Top-Down Profiling of Application Specific Many-core Neuromorphic Platforms

In this paper we present a top-down methodology aimed at evaluating the scalability of Spiking Neural Network (SNN) simulations on massively many-core and densely interconnected platforms. Spiking neural networks mimic brain activity by emulating spikes sent among neurons populations. Many-core platforms are emerging computing targets to achieve real-time simulation SNN. Neurons are mapped to parallel cores and spikes are sent over the on-chip and off-chip network. However, due to the heterogeneity and complexity of neuron population activity, achieving an efficient exploitation of platforms resources is a challenging problem, often impacting simulation reliability and limiting the biological network size. To address this challenge, the proposed methodology, based on customized SNN configurations, extracts detailed profiling information about network usage of on-chip and off-chip resources. We first show the results of the application of our methodology to the SpiNNaker neuromorphic many-core platform. Then, we demonstrate how the profiling information can be used to improve the reliability of biologically plausible SNN simulations