Implementation of Software Defined Electro-Optical Neural Networks

In recent years, the rising artificial intelligence workload has increased the demand for computational power, challenging traditional hardware like GPUs. Optical neural networks (ONNs) promise to offer faster computation with lower power consumption, but face issues such as scalability and high costs. In light of these challenges, we propose an alternative electro-optical neural network using off-the-shelf optical components, high-speed transceivers, and FPGA control, focusing on a single-layer implementation of the ONNs, to be iteratively re-used. We develop the design of the system, the encoding and decoding schemes, and the implemented training methods. Initial tests confirm the successful operation of perceptrons with TensorFlow models that consider the beforementioned physical implementation for XOR and Iris flower classification. A system efficiency of about 150 pJ/MAC is obtained from preliminary experiments with 10 Gb/s data throughput, that can be improved in future to 33.97 pJ/MAC when utilizing all system components.