How close are we to making memristors

Mott memristor: breakthrough for neuromorphic computers

So-called memristors have long been considered an elementary component of a possible hardware neuron. The memristor was theoretically described in 1971 by the American mathematician and computer scientist Leon Chua. The name is made up of the Englishmemory (Memory) andresistor (electrical resistance) together.

The way in which a memristor works is similar to a biological synapse. In a computer chip it can receive signals from other memristors in the form of ions and send them on to the next memristor.

The generated signal varies depending on the strength of the originally received signal. This means that a single memristor can assume and save different states or values. Memristor chips can simulate the brain's rapid exchange of information through short-term voltage pulses (spikes).

The first memristor variants have been around since 2007, they were built by the American nanotechnology specialist Richard Stanley Williams, who at the time worked for the IT company Hewlett-Packard.

The periodic pulses that can be reproduced by the memristor only cover part of the activity of biological neurons. A real hardware neuron has to do more, such as sending out explosive pulses in so-called “bursts” or simulating the self-sustaining oscillations of a neuron.

An artificial neuron with a memristor

Now Williams, who is now working at Texas A&M University, in cooperation with Suhas Kumar from Hewlett-Packard and Stanford researcher Ziwen Wang, is presenting a component that meets the requirements of a hardware neuron.

In their article in the specialist publication Nature, the researchers describe a component that combines electrical resistance and capacitance with a Mott memristor. A Mott memristor is a special form of the classic memristor that allows temperature-driven changes in resistance.

It took the researchers five years to find a suitable composition for the Mott memristor. The most important element is the tiny layer of niobium oxide. Image: Kumar et al.

This process takes place in a layer of niobium oxide (NbO2) in the memristor. With direct voltage, the material heats up and becomes conductive. As a result, the built-up tension is passed on.

The device then cools down again and the niobium oxide layer has an insulating effect. This release of tension corresponds to the action potential of a neuron. The use of the Mott memristor and the exact chemical composition allow the technology to reproduce a large number of different neuronal behaviors.

Hard work that takes decades

It took thirty years to develop the first memristor, and research on the Mott memristor by Williams, Kumar and Wang took five years. The exact composition of the Mott memristor cannot be found by chance, says Williams, "a lot is going on in this small piece of material at the nano-level".

There is still a lot to be done before the hardware neuron can be used. Kumar and Williams want to test other material variants for the Mott memristor, for example. Because for NbO2 the switch is only thrown at 800 degrees Celsius. This only happens in a layer of a few nanometers, but in a neuromorphic computer with millions of these components, the heat could become a problem.

Nevertheless, the hardware neuron is a real “memristor breakthrough,” writes the Institute of Electrical and Electronics Engineers. In the long term, it could overcome the computing power and energy consumption bottleneck of current von Neumann computer architecture in neuromorphic computers and enable “the energy-efficient validation of neuroscientific models”, according to the researchers.

Cover picture: MIT | Via: Nature, Readcube (Supplement)

Read more about Artificial Intelligence:


Mott memristor: breakthrough for neuromorphic computers was last modified: November 21st, 2020 by Maximilian Schreiner