Spintronic Neuromorphic Computing

As AI processors increasingly demand more energy, advances in nanotechnology and material science offer a solution. Our team harnesses the potential of spintronic nanodevices to achieve more sustainable circuits.

Spintronic materials possess rich physics that enables them to dynamically process information in intricate ways, as well as to store it locally across various time scales. This allows for the execution of complex tasks at the nano-scale within integrated circuits, and provides alternatives to the energy-intensive CMOS technology. Neuromorphic spintronics is a very active field of research, that we review in: J. Grollier et al, Nature Electronics 3, 7 (2020), [arXiv]

In the lab, we are constructing experimental neural networks that combine the unique properties of these materials and exploit the multifunctionality of spintronics. We work both at the device level, mimicking synapses and neurons, and at the system level by leveraging their capacity to dynamically interconnect and evolve collectively.

Our aim is to design specialized accelerators that outperform CMOS in efficiency and unlock capabilities previously unachievable with CMOS through the seamless integration of nanodevices in large-scale hybrid architectures.

Selected contributions:

Skyrmionic weighted sums: We showed that topological magnetic nano-particles called skyrmions can implement a key operation of neural networks, the weighted sum of neural signals by synapses, by imitating properties of biological neurotransmitters vesicles. Nature Nanotechnology (2025) [arXiv]

Fully-spintronic RF neural networks: We demonstrated with our collaborators in France and Portugal a miniature version of a neural network fully composed of magnetic tunnel junction synapses and neurons. This architecture relies on radio-frequency (RF) neural connections and can classifies both RF and DC signals without digitization, making it particularly useful for medical applications, such as microwave mammography, and airborne traffic management. Nature Nanotechnology (2023), [arXiv]

Convolutional Neural Networks: We show through numerical simulations and experiments that Radio-Frequency spintronic neural networks can implement the same operations and architectures than software neural networks, in particular convolutional neural networks which are widely used today for image recogntion. Neuromorph. Comput. Eng. 2 034002 (2021) [arXiv]; IEDM (2024) [arXiv]

Nano-neurons: With our collaborators in France, Japan and the US, we demonstrated that nanoscale spintronic oscillators can perform pattern recognition, such as classifying spoken digits and vowels, at a level of accuracy equivalent to conventional software-based neural networks. To achieve these results, we harnessed the high stability of magnetic tunnel junctions (originally developed for memory applications) and leveraged their nonlinear dynamics to emulate neural behavior. These results showed, for the first time, that fully nanometer-scale devices can execute complex neuronal computations in real-time by exploiting their physical properties—thereby demonstrating a proof of principle for truly miniaturized neuromorphic hardware. Nature 547, 428–431 (2017) [arXiv]; Nature 563, 230 (2018) [arXiv]

Spintronic memristors: we showed with our japanese collaborators that the current-controlled motion of a domain wall in a magnetic tunnel junction tunes its resistance. This memristive behaviour imitates the synaptic plasticity in the brain. Nature Physics (2011) [arXiv]; Scientific Reports (2016) [arXiv]