Ideas for multi-GPU implementation, from morrison-etal05.pdf

[stiber]

Morrison et al. describe a scheme for minimizing communication between machines in a distributed neural simulation. Their approach essentially involves partitioning the graph representing neural connectivity among the devices, then re-representing the network as a higher-level graph with vertices being devices and edges "bundling" only those connections that cross between. They also take advantage of minimum delay timing for links to buffer spikes and send them as larger packets.