Design automation for optical computing : Boolean logic and neural networks

Date

2020-06-25

Authors

Zhao, Zheng, Ph. D.

Journal Title

Journal ISSN

Volume Title

Publisher

Abstract

As a promising alternative to traditional CMOS circuits, optics has demonstrated the ability to realize ultra-high speed and low-power information processing and communications. For optical computing tasks including Boolean logic and neural networks, however, there still exist challenges such as optical power efficiency, bulkiness and noise-robustness. To address the aforementioned issues, this dissertation proposes a set of algorithms, methodology, and architectures for optical computing tasks, which include: a synthesis flow that significantly reduces the optical power loss; a set of synthesis algorithms that exploits wavelength-division multiplexing (WDM) for area-efficient optical logic construction; a hardware-software codesign methodology that generates more area-efficient and robust ONNs; and an on-chip, integratable photonic Elman RNN architecture that empowers photonic RNNs the capability of training and tuning the state transformation for the first time.

For the first work, we study the long-neglected optical power depletion problem in previous optical boolean logic synthesis, and propose graph transform techniques along with the exploitation of better optical devices to address this problem. The experiments where various sources of optical power depletion are considered, show the efficacy of our method of generating optical power efficient optical circuits, which also helps to build a much more robust and scalable integrated photonic system.

In the second work, we exploit a special property of light in optical logic synthesis to reduce the number of optical components. The great potential of adopting WDM for efficient optical logic construction, is pinpointed and a systematic synthesis flow is designed considering the practical capacity constraint. Mathematically, we demonstrate the affinity of the capacity-constrained synthesis problem to the hypergraph partitioning and the min-cost max-flow problem. The experiments show the efficiency and efficacy of our method to generate smaller optical implementations for higher optical packaging density.

In the third and fourth work, we focus on optical analog computing, more specifically, optical neuromorphic computing. We study the hardware-software co-design of a slimmed architecture for optical neural networks. The proposed methodology directly considers the structures and constraints of the optical hardware implementation during the software training process. The new design greatly reduces the number of optical components in the previous architecture, leading to a smaller optical hardware implementation. The reduction of the cascaded optical components also brings about better robustness against device and environment-related noise.

In the fourth work, we propose the first photonic RNN architecture realizing the widely-used Elman RNN model to facilitate the applications with sequential information. To compensate for the delay variation due to manufacturing imperfections and/or the environmental change, we further introduce a hardware remediation mechanism along with a software training flow to selectively apply remediation to the most sensitive parts. The simulation showed we could effectively reduce the performance degradation by using the proposed flow.

The effectiveness of proposed algorithms and techniques is demonstrated in this dissertation. These approaches can achieve the improvements regarding specific metrics and eventually advance the design of more compact, energy-efficient, and robust optical computing circuits.

Description

LCSH Subject Headings

Citation