SKM 2021 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

MA: Fachverband Magnetismus

MA 19: PhD Focus Session: Symposium on "Magnetism - A Potential Platform for Big Data?" (joint session MA/O/AKjDPG)

MA 19.2: Invited Talk

Friday, October 1, 2021, 14:00–14:30, H5

Neuromorphic computing with radiofrequency spintronic devices — •Alice Mizrahi¹, Nathan Leroux¹, Danijela Markovic¹, Dedalo Sanz Hernandez¹, Juan Trastoy¹, Paolo Bortolotti¹, Leandro Martins², Alex Jenkins², Ricardo Ferreira², and Julie Grollier¹ — ¹Unité Mixte de Physique CNRS, Thales, Université Paris-Saclay, 91767 Palaiseau, France — ²International Iberian Nanotechnology Laboratory (INL), 4715-31 Braga, Portugal

The need for energy efficient artificial intelligence has motivated research on the implementation of neural networks in hardware, using emerging technology. In particular, spintronic nano-oscillators have emerged as promising candidates to emulate neurons due to their non-linear behavior. However, in order to scale such systems to deep neural network capable of performing state of the art artificial intelligence tasks, it is necessary to have physical synapses -- which weights can be tuned --connecting the neurons. Here we propose a scalable architecture for neural networks using spintronic RF oscillators as neurons and spintronic RF resonators as synapses. First, we show how individual spintronic resonators, and in particular magnetic tunnel junctions, can multiply RF signals by a tunable weight, thus emulating synapses. Then, we show how to assemble these devices into chains performing the multiply and accumulate function, which is at the core of neural network. Finally, we show how to assemble a full neural network and perform classification tasks. These results open the path for compact and energy efficient deep neural networks.