Zero-Change CMOS Silicon Photonic Neurons for Photonic Neural Networks

Abstract

Modern microelectronics has brought us to a stage where we are able to flirt with the idea of artificial intelligence, but not necessarily use it to its fullest potential. With Moore’s law and Dennard’s scaling being pushed to their absolute limit, the semiconductor industry is well aware of the limitations of the MOSFET. With the end of Moore’s law and Dennard’s scaling coinciding with the rise of machine learning applications, and thus, the demand for information and information processing, radical approaches must be considered in order for computing platforms to keep up with the demand for greater information processing. Recent advances in photonic integrated circuit (PIC) technology have made it much more feasible to employ photonic components side-by-side with electronics, allowing us to have a CMOS compatible, silicon photonic platform that provides us with the best of both worlds — speed from light, and low cost and scalability from CMOS processes. One solution that has been proposed to overcome current limitations of microelectronic based computers is the photonic neural network. Neural networks are currently at the forefront of artificial intelligence because of their superior performance in accurately ac complishing tasks. However, they are incredibly computationally intensive and complex to implement. A CMOS-compatible, silicon photonic platform allows us to build better per forming neural networks by taking advantage of the properties of light, such better energy efficiency and bandwidth, while not imposing any significant fabrication hurdles thanks to its compatibility with CMOS platforms. The Princeton Lightwave Lab, spearheaded by Professor Prucnal, has designed a proven architecture for a photonic neuron, where when put in ensemble, will form a photonic neural network. The photonic neuron has already been demonstrated to have optical-to-optical nonlinearity, fan-in, and indefinite cascadability. However, these characteristics must also be reproducible at higher frequencies. A major factor that will determine this will be based on the photonic neuron’s ability to convert a current signal into a voltage signal that will drive the modulator. As a result, this thesis will focus on exploring and utilizing the CMOS iiicompatible, silicon photonic platform to design and characterize a transimpedance amplifier (TIA) that is able to capable of breaking the gain-bandwidth product in hopes of enabling the testing of the photonic neuron at higher frequencies.