InP photonic integrated multi-layer neural networks: Architecture and performance analysis

We demonstrate the use of a wavelength converter, based on cross-gain modulation in semiconductor optical amplifier (SOA), as nonlinear function co-integrated within an all-optical neuron realized with SOA and wavelength-division multiplexing technology. investigate impact fully monolithically integrated linear functions output respect to number synapses/neuron data rate. Results suggest that inputs can scale up 64 while guaranteeing large input power dynamic range 36 dB neglectable error introduction. also performance its transfer by tuning total rate: The performs about 10% better accuracy than corresponding hybrid device for same These neurons are then used simulate 64:64:10 two-layer photonic deep neural network handwritten digit classification, which shows 89.5% best-case at 10 GS/s. Moreover, we analyze energy consumption synaptic operation, considering full end-to-end system, includes transceivers, network, electrical control part. This investigation when is >18, per operation <20 pJ (6 times higher only engine). computation speed this system 47 TMAC/s, 2.5 faster state-of-the-art graphics processing units, efficiency 12 pJ/MAC, 2 better. result underlines importance scaling networks chip.