Quantitative simulations of Spiking Neural Networks on an event-driven FPGA cluster

Abstract

Modern digital neuromorphic system design increasingly demands parallel architectures capable of scaling Spiking Neural Network (SNN) simulations to levels of complexity comparable to the human brain. As a result, developing SNN models that can efficiently leverage such architectures has emerged as a critical research challenge. This work presents a novel Synfire Ring SNN model specifically designed for event-driven, parallel neuromorphic platforms to explore the capacity of neurons and synapses, addressing key limitations related to communication latency and resource inefficiency observed in previous implementations. The proposed model employs uniform neuron pools, thus simplifying network topology while improving scalability and ensuring more predictable performance. Experimental validation in a single POETS box, comprising six DE5-Net FPGAs, demonstrates real-time operation that involves 509,648 neurons and 509,648,000 synapses. Comparative analysis with quantitative simulations of SpiNNaker, evaluated under a fixed wallclock simulation window, highlights the suitability of the proposed Synfire Ring topology for the event-driven parallel architecture. Furthermore, an analytical scalability model, grounded in the experimental data, forecasts near-linear scaling up to 4.1 million neurons and 4.1 billion synapses on the full 48-FPGA POETS system.

More information

Identifiers

ORCID for David B. Thomas: orcid.org/0000-0002-9671-0917

ORCID for Graeme M. Bragg: orcid.org/0000-0002-5201-7977

Catalogue record

Export record

Altmetrics