Flexible Photonic Architectures

This research area aims to explore the design space of flexible accelerator network architectures by leveraging reconfigurable wavelength selective switches and Dense Wavelength-Division Multiplexing (DWDM) transceivers. The proposed architecture is evaluated using network-level simulations and testbed experiments, and its performance is compared with state-of-the-art accelerator architectures. A key objective of our work to develop a flexible, photonic interconnect that enhances communication efficiency in AI/ML applications. 

Current Projects:

  • Center for Ubiquitous Connectivity (CUbiC) - FlexPAC Architectural Design

    • The objective of this project is to design an accelerator network architecture that is able to accelerate the communication-intensive collectives in AI/ML workloads using silicon photonic technologies. This architecture will leverage novel optical comb sources and fine-grained, highly flexible optical connectivity through multi-wavelength selective switching and multicasting.

    FlexPAC System Architecture
    • Photonically Interconnected datacenter Elements (PINE)  - Transceiver/Switch Testbed Experiments

      • The objective of the PINE project is to develop an energy-efficient and high-performance data center network using silicon photonic hardware components. Our team conducts system design, modeling, simulations and testbed experiments of the proposed architecture and and is working to demonstrate its energy efficiency on both the device/link level as well as on the end-to-end system level.
    • Laboratory for Physical Sciences (LPS/NSA) - Simulation Platform Development

      • The objective of this project is to develop a simulation platform that can be leveraged to conduct large-scale network-level simulations. Our team is building network architectures and experimentation pipelines on this platform to demonstrate the performance improvement and scalability of the proposed silicon photonic network architectures for data centers, high performance computing systems, and accelerator clusters.