Silicon Photonics Building Blocks for AI Accelerators - An Integrated Approach

  • Kanhaya Sharma

Student thesis: Master's Thesis

Abstract

Silicon photonics is an emerging technology in computing. Photonic integrated circuits consume less power and can achieve much higher data rates and computations compared to their electronic counterparts. In the current digital electronics, Artificial Neural Network (ANN) requires considerable amount of multiply-accumulate (MAC) operations, which can be done by using optical interferometers and interconnects on silicon chips with the present CMOS fabrication processes. It is estimated that with silicon photonics, these MAC operations can be done million times faster and with one thousand times less power than the current state of-the-art Google TPU. As a result, Silicon photonics is considered as an alternative approach to accelerate AI (Artificial Intelligence) hardware which are conventionally time and energy consuming. In the last demi-decade, an extensive effort has been taken by groups at MIT, Princeton, and Stanford to demonstrate photonic neural networks (PNNs) for applications like image and speech recognition. In this work, we demonstrated a programmable PNN for nonlinearly separable applications such as XOR and XNOR gate with reduced footprint and better power efficiency; and studied its tolerance to fabrication and thermal variability on state of-the-art CMOS fabrication platform.
Date of AwardJul 2022
Original languageAmerican English

Keywords

  • Silicon Photonics
  • Neuromorphic Computing
  • AI Hardware
  • Photonic Neural Network
  • Integrated Photonics.

Cite this

'