EdgeSNN-RT proposes a Python-managed, GPU-accelerated spiking neural network simulation method optimized for edge platforms, integrating their definition, execution, and, critically, low-cost host-device synchronization. A dense spike bitfield logging tool groups spikes on the device, obviating the need for one-timestep communication, yielding up to 10x cost savings for logging spikes. Experiments conducted for full-scale cortical microcircuit simulation and long-term conditioning tasks with large time horizons range across variant models of diversified, consumer-class, and larger-scale, heterogeneous GPUs, from the low-power, as-specified, 15W Jetson Xavier NX, to measure the latency, kernel execution time, and interpretative overheads of their execution in the real-time regime specifically for the purposes of the challenge. The system supports real-time or faster simulation execution for today’s leading-class GPUs, and, for shared-memory architectures of embedded platforms, preserves their leading performance with direct, low-level control, NumPy direct views to the managed buffers, and an events-based representation of plasticity within a convenient API. On-device inference and learning of spiking neural networks for real-world tasks will become feasible for networks of the described architecture based on this work, mapping telemetry, memory representation, and scheduling approaches to fit within limitations imposed by embedded platforms, supplanting today's restricted assumptions about direct, one-step execution communication and/or awaiting improvements in physical implementation technology for general programmed computation targets from desktop workstations back toward rudimentary, network-edge platforms in an industry led by programmability for applied innovation.

Spiking Neural Networks GPU Acceleration Edge Computing Embedded Systems Python CUDA