This paper is a contribution to the Physical Review Applied collection titled Physics-Inspired Computing.

Spin-orbit-torque (SOT) devices based on a single ferromagnetic layer (free layer) and a single heavy metal layer, popular for neuromorphic synapse device applications, suffer from a trade-off between thermal stability and current density for switching (which determines synaptic programming energy) due to the interfacial nature of SOT in these devices. In this article, first, magnetic switching is experimentally demonstrated in a heavy-metal($\mathrm{Pt}$)/ferromagnetic-metal($\mathrm{Co}$)-gradient-multilayer device instead, where the thickness of each $\mathrm{Co}$ layer remains fixed, but that of $\mathrm{Pt}$ increases from bottom to top. By applying the thermally assisted SOT switching model on pulse-duration-dependent switching measurements, its thermal stability factor is determined to be higher than 60 (across multiple devices), much higher than that of more common $\mathrm{Pt}$/$\mathrm{Co}$-bilayer-based SOT devices (approximately equal to $30$). Also, the switching current density remains similar in both devices, thereby showing elimination of the aforementioned trade-off in gradient-multilayer stacks. Next, synaptic behavior of this device is experimentally demonstrated: 30 stable and distinguishable states can be controlled through positive current pulses [long-term potentiation (LTP)], and 24 such states through negative pulses [long-term depression (LTD)]. In congruence with the enhanced thermal stability of the device, as reported earlier through the switching measurement, all the above synaptic states are demonstrated to be stable up to $1.2\times {10}^3$ s (stability of some states shown up to $5\times {10}^4$ s as well). Using nonideality coefficients extracted from experimentally obtained LTP and LTD plots, on-chip inference and on-chip learning performance of neuromorphic crossbar arrays of such devices are also estimated through system-level simulations.

Close

PrevNext

View figure in article

• Neuromorphic computing
• Physics of computation
• Spin-orbit torque
• Synapses
• Artificial neural networks
• Devices for digital logic, storage & processing
• Heterostructures
• Magnetic multilayers