SpiNNaker is a super-computer designed for the large-scale simulation of spiking neural networks and consists of a custom network of up to one million ARM CPUs each of which simulating the behaviour of hundreds or thousands of neurons. The state of each neuron is (typically) recomputed every millisecond and any resulting spikes are multicast throughout the network to other connected neurons. This millisecond update is controlled by a simple timer-counter in each CPU core. Since these timers are all independent, they may be initially out-of-phase and, if the underlying oscillator is not shared, their frequencies may cause them to drift appart over time. Because of this, simulations cannot be deterministic and the models behaviour is nolonger consistent (since the model assumes a globally synchronised state update computation.

This repository contains a number of experiments attempting to solve this problem by synchronising the clocks throughout a SpiNNaker system. Also included is an initial implementation of a synchronisation algorithm and a suitable testbench for SpiNNaker along with a standalone software testbench.