Artificial intelligence is currently facing a real “energy barrier”: in conventional architectures, data flows continuously between memory and processing units, resulting in high energy consumption and limiting system speed. For Dominique Drouin, a research professor in the Department of Electrical and Computer Engineering at the Université de Sherbrooke and his team, the issue is clear. The aim of their work is to design systems capable of processing information more efficiently by drawing inspiration from the human brain, where memory and computation are closely intertwined. By replicating this organization at the hardware level, the research paves the way for smarter, more autonomous and significantly more energy-efficient connected devices.
To meet this challenge, the team is focusing on an “in-memory computing” approach developed in collaboration with Fabien Alibart, a researcher at the CNRS in France and at the Institut interdisciplinaire d’innovation technologique (3IT) at the Université de Sherbrooke. This method involves performing operations directly within storage devices, thereby eliminating energy-intensive data transfers. This innovation relies on memristors, electronic components that act as artificial synapses that “remember” the current that flows through them. These miniature 100-nanometre memristors are highly stable and durable and can therefore simulate biological learning with minimal energy consumption, in the order of a few femtojoules. Experimental prototypes have proven their capacity to perform inference directly in memory, and to reproduce synaptic plasticity mechanisms that depend on the time and frequency of the voltage pulses.
These advances pave the way for practical applications. A microchip incorporating these artificial neurons has been designed, enabling learning directly on the hardware – a rarity worldwide. This work has also contributed to the creation of technology companies such as AMT, which specialises in equipment for characterizing these systems, now marketed internationally. Furthermore, industrial collaborations with the company Irréversible are already exploiting these architectures to develop innovative solutions in artificial intelligence. Ultimately, these innovations could transform many sectors, particularly where energy constraints are critical, such as the Internet of Things, autonomous systems, and certain biomedical applications.
References
El Mesoudy, A., Lamri, G., Dawant, R., Arias-Zapata, J., Gliech, P., Beilliard, Y., Ecoffey, S., Ruediger, A., Alibart, F., & Drouin, D. (2022). Fully CMOS-compatible passive TiO₂-based memristor crossbars for in-memory computing. Microelectronic Engineering, 255, 111706. https://doi.org/10.1016/j.mee.2021.111706
Amirsoleimani, A., Alibart, F., Yon, V., Xu, J., Pazhouhandeh, M. R., Ecoffey, S., Beilliard, Y., Genov, R., and Drouin, D. (2020). In-memory vector-matrix multiplication in monolithic complementary metal–oxide–semiconductor-memristor integrated circuits: Design choices, challenges, and perspectives. Advanced Intelligent Systems, 2(11), 2000115.
https://doi.org/10.1002/aisy.202000115