Computing With Physics

The Spin Torque Nanooscillators that we use are magnetic tunnel junctions -- the same nanodevices as those used in spin torque MRAM. However, when we apply a current to them these devices do not switch; instead their magnetization precesses at radio frequency, following a beautiful nonlinear differential equation. Additionally, these devices can couple to each other and to analog inputs. In a way, Spin Torque Nanooscillators are reminiscient of the nanodevices of the brains: they are comapct, but with a rich dynamical physics.

In recent years, we have worked with the group of Julie Grollier (CNRS/Thales lab) to show that the physics of these devices can be used to perform neuromorphic computation. We have used them as artifical neurons, and showed that the dynamics of he devices, their outstanding tunability and capability to couple to each other, can allow them to perform tasks at competitive accuracy.

Superparamagnetic tunnel junctions are magnetic tunnel junctions that switch very easily: thermal noise is sufficient to switch their magnetization. Their electrical resistance, therefore, behaves as a telegraph noise. Far from being useless, such devices provide a source of true randomness, which can be tuned!

They can also be used to implement many brain-inspired (and non brain-inspired) ideas for the stochasticity of the basic units is used as a source of computation. We are exploring this lead with the group of Julie Grollier!