A French team led by CEA-Leti, with participation from C2N and several other laboratories*, has developed a hybrid memory technology that allows artificial neural networks to learn and make decisions directly on-chip.
The new system combines ferroelectric capacitors and memristors in a single CMOS-compatible stack. The memristors are suited for inference, while the FeCAPs enable precise updates for learning. The hybrid device switches between these modes, allowing efficient learning without relying on the cloud.
This innovation could benefit autonomous vehicles, medical sensors, and industrial systems by enabling them to adapt in real time to incoming data.
The work is presented in Nature Electronics and was supported by the European Research Council and the PEPR Electronics program, as part of the France 2030 initiative.
*University Grenoble Alpes, CEA-List, CNRS, University of Bordeaux, Bordeaux INP, IMS France, and University Paris-Saclay.
References
A ferroelectric–memristor memory for both training and inference
Michele Martemucci1,2, François Rummens2, Yannick Malot2, Tifenn Hirtzlin1, Olivier Guille1, Simon Martin1, Catherine Carabasse1, Adrien F. Vincent3, Sylvain Saïghi3, Laurent Grenouillet1, Damien Querlioz4 & Elisa Vianello1
Nature Electronics (2025)
https://doi.org/10.1038/s41928-025-01454-7
Affiliations
1 Université Grenoble Alpes, CEA-Leti, Grenoble, France
2 Université Grenoble Alpes, CEA-List, Grenoble, France
3 Université de Bordeaux, CNRS, Bordeaux INP, IMS, Talence, France
4 Université Paris-Saclay, CNRS, Centre de Nanosciences et de Nanotechnologies, Palaiseau, France
Contact C2N : Damien Querlioz
Photo : A single memory, which functions as both memristor and FeCAP, for neural network inference and training
Photo credit: © E.VIANELLO-M.PLOUSEY DUPOUY/CEA
