A 336-neuron 28k-synapse Self-learning Neural Network Chip With Branch-neuron-unit Architecture

Present VLSI technology can integrate up to several hundred neurons with fully connected synapses on a single chip, a number insufficient to realize a practical neural network system. The chip reported here uses I.Om CMOS technology to integrate 336 neurons and 28k synapses, equivalent to 56k symmetrical connections. The chip uses branch-neuron unit (BNU) architecture. Interconnection of 200 chips is believed possible with this BNU architecture, based on the assumption of a 30% firing rate and 1% fluctuation of each neuron unit. Interconnection of 200 chips realizes a neural network system with almost 3,300 neurons and 5.6M synapses (11.2M symmet­rical connections).