Back to search

Aristotle University Of Thessaloniki (AUTH)

Photonic AI accelerators and optical memory architectures

Abstract

dc:description

The rapid growth of artificial intelligence (AI) and deep learning (DL) workloads has exposed fundamental limitations in conventional electronic computing platforms in terms of computational throughput, energy efficiency, and memory bandwidth. Photonic computing offers a promising alternative by exploiting the inherent parallelism and high bandwidth of light. This dissertation presents a comprehensive investigation of photonic hardware building blocks for next-generation AI acceleration, addressing linear computation, nonlinear activation, and memory subsystems within a unified optical computing framework. First, novel photonic accelerator architectures are introduced for high-speed matrix and tensor multiplications, targeting the linear transformations required in neural network (NN) inference. The proposed architectures are based on optical arrayed waveguide grating routers (AWGRs) and exploit optical parallelism through wavelength-, space-, and time-division multiplexing. Experimental evaluation using benchmark NN models demonstrates accurate inference on the Fashion-MNIST and MNIST datasets, while achieving record-high computational performance of up to 262 tera operations per second (TOPS), representing a 24× improvement over prior state-of-the-art photonic implementations. System-level modeling further projects the scalability of the proposed architectures toward peta-scale computing, reaching up to 3.2 peta operations per second (POPS) with energy efficiency at the femtojoule-per-operation level. To enable fully hardware-based NN inference, the dissertation further investigates programmable optical nonlinear activation functions (NLAFs). Two integrated opto-electronic architectures are developed, supporting multiple reconfigurable nonlinear activation functions with tunable characteristics. Experimental results demonstrate the realization of both optical–electrical (O-E) and optical–electrical–optical (O-E-O) nonlinearities operating at symbol rates up to 10 GBd. When incorporated directly into NN training workflows, the proposed activation functions achieve classification accuracies exceeding 90%, highlighting their potential to replace conventional software-based nonlinearities. Finally, the dissertation addresses the critical challenge of memory bandwidth in optical computing systems through the design and experimental demonstration of all-optical random-access memory (RAM) and cache architectures. A fully addressable optical RAM bank composed of monolithically integrated indium phosphide (InP) optical flip-flops (FFs) is demonstrated, followed by its extension to the first fully optical cache memory incorporating an optical tag-comparison subsystem. Both memory architectures operate at multi-gigabit-per-second data rates with error-free performance, enabling high-speed data access without repeated optical–electrical–optical conversions. Overall, this dissertation demonstrates a cohesive set of photonic computing technologies that collectively enable high-performance, energy-efficient, and scalable AI acceleration. The presented architectures establish key building blocks toward fully integrated photonic computing systems for future artificial intelligence applications.

Degree

thesis:*
Grantor dc:publisher
Aristotle University Of Thessaloniki (AUTH)
Year dc:date
2026

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Pappas, Christos
  • Παππάς, Χρήστος

Subjects

dc:subject × 20

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/61882

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Pappas, Christos; Παππάς, Χρήστος. Photonic AI accelerators and optical memory architectures. Aristotle University Of Thessaloniki (AUTH), 2026. http://hdl.handle.net/10442/hedi/61882