SpiNNaker: Distributed Computer Engineering for Neuromorphics
David Lester, Steve Furber
Abstract
SpiNNaker is a biologically-inspired massively-parallel computer architecture optimized specifically for modeling large-scale systems of spiking neurons in biological real time. The biological inspiration is manifest in the lightweight inter-processor communications architecture, which enables a “spike” generated by a neuron modeled on one processor to be transmitted to all of the processors modeling neurons that the source neuron connects to in zero biological time (which we take to be a time much less than a millisecond). The design of such a machine is based upon achieving a balance between the processing power required to execute the neuron and synapse models, the memory hierarchy required to maintain neuron and synapse state sufficiently close to the respective processor, and a communications infrastructure that can meet the latency and bandwidth constraints of the target application.