Neurogrid: A Mixed-Analog-Digital Multichip System for Large-Scale Neural Simulations
We present a novel Address-Event Representation (AER) transmitter circuit to communicate pulses of neural activity (spikes) within a neuromorphic system. AER circuits allow an ensemble of neurons to achieve large scale timemultiplexed connectivity through a shared communication channel. Our design makes use of token-ring mutual exclusion where two circulating tokens in a 2D array of neurons provide exclusive access to the shared channel. Compared to traditional arbitration-tree-based designs, our design has a higher throughput and lower latency during high spiking activity. This allows the circuit to serve larger neuronal densities and higher spiking activity while maintaining the temporal precision required by the system. Our design also eliminates the address line loads that restricted scalability in previous designs. In addition, our simpler circuit topology leads to area and power savings.