Optics Primer, Part 1: Traditional Pluggable Optics (www.chipstrat.com)

A new series exploring optical interconnects has begun with a focus on traditional pluggable optics, which are essential for data communication between GPUs in AI training workloads. As large models are distributed across multiple GPUs, efficient data transfer via optical interconnects is crucial. This method outperforms copper in terms of distance and bandwidth, allowing for the rapid exchange of activations, gradients, and parameters across racks and data centers. However, within the rack, data must be converted from optical to electrical signals via pluggable transceivers, which facilitate this conversion while being removable for ease of maintenance and flexibility. The significance of pluggable optics lies in their ability to decouple optical components from switches, enabling operators to mix vendors and upgrade bandwidth without redesigning switch infrastructure. While their design offers flexibility, it introduces challenges such as increased power consumption and latency, particularly as data rates reach 100G per lane and higher. The reliance on digital signal processing (DSP) to manage signal integrity along long copper traces adds to the inefficiencies. Future articles in this series will explore innovations like linear-drive optics and co-packaged optics, aiming to address these limitations by minimizing DSP overhead and enhancing system performance.