Hybrid Chip Set for Artificial Neural Network Systems

This paper proposes a complete self-contained chip set for ANN systems based on back-propagation model with on-chip learning. We mixed both technologies, digital and analog, as to combine the best of the two worlds. Our system which is constructed from two distinct chips, SynChip and NeuChip, performs the entire neural computations in analog mode. Interface and communications with host and I/O systems is totally digital. A representative system of 3 layers and 100K analog (16-level) synapses per layer would require 300 SynChips and 40 NeuChips and sustain a performance of 31.46 × 10¹⁵ CPS. Moreover, such a system acquires direct input and output ports with a bandwidth of 12800 digital bits or 3200 analog channels.