Memristors, or memory resistors, are an emerging form of non-volatile memory considered a viable alternative to traditional memory storage. These devices act as switches that can recall their electric state even after their power is turned off, allowing them to behave like human neurons—both computing and storing data.

Memristors are a hot research topic for their numerous unique properties, including scalability to lower process nodes, which enable smaller, faster, and more energy-efficient devices.   

Cea-Leti's wafer-level memristor crossbar arrays using a prober. Image (modified) used courtesy of Cea-Leti

Now, as the demand for more sophisticated and power-efficient computing grows, memristors hold promise in various applications, including artificial intelligence, neuromorphic computing, and high-density data storage. In recent months, several international research teams have published findings about memristors that may help push this technology from lab to market.

The University of Rochester: Strain-Engineered Phase-Change Memristors

Researchers at the University of Rochester recently developed high-performance, bipolar phase-change memristors based on strain-engineered multilayer molybdenum ditelluride (MoTe2).

The strain-based memristor. Image used courtesy of Nature

These devices, distinguished by their bipolar phase-change characteristics, represent a significant leap from conventional filamentary conduction mechanisms. The study successfully utilizes process-induced strain engineering to induce a semimetallic-to-semiconducting phase transition in MoTe2.

This transition is key to achieving the memristor's remarkable switching characteristics, including low voltages of 90 mV, high on/off ratios of 10^8, swift switching times of 5 ns, and prolonged retentions over 10^5 seconds. Such performance metrics indicate the potential of these memristors for applications in robust, fast, low-power, and scalable memory and neuromorphic computing systems.

The device's performance can also be finely tuned by varying a single process parameter: the contact metal film force, a product of film stress and thickness. This control over the device's switching voltage and on/off ratio underlines the memristor's adaptability and potential for diverse applications.

Cea-Leti: Using Memristors for Bayesian Neural Networks

In France, a group of Cea-Leti researchers announced that they have successfully integrated memristor hardware into artificial intelligence systems. Specifically, the team leveraged the inherent probabilistic nature of memristors to effectively execute the computationally intensive tasks of Bayesian neural networks.

Filamentary memristor in the back end of the line of Cea-Leti's hybrid memristor/CMOS process. Image used courtesy of Nature

In the study, the team programmed a Bayesian neural network on an assembly of 75 crossbar arrays, each containing 1,024 memristors, complemented by CMOS periphery for in-memory computing. The experiment involved a two-layer Bayesian neural network trained to differentiate nine classes of heart arrhythmia from electrocardiogram (ECG) recordings.

The experimental results were compelling, with the memristor-based Bayesian neural network nearly matching the best performances of a software neural network in terms of both aleatoric (area under the curve of 0.91) and epistemic uncertainty evaluation (area under the curve of 0.99), albeit with a slight reduction in raw accuracy. This performance sharply contrasted conventional neural networks, which exhibit little ability to recognize unknown data and evaluate uncertainties.

Tsinghua University: 2D-Material-Based Memristors

Finally, in China, a team at Tsinghua University developed a new system-integrated memristor computing-in-memory chip the researchers claim is the world's first of its kind to support efficient on-chip learning. This chip, still in the laboratory phase, is expected to significantly advance developments in fields such as artificial intelligence, autonomous driving, and wearable devices, among others

The memristor-based integrated circuit. Image used courtesy of Tsinghua University

The chip is notably energy efficient, consuming only 3% of the energy of an application-specific integrated circuit (ASIC) system for on-chip learning. According to the team, this innovation started with research on memristors for storage applications and took about 11 years to develop to this point.

Memristors Gain International Steam

Memristors converge computing and memory within a single device, opening up new frontiers for energy-efficient, high-speed, and intelligent computing solutions. As international researchers continue to explore and refine memristor technology, they pave the way for more advanced and efficient computing systems that could define the next generation of electronics.