New type of memristor for AI in-memory compute
Researchers in Germany have developed a new type of memristor so that low power edge AI chips do not ‘forget’ when they change from one AI model to the next.
Agentic AI uses optimised models for different tasks, but are finding a major challenge in losing all the data when switching from one model to the next. Memristors such as ReRAM can dramatically reduce the power consumption of edge AI chips by processing in memory, but still struggle with the transition between models. Reliable arrays of these memristors can store both the inference weights for the AI models but also the hidden weights that are used across the models.
The team at the Fundamentals and Applications of Nanoelectrochemistry group at the Peter Grünberg Institute (PGI-7) of Forschungszentrum Jülich have developed memristors, or resistors with memory, built with what they call filament conductivity modification mechanism (FCM).
“We have discovered a fundamentally new electrochemical memristive mechanism that is chemically and electrically more stable,” Prof Ilia Valov who heads up the group. “Its unique properties allow the use of different switching modes to control the modulation of the memristor in such a way that stored information is not lost,” he says.
“Basic research is essential to better control nanoscale processes,” says Valov, who has been working in this field of memristors for many years. “We need new materials and switching mechanisms to reduce the complexity of the systems and increase the range of functionalities.”
Two main mechanisms have been identified for the functioning of so-called bipolar memristors: ECM and VCM but each has its own advantages and disadvantages.
ECM memristors use Electrochemical Metallization to form a metallic filament between the two electrodes—a tiny “conductive bridge” that alters electrical resistance and dissolves again when the voltage is reversed. The critical parameter here is the energy barrier (resistance) of the electrochemical reaction. This design allows for low switching voltages and fast switching times, but the generated states are variable and relatively short-lived.
VCM memristors use a valence change mechanism to change the resistance through the movement of oxygen ions at the interface between the electrode and electrolyte by modifying the Schottky barrier. This process is comparatively stable but requires high switching voltages.
“We therefore considered designing a memristor that combines the benefits of both types,” said Valov. This uses a filament made of metal oxides rather than a purely metallic one like ECM, formed by the movement of oxygen and tantalum ions and is highly stable—it never fully dissolves. “You can think of it as a filament that always exists to some extent and is only chemically modified,” says Valov.
This makes the switching mechanism robust. The scientists also refer to it as a filament conductivity modification mechanism (FCM). Components based on this mechanism have several advantages: they are chemically and electrically more stable, more resistant to high temperatures, have a wider voltage window and require lower voltages to produce. As a result, fewer components burn out during the manufacturing process, the reject rate is lower and their lifespan is longer.
On top of that, the different oxidation states allow the memristor to be operated in a binary and/or analog mode for edge AI chips.
The combination of analog and digital behaviour is particularly interesting for neuromorphic chips because it can help to avoid the overwriting of the models.
The researchers have implemented the new memristive component in a model of an artificial neural network in a simulation. In several image data sets, the system achieved a high level of accuracy in pattern recognition. In the future, the team wants to look for other materials for memristors that might work even better and more stably than the version presented here.
“Our results will further advance the development of electronics for ‘in-memory compute’ applications,” says Valov.
The paper is at 10.1038/s41467-025-57543-w
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
