Connectivity effects on energy and area for neuromorphic system with high speed asynchronous pulse mode links

Abstract

Hardware neuromorphic systems are challenged to achieve biologically realistic levels of interconnectivity. When building a physical implementation of a neural net, the properties of the media immediately impose limits on the number of interconnects and available timing options. The design of any system must consider the energy and area costs associated with the physical layout of neuron core connectivity, rst, by accepting the wiring limits imposed by Rent's rule and second, by understanding the temporal overhead introduced by routing. The presented results show the energyarea trade-o for a model of a neuromorphic system with event driven interconnections. The low area overhead of the asynchronous pulse-mode links create an attractive opportunity for a digital neuromorphic system with a connectivity model closer to the existing software models of neural nets.