Enabling technologies for 800Gb/s and 1.6Tb/s optical interconnects for future data center
Maxim Kuschnerov, Huawei
The path towards 1.6TbE will require 200G/lane optical technology, which could be used both for intensity modulated direct detection (IMDD) schemes and coherent optics. A higher parallelization will be needed going forward and the baud rate vs. number of lane trade-off is driven by the availability requirement of the market. Going forward IMDD technology is facing basic limitations and the question needs to be analyzed whether these can be solved or the transition towards coherent optics is inevitable.
Photonic networks with nanoseconds switching and control for distributed machine learning systems
Dr. Nicola Calabretta, Eindhoven University of Technology
Diverse cloud workloads and AI applications with different requirements result in a 40% underutilization of the servers in conventional server-centric architecture with huge waste of resources and power consumption as well degradation of the application performance when available resources cannot meet application requirements, particularly for AI computing clustering. To flexibly accommodate various computing nodes and memory intensive applications, a novel disaggregated DCN architecture consisting of independent hardware (like XPU, memory, and storage) nodes, innovative DC networks with low latency and high bandwidth are crucial to support the network performance of disaggregated architecture. The talk will discuss the requirements of latency and bandwidth as well as the photonic switching technology to implement the disaggregated network infrastructure. We present an experimental prototype rack-scale disaggregated architecture based on nanoseconds optical switches and discussed the network performance assessment in terms of latency and bandwidth.
Enabling low and stable latency communication using clock and frequency referenced access networks
Zhixin Liu, UCL
The rise of timing-critical applications such as virtual reality and connected car fleets, combined with the rapid growth of the number of user devices, creates new challenges for the latency and reliability of user-cloud data communications. Currently user-cloud communications rely on time-scheduled data frames through tree-topology fibre networks, incapable of assuring guaranteed connections with low or stable latency, which is necessary for, e.g. remote surgeries and safe operations of self-driven cars. Besides, their scalability to a larger user count is limited. Here we show that clock and optical frequency synchronisation, enabled by burgeoning frequency comb and signal processing techniques, can provide each user with dedicated optical bandwidth to enable scalable user-cloud communications that guarantees simultaneously high per-use data rate and low latency. Our approach provides accurate clock and optical frequency synchronisation over deployed optical fibre links, which will be beneficial for many applications including accurate navigation, quantum communications, and astronomy.