Modulation of nonlinearity and asymmetry in a spin–orbit torque driven artificial synapse

Unconventional computing schemes inspired by biological neural networks are being explored with ever growing interest to eventually replace traditional von Neumann architecture-based computation. Realization of such schemes necessitates the development of device analogs to biological neurons and synapses. Particularly, in spin-based artificial synapses, the spin–orbit torque (SOT) can be utilized for changing between multiple resistance states of the synapse. In this work, we demonstrate synaptic behavior, namely long-term potentiation and long-term depression in a ferrimagnet (GdFe) via SOT generated using a heavy metal (Pt). The dependence of the synapse-like output on the input parameters is extensively investigated. Synaptic arrays based on experimental results are simulated and used to perform the classification of a handwritten digit dataset. Correlating the classification accuracy with the experimentally observed synaptic behavior, the performance of the synapse is found to depend on the critical switching currents. Understanding the correlation between the input parameters and synaptic performance could accelerate the development of artificial spintronic synapses possessing high operation speed, nonvolatility and plasticity, thereby enabling efficient compute in-memory systems in the near future.