bitstreams and multi-level "analog" synapses using arrays of binary memristors.

Introduction: The Challenge of Stochasticity

As conventional silicon-based transistors approach their physical scaling limits, the industry is exploring alternative architectures. Memristors are at the forefront of this search due to their high density and low power consumption. However, at the nanoscale, these devices are notoriously "unreliable." Because their switching mechanism typically involves the random formation of a microscopic metal filament, the exact moment a device transitions from a high-resistance state (OFF) to a low-resistance state (ON) is unpredictable.

Traditionally, engineers have attempted to mitigate this stochasticity through "brute-force" methods—applying excessively high voltages or long pulses to ensure a switch occurs. While effective for reliability, this approach significantly increases power consumption and accelerates device degradation. This paper proposes a paradigm shift: instead of fighting the inherent randomness of memristors, we can utilize it as a functional feature for specialized computing tasks.

Methodology and Device Architecture

The researchers fabricated two-terminal crossbar devices featuring an amorphous silicon (a-Si) switching layer sandwiched between a silver (Ag) top electrode and a polysilicon bottom electrode. To characterize the device's randomness, the team applied constant voltages below the standard threshold and measured the "wait time"—the duration between the application of voltage and the actual switching event.

By repeating these measurements hundreds of times, the researchers built a statistical profile of the switching behavior. The technical approach relied on three pillars:

  1. Poisson Modeling: The switching probability was modeled using a Poisson distribution, which is ideal for processes governed by a single dominant energy barrier, such as the nucleation of a metal filament.
  2. Bitstream Generation: Random switching events were digitized into bitstreams, where the ON state represented a "1" and the OFF state represented a "0."
  3. Parallel Array Integration: To move beyond binary states, the team constructed a 4-device array. Each device was paired with a local series resistor to prevent a "winner-take-all" scenario, allowing multiple devices to switch independently and represent a range of values.

Key Insights: Predictable Randomness

The study revealed that while individual events are random, the collective behavior is highly predictable. The wait times followed a Poisson distribution almost perfectly, requiring only a single fitting parameter (τ).

One of the most significant findings was the sensitivity of the characteristic wait time to voltage. A modest increase in bias (from 2.5V to 4.5V) reduced the average switching time by three orders of magnitude, suggesting that voltage acts as a logarithmic "knob" for controlling switching probability. Furthermore, the researchers discovered that switching probability is cumulative. The device effectively "remembers" the total duration of stress; the probability of switching depends on the total time voltage is applied, regardless of whether it is delivered in a single long pulse or a series of short pulses with significant gaps between them.

Applications in Stochastic and Neuromorphic Computing

By leveraging these insights, the researchers demonstrated two primary applications:

1. Stochastic Bitstreams

The team successfully generated bitstreams with controlled "bias ratios" (the frequency of 1s). By adjusting the pulse width, they could tune the ratio; for instance, 300ms pulses produced a ratio of 0.4, while 1000ms pulses increased it to 0.76. This capability is essential for stochastic computing, where numbers are represented by the probability of bits being "1" in a stream, allowing complex arithmetic to be performed with simple logic gates.

2. Multi-level Neuromorphic Synapses

In neuromorphic AI, synapses often require analog or multi-level states to represent weights. By grouping four binary memristors into a single array, the authors created a "multi-bit" synapse. Because each device in the array flips randomly and independently, the total conductance of the array can represent 16 different states ($2^4$). This demonstrates how robust, binary devices can collectively perform the complex analog tasks required for brain-like computing.

Conclusion and Future Outlook

This research proves that the stochastic nature of memristors is not a flaw to be eliminated, but a resource to be exploited. By using the inherent physics of filament formation, researchers can create efficient random number generators and multi-level synapses for AI.

However, challenges remain. While temporal randomness is now well-understood, spatial variation—the differences between individual physical devices—must be further minimized through improved fabrication. Additionally, while low-voltage operation may improve device longevity, long-term endurance and the complexity of peripheral circuitry for large-scale arrays are critical areas for future investigation. Embracing the "noise" of the nanoscale may well be the key to the next generation of efficient, bio-inspired hardware.

Summary updated

Abstract

Nanoscale resistive switching devices (memristive devices or memristors) have been studied for a number of applications ranging from non-volatile memory, logic to neuromorphic systems. However a major challenge is to address the potentially large variations in space and time in these nanoscale devices. Here we show that in metal-filament based memristive devices the switching can be fully stochastic. While individual switching events are random, the distribution and probability of switching can be well predicted and controlled. Rather than trying to force high switching probabilities using excess voltage or time, the inherent stochastic nature of resistive switching allows these binary devices to be used as building blocks for novel error-tolerant computing schemes such as stochastic computing and provides the needed "analog" feature for neuromorphic applications. To verify such potential, we demonstrated memristor-based stochastic bitstreams in both time and space domains, and show that an array of binary memristors can act as a multi-level "analog" device for neuromorphic applications.
BibTeX
@inproceedings{gaba2013stochastic,
  author = {Siddharth Gaba and Patrick Sheridan and Jiantao Zhou and Shinhyun Choi and Wei Lü},
  title = {{Stochastic memristive devices for computing and neuromorphic applications}},
  booktitle = {Nanoscale},
  year = {2013},
  doi = {10.1039/c3nr01176c},
}
Figure