Abstract
Spin-transfer-torque magnetic random-access memory (STT-MRAM) is a proven technology for embedded non-volatile memory applications. The backhopping phenomena in STT-MRAM, whereby the resistance of the device oscillates under higher current, has been recently explored for emerging spiking neural network applications. We report a detailed characterization of backhopping in foundry compatible STT-MRAM having ~15kb bit-cell arrays by analyzing the behavior of backhopping spike rate versus applied current and temperature. Our study shows that the backhopping in STT-MRAM exhibits the Poisson statistics with a controllable spike rate with current that displays three regimes: non-backhopping, exponential and linear. This mimics the behavior of a rectified linear unit (ReLU) neuron, a commonly used activation function in deep learning models. A spiking neural network (SNN) communication channel is simulated using the derived statistics and a first principles mathematical framework to analyze the reliability performance of backhopping-based SNN in terms of trading-off the accuracy and applied current.