Power efficient ReLU design for neuromorphic computing using spin Hall effect

Abstract

We demonstrate a magnetic tunnel junction injected with spin Hall current to exhibit linear rotation of magnetization of the free-ferromagnet using only the spin current. Using the linear resistance change of the MTJ, we devise a circuit for the rectified linear activation (ReLU) function of the artificial neuron. We explore the role of different spin Hall effect (SHE) heavy metal layers on the power consumption of the ReLU circuit. We benchmark the power consumption of the ReLU circuit with different SHE layers by defining a new parameter called the spin Hall power factor. It combines the spin Hall angle, resistivity, and thickness of the heavy metal layer, which translates to the power consumption of the different SHE layers during spin-orbit switching/rotation of the free FM. We employ a hybrid spintronics-CMOS simulation framework that couples Keldysh non-equilibrium Green's function formalism with Landau-Lifshitz-Gilbert-Slonzewski equations and the HSPICE circuit simulator to account for diverse physics of spin-transport and the CMOS elements in our proposed ReLU design. We also demonstrate the robustness of the proposed ReLU circuit against thermal noise and non-trivial power-error trade-off that enables the use of an unstable free-ferromagnet for energy-efficient design. Using the proposed circuit, we evaluate the performance of the convolutional neural network for MNIST datasets and demonstrate comparable classification accuracies to the ideal ReLU with an energy consumption of 75 $pJ$ per sample.