Impact of Photonic Integration Platforms on the Performance of Neuromorphic Accelerators

Artificial Intelligence (AI) has recently proven to be a powerful and versatile tool, able to achieve super-human capabilities in an ever increasing number of fields such as image recognition, game playing, and text generation [1]. Fig. 1A depicts a typical Deep Neural Network (DNN), the underpin of modern AI: a layered structure built upon a simple computing primitive. The deployment of DNNs represents a major challenge where cumbersome and energy-intensive CPUs/GPUs cannot be exploited, such as in edge computing [2]. In this scenario, analog photonics is promising for realizing AI accelerators meeting the bandwidth and power consumption requirements. Several photonic neuromorphic processors have been recently demonstrated, proving advantages in respect to electronic solutions in terms of bandwidth, latency, and power consumption [2].