Comment les memristors rendent l’IA plus semblable à l’humain

L’essor du matériel d’IA neuromorphique, semblable au cerveau

As AI becomes the center of the tech industry, a growing problem has emerged: the massive computing and energy demand of AI when performed using CPUs and GPUs.

As a result, researchers are working hard on Neural Processing Units (NPUs), also called neuromorphic chips, a type of AI hardware that mimics the brain’s neurons.

“Ce n’est pas que nos puces ou ordinateurs ne sont pas assez puissants pour ce qu’ils font. C’est qu’ils ne sont pas assez efficaces. Ils consomment trop d’énergie.”

Professeur Joshua Yang – University Of Southern California

Le passage au matériel inspiré du cerveau pourrait transformer notre approche de l’intelligence artificielle. Les conceptions neuromorphiques offrent trois avantages majeurs par rapport aux puces conventionnelles:

• Architecture adaptative: circuits pouvant se reconfigurer en fonction des données d’entraînement.
• Efficacité énergétique radicale: dans certains cas, n’utilisant que 1/100^th de la puissance d’un GPU.
• Moins de chaleur dégagée: réduction des coûteux besoins en refroidissement qui affectent les centres de données IA actuels.

(Vous pouvez en savoir plus sur le matériel spécialisé IA, y compris les NPU, dans notre rapport dédié sur le sujet.)

“Être capable de développer des micro-puces qui imitent l’activité neuronale réelle signifie que vous n’avez pas besoin de beaucoup d’énergie en veille ou lorsque la machine n’est pas utilisée.

C’est quelque chose qui peut représenter un énorme avantage potentiel sur le plan computationnel et économique.

John LaRocco – chercheur en psychiatrie à Ohio State’s College of Medicine.

Researchers are testing several promising methods for creating neuromorphic chips. One approach involves exploiter la ferroélectricité naissante—a still poorly understood phenomenon that could allow materials to spontaneously switch their electric polarization under the right conditions. Another focuses on substrats actifs à base de vanadium ou de titane, materials that can dynamically change their electrical properties to mimic brain-like signaling.

Perhaps the most widely discussed path is the use of memristors—a revolutionary class of electronic components capable of storing information through resistance changes. These devices can perform AI tasks at as little as 1/800^th the normal power consumption, making them one of the most energy-efficient solutions under development.

Comment les memristors imitent les synapses

Glissez pour faire défiler →

Les memristors sont des composants électroniques qui imitent les synapses reliant les neurones en se souvenant de l’état électrique auquel ils ont été basculés après la coupure de l’alimentation.

This can greatly reduce the energy and time lost from shuttling data back and forth between processors and memory.

One of the key strengths of memristors is their capacity for efficient and self-adaptive in situ learning, which is critical for applications in robotics and autonomous vehicles.

Moreover, the low power consumption of memristors is particularly beneficial in robotics and autonomous vehicles, where energy efficiency is paramount.

Many paths are being explored on how to create the best memristors, from relatively conventional titanium oxide memristors to using actual human neurons (organoids) or even mushrooms.

The idea of using organic material, including actual neurons, to mimic the activity of neurons makes sense on a theoretical level. However, in practice, interfacing such a “computer” to traditional IT systems can be challenging.

 “Our existing computing systems were never intended to process massive amounts of data or to learn from just a few examples on their own.

One way to boost both energy and learning efficiency is to build artificial systems that operate according to principles observed in the brain.

Professeur Joshua Yang – University Of South California

This lack of efficiency is staggering when compared to the human brain. A young child can learn to recognize handwritten digits after seeing only a few examples of each, whereas a computer typically needs thousands to achieve the same task.

And the human brain performs this feat while consuming around 20W of power, while the latest AI data centers are looking at GW-scale, or almost a hundred million times more power.

A new intermediary option could be to create artificial chips that act like neurons in their basic principle. This is the path taken by researchers at the University of Southern California, University of California, University of Massachusetts, Syracuse University, Air Force Research Laboratory, and NASA Ames Research Center.

They published their results in Nature Electronics¹, under the title “Un neurone artificiel à pointes basé sur un memristor diffusif, un transistor et une résistance”.

Reproduire le déclenchement des neurones à l’aide de memristors diffusifs

Comment les neurones fonctionnent-ils ?

The way neurons interact with each other, and ultimately process information, is by using both electrical and chemical signals.

If a signal is strong enough, it generates an electrical impulse called an action potential by allowing positively charged sodium ions to flood into the cell.

Source: Nature Electronics

When this electric signal is received, it causes the release of neurotransmitters.

Until now, electronic memristors and complementary metal–oxide–semiconductor (CMOS) circuits have used electric signals to virtually simulate such functioning, requiring hundreds of transistors to simulate a neuron.

Instead, the researchers developed a device called a “diffusive memristor”, which also uses actual chemical interactions to start computational processes.

Qu’est-ce que les memristors diffusifs et comment fonctionnent-ils ?

While traditional silicon systems rely on electrons to perform computations, diffusive memristors use the motion of atoms instead. They use silver ions embedded in oxide materials to generate electrical pulses that mimic natural brain functions.

Source: Nature Electronics

Of course, this does not replicate exactly how a neuron works, but the principle is very similar.

“Even though it’s not exactly the same ions in our artificial synapses and neurons, the physics governing the ion motion and the dynamics are very similar.”

Professeur Joshua Yang – University Of South California

In part, this similarity comes from the fact that silver ions are easy to diffuse in this memristor system, similar to how sodium ions can move in organic cells.

Besides silver, the memristor also uses palladium, silicon, titanium, and hafnium. The research could visualize in real time the diffusion of silver in response to an electric stimulus.
Glissez pour faire défiler →

L’avenir : puces neuromorphiques et IA à faible consommation d’énergie

A key advantage of this new type of memristor is that it fits within the footprint of a single transistor, whereas older designs required tens or even hundreds.

The initial test used only a handful of such diffusive memristors, demonstrating that they can be used for building the typical multi-level neural network used by almost all AI systems today.

Source: Nature Electronics

The next step will be to assemble many more of such systems to test how efficient they can be at actually performing AI tasks.

“We are designing the building blocks that eventually led us to reduce the chip size by orders of magnitude, and reduce the energy consumption by orders of magnitude.

So it can be sustainable to perform AI in the future, with a similar level of intelligence without burning energy that we cannot sustain.

Professeur Joshua Yang – University Of South California

Finding out if other ions can be used could also be useful, as silver ions are not commonly used in semiconductor manufacturing, which could limit the speed of adoption of this design by the industry.

Another effect of diffusive memristors is that they could help better understand how biological brains work.

In the long run, they are likely to be especially useful for so-called “edge computing”, where computation is done directly on site, like for example with a robot or self-driving car having to make a decision without connection to an AI data center.

Investir dans les fabricants de puces neuromorphiques

Intel

Intel (INTC ) is a giant in the semiconductor sector and has evolved over the years from a founder of the industry to a scientific and innovation leader, losing the top spot of manufacturing volume to companies like Taiwan’s TSMC.

Intel is a leader in neuromorphic computing, including through its Loihi 2 chip.

It also created the Intel Neuromorphic Research Community, which includes Pennsylvania State University, involved in vanadium dioxide research, as well as many other potential neuromorphic designs, and 75+ other research groups.

Intel is also very active in mimicking biological sense through replicating the way our brain works (itself a branch of neuromorphic computing), something we discussed further in our article “Puces olfactives biomimétiques : L’intelligence artificielle et les nez électroniques sont-ils le prochain canari dans une mine de charbon ?”.

Overall, research from Intel Lab is at the forefront of semiconductor innovation, including AI, quantum computing, neuromorphic computing, etc. (We discussed Intel advances in quantum computing in our article “L’état actuel de l’informatique quantique”).

You can also read more about Intel’s current business and R&D programs in our dedicated investment report.

Dernières actualités et développements de l’action Intel (INTC)

Étude référencée

^{1. Zhao, R., Wang, T., Moon, T. et al. Un neurone artificiel à pointes basé sur un memristor diffusif, un transistor, et une résistance. Nature Electronics (2025). https://doi.org/10.1038/s41928-025-01488-x}