Hardware Implementation of a Fully Functional Stochastic p-STT Neuron for Probabilistic Computing

Abstract

A stochastic binary neuron is developed for probabilistic computing using a perpendicular spin-transfer torque (p-STT) neuron device and its associated peripheral circuits. The p-STT neuron device, featuring a 300-nm-diameter pillar structure and a perpendicular magnetic-tunnelling-junction spin valve, is fabricated to exhibit stochastic switching behavior characterized by a sigmoidal function. The stochastic switching mechanism is influenced by the presence of defect sites within the device. The hardware implementation encompasses the construction of the neuron peripheral circuit, including the fire-spike generation, sensing amplifier, and reset circuit, using a 28-nm-design-rule complementary metal-oxide-semiconductor (C-MOSFET) process. This hardware realization also demonstrates stochastic switching behavior with a sigmoidal function. However, a slight shift toward higher input voltage spike amplitudes is observed compared to the p-STT neuron device alone, owing to voltage drop within the peripheral circuit. Furthermore, the software-based investigation focuses on validating the feasibility of a spiking neural network. The temporal correlation detection is estimated using the stochastic switching probability derived from the hardware implementation, indicating successful synchronization between the correlated input neuron and the output p-STT neuron. Probabilistic-bits can be generated using the firing signal of stochastic binary neurons to overcome conventional deterministic computing. Here, a fully functional stochastic p-STT neuron is developed by integrating a stochastic p-STT neuron device with its corresponding peripheral circuitry fabricated by a 28-nm CMOS process in hardware and a temporal correlation detection task is demonstrated using p-STT neuron.