What question did this study set out to answer?

The aim is to develop a spiking neural network runtime that is energy-efficient for edge AI applications.

April 8, 2026Open Access

NeuromorphicRT: A Production-Grade Spiking Neural Network Runtime for Energy-Efficient Edge AI

MPMichael PendletonUnisys (United States)

Key Points

The aim is to develop a spiking neural network runtime that is energy-efficient for edge AI applications.
Developed a framework architecture supporting five neuron models
Implemented 11 model architectures for vision, audio, and anomaly detection
Designed Temporal Event Neural Network (TENN) with learnable temporal kernels
Performed DNN-to-SNN conversion and delta inference on edge devices
Achieved up to 1000x energy reduction compared to conventional deep neural networks
Reported 61.76% validation accuracy on CIFAR-10 using a 129K-parameter SpikeNet7 model
Framework effective on resource-constrained hardware like Jetson Orin Nano and Raspberry Pi 5

Abstract

NeuromorphicRT is a production-grade spiking neural network (SNN) runtime for energy-efficient edge AI on resource-constrained hardware. This paper presents the framework architecture, which supports five neuron models (LIF, AdEx, Izhikevich, TTFS, Surrogate Gradient), 11 model architectures spanning vision, audio, and anomaly detection, and a Temporal Event Neural Network (TENN) with learnable temporal kernels. We demonstrate up to 1000x energy reduction compared to conventional DNNs on edge targets including Jetson Orin Nano and Raspberry Pi 5. Training results on CIFAR-10 using surrogate gradient backpropagation through time are reported, achieving 61.76% validation accuracy with a 129K-parameter SpikeNet7 model trained on Jetson AGX Orin. The framework includes DNN-to-SNN conversion, event-driven delta inference, and a selective state-space model (Mamba-style) with spike I/O.

Demander à l'IA

Bookmark

View Full Paper