Modified Communication Networks for the Simulation of Neuromorphic Systems

Neuromorphic computing systems have been introduced in the past few decades as a paradigm shift in computing architectures, commonly used for biological and artificial neural network simulations and artificial intelligence applications. In order to study the long-term dynamics in neuronal networks (e.g. biological learning), the existing systems lack in a significant acceleration compared to biological time for the simulation of large scale networks, due to the increase of latency and the unfulfilled bandwidth demands in the communication network architectures. This work investigates the traffic load distribution and latency within a communication network, used to simulate an arbitrary neural network with a fixed size and connectivity probability. We report a significant decrease in the latency and bandwidth requirements by modifying the commonly used square mesh grid with additional short- and/or long-range connections between nodes