Stochastic Memristive Devices for Computing and Neuromorphic Applications (1304.5993v1)

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 in 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 excessive 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 provide a needed "analog" feature in 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.

• The paper demonstrates that metal-filament memristive devices exhibit fully stochastic switching behavior that follows a Poisson distribution.
• The paper shows that leveraging memristors’ intrinsic randomness enables analog emulation from binary arrays, enhancing energy efficiency and robustness in neuromorphic systems.
• The paper verifies its claims through experimental fabrication and testing of crossbar devices, establishing a foundation for scalable stochastic and neuromorphic architectures.

Analysis of Stochastic Memristive Devices for Computing and Neuromorphic Applications

The paper presents a detailed investigation into nanoscale resistive switching devices, commonly known as memristors, and their potential applications in neuromorphic computing and stochastic computing paradigms. The central challenge addressed is the inherent temporal and spatial variability in these devices. The authors demonstrate that the unpredictable switching behavior of metal-filament-based memristive devices can be leveraged rather than suppressed, offering new avenues for computation that capitalize on stochastic switching.

Memristors have been highlighted for their scalability, low power consumption, and extensive connectivity capabilities, marking them as promising candidates for non-volatile memory, logic, and neuromorphic systems. Traditional approaches to electronic components, notably transistors, are approaching their physical limitations, making the exploration of alternative technologies like memristors a high-priority area of research.

Key Technical Contributions and Findings

1. Stochastic Nature of Switching: The authors demonstrate that the switching events in metal-filament-based memristive devices are fully stochastic, while they contend that the statistical distribution and probability of these events can be adequately predicted and controlled. Empirical evidence shows that the switching times follow a Poisson distribution, allowing for the prediction of switching probabilities through statistical modeling.
2. Implications for Neuromorphic Computing: The intrinsic stochastic qualities of memristors can introduce an essential analog component in neuromorphic applications. A binary memristor array can emulate a multi-level "analog" device, improving the robustness and versatility of models of neural networks implemented in hardware.
3. Stochastic Computing Application: The paper explores how stochastic switching behavior in memristors can facilitate stochastic computing, where analog values are processed as probabilities in bitstreams. This approach exploits the random nature of switching to generate stochastic number distributions without the need for complex stochastic number generators traditionally required in CMOS technology.
4. Experimental Validation: The research includes thorough experimental validation through the fabrication of two-terminal crossbar devices using silver and amorphous silicon electrodes. The experiments confirm the random nature of switching events and illustrate a method to predict the probability of switching in single devices and arrays.
5. Device Fabrication and Methodology: The memristive devices are realized using electron beam lithography and include a layer configuration conducive to exploring resistive switching phenomena. The paper highlights the challenges of device fabrication and the necessity for tight process control to minimize spatial variations that could affect device reliability.

Theoretical and Practical Implications

The work provides a foundational understanding of using stochastic processes in hardware computation, presenting a significant shift from deterministic digital logic to probabilistic models. The research implies that instead of coercing memristors into deterministic roles, embracing their stochastic behavior opens up innovative computational frameworks. These could significantly enhance energy efficiency and fault tolerance in applications such as artificial intelligence, sensory data processing, and neuromorphic devices.

Prospective Research Directions

Future investigations could focus on addressing issues such as spatial variations, device endurance, and reliability to make stochastic computing with memristors a viable technology at scale. Exploring persistent reliability under reduced power conditions and leveraging peripheral circuitry to optimize performance remains crucial. Additionally, further research could assess the utility of memristors in complex logic tasks and evaluate the design of neuromorphic systems that leverage multi-bit synaptic weights to mimic more closely biological neural networks.

By demonstrating a pragmatic approach to utilizing stochastic processes in memristors, this paper lays the groundwork for future advancements in computing architectures that transcend traditional von Neumann paradigms.