Aristotle University Of Thessaloniki (AUTH)
Photonic integrated circuits for routing and computing applications
Abstract
dc:descriptionThe unprecedented explosive growth of mobile and Data Center (DC) traffic is currently driving the need for an efficient mobile hardware transportation, extreme capacity DC interconnects and innovative high-speed computing engines, enforcing ultra-low latencies, high bandwidth, seamless connectivity, high spectral and energy efficiencies. In this context, photonic integrated circuits (PICs) arise as a promising technological candidate for future cellular, DC and High-Performance Computing (HPC) environments, aiming to replace the existing electronic infrastructure with high-bandwidth, low-power and reconfigurable embedded PICs across the cellular, DC and HPC network hierarchy. Hence, the research contributions of this dissertation have been revolved around the development of fully-reconfigurable Fiber Wireless (FiWi) and optical Burst-Mode Wavelength Converter (BMWC) architectures for spectrally efficient, flexible, and low-cost 5G fronthaul and Time Division Multiplexing (TDM) - Passive Optical Networks (PONs). Additionally, towards addressing the need for high capacity, high radix and low latency DC switching infrastructure, an optical packet switch architecture is developed. Finally, in view of sustaining the massive computations executed in HPC engines, two linear optics architectures are proposed. First, towards meeting the TDM-PON requirements in terms of high-dynamic range photoreceivers and high data rate, a BMWC architecture is theoretically and experimentally demonstrated. The BMWC comprises a differentially-biased Semiconductor Optical Amplifier- Mach-Zehnder Interferometer (SOA-MZI) operating in deeply saturated regime in order to provide simultaneously both power equalization and wavelength conversion when burst-mode operation is targeted. The theoretical analysis reveals that the BMWC can provide power equalization up to 9 dB, while this performance is then experimentally verified at 10 Gb/s operation, relaxing in this way the dynamic range photoreceiver requirements. Additionally, in view of validating the high-speed credentials of the proposed architecture, the BMWC is experimentally evaluated under 20 Gb/s non-Return to Zero (NRZ) burst-mode operation, exhibiting a loud-soft ratio of up to 5 dB. Finally, in order to emulate real TDM-PON burst-mode traffic conditions, the BMWC is tested under non-dispersion compensated transmission links with differential distances, providing error-free operation in all cases. Secondly, acknowledging the requirement of reconfigurability and spectral efficiency in 5G fronthaul networks, a bandwidth reconfigurable FiWi architecture for flexible Centralized- Radio Access Networks (C-RANs) is presented. It comprises a fronthaul bus topology with four 1 Gb/s independent channels that can be re-allocated among two cascaded Silicon Photonic (SiPho) Reconfigurable Optical Add Drop Multiplexer (ROADM) nodes, feeding in parallel 8 independent terminals. Each terminal comprises a Phased Array Antenna (PAA) with beam steering capabilities. The ROADM architecture relies on cascaded MZI interleavers and is fabricated on the low loss Si3N4 TriPleX platform yielding 5 dB insertion losses. The proposed fronthaul bus is experimentally demonstrated employing four FiWi links of 250 MBd QAM16 each, enabling an aggregate rate of 1 Gb/s per PAA terminal, meeting the 5G Key Performance Indicators (KPI) user requirements. Towards validating the reconfigurability of the FiWi architecture, the deployed four channels are reconfigurably dropped either at the first or the second ROADM stage of the bus topology, before being wirelessly transmitted through the PAA terminal, revealing Error Vector Magnitude (EVM) values that meet the 3GPP limits. Motivated by the challenges in terms of capacity and latency faced by currently deployed DC electronic switches and the new requirements imposed by emerging DC disaggregation paradigms, an optical packet switch architecture named Hipoλaos (High Port λ-routed All Optical Switch) is presented and experimentally demonstrated. Hipoλaos integrates a Spanke switching architecture with Arrayed Waveguide Grating Router (AWGR)-based wavelength routing and optical feedforward buffering, towards providing high-port layouts with sub-μs latency. A prototype of a 1024-Hipoλaos layout was experimentally demonstrated, yielding a total capacity of 25.6 Tb/s and latency value of only 56 ns. The switch capabilities of the Hipoλaos prototype were experimentally validated at 25 Gb/s line-rates in unicast and multicast scenarios, with the Bit Error-Rate (BER) measurements revealed error-free operation in both cases. Finally, towards meeting the requirements of an asynchronous communication in optical switching, a Burst-Mode-Clock and Data Recovery (BM-CDR) operating at 25 Gb/s was incorporated at the receiver site of the Hipoλaos architecture, revealing error-free operation with an average power penalty of only 2.88 dB. Demarcating from the data transportation hardware solutions for cellular and DC networks, towards high-speed computing engines, two universal linear optics architectures are introduced and theoretically demonstrated. The first architecture is based on N cascaded size-augmenting Generalized Mach-Zehnder Interferometers (GMZIs) interposed by N phase shifters, constructing in this way a Universal GMZI (UGMZI), that has been used in a comprehensive mathematical formula towards theoretically validating its credentials to support any unitary matrix decomposition. The performance of theUGMZI architecture is evaluated using the fidelity metric, which is a measure of discrepancy between the achieved and targeted matrix element values. In comparison with its counterparts, UGMZI fidelity performance can always be restored to 100% without altering its architectural structure. Migrating to the implementation of arbitrary matrices, two architectures are proposed. The first utilizes UGMZI in a Singular Value Decomposition (SVD) scheme, while the second one comprises a direct elements mapping architecture realized in a coherent crossbar (Xbar) layout. While both designs yield 100% fidelity that owes to their loss-balanced configuration, Xbar layout offers the best scalability credentials as it avoids the node loss cascadability. Finally, the last chapter presents an experimental demonstration of the Xbar architecture (theoretically investigated previously), towards implementing a neuromorphic photonic accelerator that can be incorporated in next-generation computing Deep Neural Network (DNN) engines. A review of state-of-the-art photonic Matrix-Vector-Multiplication (MVM) and Neural Network (NN) engines is presented with respect to both their architectural and technological basis, discussing the pros and cons of the different implementations and analyzing the different perspectives offered by every architectural and technological platform. Exploiting the low-loss and high-fidelity performance of the Xbar architecture described previously, we proceeded with the fabrication of a 2:1 single column Xbar using Silicon Germanium (SiGe) Electro-Absorption Modulators (EAM)s as its computing cells. The synergy of the Xbar architecture with EAM-based fan-in and weighting stages is subsequently analyzed, proceeding also to the experimental deployment that allows for high-accuracy (95.24%) Modified National Institute of Standards and Technology (MNIST) dataset classification at speeds up to 50 GHz. Adopting the experimentally obtained performance specifications into a SiPho NN engine scalability study leads to projected energy and area efficiency values of 54 fJ/MAC and 1.54 TMAC/s/mm2, respectively, stimulating significant breakthroughs in the area of neuromorphic computing. In the same chapter, a 10 GHz all-optical Random Access Memory (RAM) cell is experimentally demonstrated, presenting the fastest read/write operation so far in the literature. The proposed RAM was experimentally evaluated through BER measurements, revealing an error-free operation in both write and read operation with 6.2 dB and 0.4 dB power penalty, respectively.
Degree
thesis:*- Grantor dc:publisher
- Aristotle University Of Thessaloniki (AUTH)
- Year dc:date
- 2023
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Tsakyridis, Apostolos
- Τσακυρίδης, Απόστολος
Subjects
dc:subject × 16- Οπτική μεταγωγή πακέτων
- Ολοκληρωμένα φωτονικά κυκλώματα
- Νευρωνικά δίκτυα
- Οπτικές μνήμες τυχαίας προσπέλασης
- Κέντρα δεδομένων
- Optical packet switching
- Photonic integrated circuits (PIC)
- Artificial neural networks
- Optical RAM
- Data centers
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Φυσικές Επιστήμες
- Δίκτυα ηλεκτρονικών υπολογιστών και Επικοινωνίες
- Computer and Information Sciences
- Natural Sciences
- Computer Networks and Communications
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/54404