Fast Switching Serial and Parallel Paradigms of SNN Inference on Multi-core Heterogeneous Neuromorphic Platform SpiNNaker2

TL;DR

The paper addresses the challenge of efficiently deploying SNNs on the heterogeneous SpiNNaker2 platform by deciding between serial ARM-based and parallel MAC-array execution. It proposes a classifier-driven fast-switching compiler that predicts the better paradigm before host compilation, achieving substantial reductions in compilation time and RAM usage, with Adaptive Boost delivering 91.69% accuracy. The work shows how to model memory costs across layer configurations, train and select a high-performing classifier, and integrate it into a switching system that outperforms either paradigm alone in memory efficiency. This approach enables mapping larger-scale, multi-task SNNs on SpiNNaker2, highlighting how heterogeneous hardware can be exploited through predictive paradigm selection.

Abstract

With serial and parallel processors introduced into Spiking Neural Networks (SNNs) execution, more and more researchers are dedicated to improving the performance of the computing paradigms by taking full advantage of the strengths of the available processor. In this paper, we compare and integrate serial and parallel paradigms into one SNN compiling system. For a faster switching between them in the layer granularity, we train the classifier to prejudge a better paradigm before compiling instead of making the decision afterward, saving a great amount of compiling time and RAM space on the host PC. The classifier Adaptive Boost, with the highest accuracy (91.69%) among 12 classifiers, is integrated into the switching system, which utilizes less memory and processors on the multi-core neuromorphic hardware backend SpiNNaker2 than two individual paradigms. To the best of our knowledge, it is the first fast-switching compiling system for SNN simulation.

Figures (5)

• Figure 1: Overview of SpiNNaker2 architecture.
• Figure 2: Schematic of mapping the SNN model on SpiNNaker2 with the switching system, and the details of layer granularity switching.
• Figure 3: The marginal distribution of four univariables based on the acquired dataset. The orange line represents the parallel paradigma, and blue line for serial paradigma.
• Figure 4: Accuracy comparison among 12 classifiers. MLP x means the multilayer perceptron model with x neurons in hidden layer. The red lines mark the accuracy range of training with 20 different random seeds.
• Figure 5: Memory performance comparison among two paradigms, trained classifier (Adaptive Boost, estimate before compiling both paradigms), and the ideal situation (classify after compiling both paradigms).