Scaling AI Infrastructure With Optical Interconnects

Author: Joseph Ber, Marvell Collection: 2025 OCP Global Summit - Expo Hall Sessions

Summary

Marvell presents a comprehensive overview of optical interconnect technologies required for exascale AI factories, covering three distinct application domains: scale-up (XPU-to-XPU within and beyond racks via CPO), scale-out (rack-to-rack via PAM4 pluggable transceivers), and scale-across (site-to-site via coherent and ZR/ZR+ pluggables). The vision targets AI factories with over 1 million GPUs, 10 exabits of bandwidth between sites, and multi-gigawatt data center regions. Marvell's CPO (co-packaged optics) technology uses a 3D integrated silicon photonics light engine with electronics, photonics, substrate, and integration layers, delivering 6.4T per engine (32 x 224G optical lanes), with four engines per XPU substrate providing 25.6T optical connectivity in a compact form factor.

For scale-out, the industry is currently in high-volume 800G shipping, ramping 1.6T (8x200G), with 3.2T (requiring 400G/lane optics and electrical) in development and 6.4T+ on the horizon. The generation cadence is accelerating from roughly two years to one year between generations. For scale-across (DCI), coherent optics and ZR/ZR+ pluggables enable multi-site AI clusters across campus, regional, and metro distances, addressing the fundamental constraint that power delivery limits single-site capacity. Marvell demonstrated 448G PAM optical performance at OFC earlier in 2025, showing the path to 3.2T transceivers.