Image
The results of this research pave the way for the deployment of Artificial Intelligence in energy-autonomous embedded systems. They are published in the journal Nature Communications.
Powering artificial intelligence with natural energies
Background
Artificial intelligence (AI) is widely used in various embedded applications, such as patient monitoring. To guarantee safety and minimize power consumption due to communication, it would be preferable to process data directly in these embedded systems. However, deploying AI in extreme environments poses a problem due to its high power consumption. A promising solution to this problem is the design of systems based on memristors, electronic components that are electrically programmable and can therefore store information by modulating their resistance value. The use of these memristors can considerably reduce AI energy consumption, making it even conceivable to create self-powered AI systems by collecting energy directly from their environment, enabling the design of energy-autonomous AIs that do not require batteries.
Most memristor-based AI circuits are based on an analog memory computing concept, exploiting the classical laws of electricity (Ohm's and Kirchhoff's laws) to perform the fundamental neural network operation of multiplication and accumulation (MAC). This concept is difficult to put into practice due to the high variability of memristors, the imperfections of CMOS analog circuits and the effects of supply voltage variation. To overcome these difficulties, memristor-based integrated AI systems use complex peripheral circuits, which are tuned for a particular supply voltage.
This requirement for supply voltage stability is in direct contradiction with the properties of energy collectors such as miniature solar cells, which provide fluctuating voltage and energy, posing a significant obstacle to the realization of memristor-based self-powered AI.
The project in brief
Researchers at the Institut des Matériaux, de Microélectronique et des Nanosciences de Provence (IM2NP, CNRS/Aix-Marseille Université), in collaboration with researchers at the Centre de Nanosciences et de Nanotechnologies (C2N, CNRS/Université Paris-Saclay) and the Commissariat à l'énergie atomique et aux énergies alternatives (CEA-Leti) have designed a binarized neural network, fabricated in a CMOS/memristor hybrid process, designed with an alternative approach that is particularly resistant to power supply fluctuations. This robustness was illustrated by powering the circuit with a miniature, high-bandwidth solar cell, optimized for indoor applications. Remarkably, the circuit remains functional even in low-light conditions equivalent to 0.08 times the average solar flux, suffering only a modest drop in neural network accuracy.
These remarkable results pave the way for the deployment of Artificial Intelligence in energy-independent embedded systems.
TRL level
3