Orthogonal Spin Current Injected Magnetic Tunnel Junction for Convolutional Neural Networks

We propose that a spin Hall effect driven magnetic tunnel junction device can be engineered to provide a continuous change in the resistance across it when injected with orthogonal spin currents. Using this concept, we develop a hybrid device-circuit simulation platform to design a network that realizes multiple functionalities of a convolutional neural network. At the atomistic level, we use the Keldysh non-equilibrium Green's function technique that is coupled self-consistently with the stochastic Landau-Lifshitz-Gilbert-Slonczewski equations, which in turn is coupled with the HSPICE circuit simulator. We demonstrate the simultaneous functionality of the proposed network to evaluate the rectified linear unit and max-pooling functionalities. We present a detailed power and error analysis of the designed network against the thermal stability factor of the free ferromagnets. Our results show that there exists a non-trivial power-error trade-off in the proposed network, which enables an energy-efficient network design based on unstable free ferromagnets with reliable outputs. The static power for the proposed ReLU circuit is $0.56\mu W$ and whereas the energy cost of a nine-input rectified linear unit-max-pooling network with an unstable free ferromagnet($\Delta=15$) is $3.4pJ$ in the worst-case scenario. We also rationalize the magnetization stability of the proposed device by analyzing the vanishing torque gradient points.