Full-scale simulation of a cortical microcircuit on SpiNNaker

SpiNNaker is a digital neuromorphic hardware designed to reduce simulation time and power consumption compared to traditional computing architectures. While it is able to simulate artificial neural networks with 1 ms resolution in real time, its performance for biologically realistic models necessitating shorter integration time steps and featuring a convergence of on the order of 10,000 synapses per neuron has not been fully tested. Furthermore, simulations on SpiNNaker have previously been downscaled compared to the biological numbers of neurons and synapses [1, 2].We here describe the first full-scale simulations of a cortical microcircuit model [3] on SpiNNaker, comparing performance with that of NEST [4]. With approximately 80,000 leaky integrate-and-fire neurons and 0.3 billion synapses, this model is the largest simulated on SpiNNaker to date. The scale-up is enabled by recent developments in the SpiNNaker software stack (https://github.com/SpiNNakerManchester) that allow simulations to be spread across multiple boards. The implementation uses the simulator-independent description language PyNN [5].We consider two types of NEST simulations: one with spikes constrained to a 0.1 ms grid, and one with spikes in continuous time. The comparison with the latter provides a sensitive test for the correctness of SpiNNaker after a major reorganization of its software stack. We describe deterministic comparisons of the single-neuron dynamics and statistical comparisons of the network dynamics, quantified in terms of firing-rate distributions, spiking irregularity, and correlations in the activity of different neurons. These comparisons reveal close correspondence of the NEST and SpiNNaker results, demonstrating the usability of SpiNNaker for large-scale networks with biological time scales