Gyro: A Digital Spiking Neural Network Architecture for Multi-Sensory Data Analytics

Abstract

Unmanned Aerial Vehicles (UAVs) that interact with the physical world in real-time make use of a multitude of sensors and often execute deep neural network workloads for perceiving the state of the environment. To increase UAV’s operations, it is required to execute these workloads in the most power-efficient manner. Spiking Neural Networks (SNNs) have been proposed as an alternative solution for the execution of deep neural networks in an energy-efficient way. We introduce Gyro, a digital event-driven architecture capable of executing spiking neural networks. The architecture is tailored towards sensory fusion applications and it is optimized for Field-Programmable Gate Arrays (FPGAs). In hardware, we demonstrate the performance of a sensory fusion task using a public dataset of bi-temporal optical-radar data for pixel-wise crop classification. We achieve an accuracy of 99,7%, a peak throughput of 31,82 Giga Synaptic Operations per Second (GSOPS) while consuming 50 pico Joule / Synaptic Operation (pJ/SO).