Maximized Lateral Inhibition in Paired Magnetic Domain Wall Racetracks for Neuromorphic Computing

Abstract

Lateral inhibition (LI) is an important neuromorphic network functionality, modeled after the biological neuron behavior that a firing neuron deactivates its neighbors and prevents their firing. In most neuromorphic hardware platforms LI is implemented by external circuitry, thereby decreasing the energy efficiency and increasing the area overhead. Recently, the domain wall – magnetic tunnel junction (DW-MTJ) neuron is shown in modeling to be intrinsically inhibitory1. LI in DW-MTJ neurons results from magnetostatic interaction between adjacent neurons (Fig.1): the two DWs, DWN and DWI, are driven by electrical current of densities JeI<JeN, so that the DW velocities vDWI<vDWN and the active Neuron N fires first. DWI falls behind DWN and is subjected to the stray field of Neuron N in the −z direction. It is demonstrated that such stray field can decrease DWI velocity and prevent Neuron I from firing, thus emulating LI. However, the LI mechanism in DW-MTJ neurons has not been studied thoroughly, leading to weak inhibition only in very closely spaced devices. In our work2 we model the current- and field-driven DW motion in two adjacent DW-MTJ neurons to approach these problems. LI is maximized by an optimal magnetic interaction strength, determined by tuning the spacing between the two neurons (Fig.2). The results are explained by the current- and field-driven DW velocity characteristics and are quantified by an analytical model. We further study the influence of device parameters and demonstrate the LI behavior in an array of 1000 neurons, which renders our proposed DW-MTJ array promising for implementing a k-winner-take-all layer. With strong LI achieved, a path towards competitive learning algorithms are made possible on the DW-MTJ neurons. This work is funded by Sandia National Lab. Sandia National Lab (SNL) is managed and operated by NTESS under DOE NNSA contract DE-NA0003525