Magnetic tunnel junction synapses for neuromorphic computing

Deep learning algorithms are now widely used. However, the amount of computation power they require to run on conventional CMOS electronics remains high. Consequently, there is an important need for specialized fast and energy-efficient processors tailored for deep learning. We believe that magnetic tunnel junctions in a crossbar array potentially have all the required characteristics of an ideal synapse: high resistance (kOhms), gigahertz speed and symmetric and bi-directional partial switching behavior.
Following up on the pioneering work of Lequeux et al. [1], we have studied different device geometries and material sets. Here we demonstrate nanodevices with resistances on the order of several kiloohms in which several intermediate resistance steps were obtained and their controlled partial switching behavior were achieved using nanosecond pulses.

[1] Lequeux, S., Sampaio, J., Cros, V., Yakushiji, K., Fukushima, A., Matsumoto, R., ... & Grollier, J. (2016). A magnetic synapse: multilevel spin-torque memristor with perpendicular anisotropy. Scientific reports, 6, 31510.