Breaking the GPU Cluster Bottleneck: High-Capacity Optics for the 1.6T Era - Lumentum, VIAVI, AFL & Genuine Optics
With AI model training scaling far faster than network bandwidth, the optical layer has become the primary bottleneck for the performance and operating efficiency of next-generation GPU clusters. In this in-depth discussion, Lumentum and test leader VIAVI Solutions, joined by experts from AFL and Genuine Optics, mapped out the technology roadmap toward 1.6T and 3.2T, and argued that AI infrastructure is shifting from a pure "bandwidth race" to an all-round contest over density, power and reliability.
Core Technology Insights: From 400G per Lane to a Practical CPO Path
Lumentum: Optimizing AI Optical Networking Along Three Dimensions
Christian Urricariet, Senior Director of Strategic Marketing at Lumentum, stressed that the AI infrastructure challenge is not just raw bandwidth but energy efficiency, reach, and bandwidth density at the ASIC edge (shoreline density). He outlined three technology pillars for next-generation clusters:
Pushing the limits of indium phosphide (InP) EMLs: 100G/lane is deployed at scale and 200G/lane is entering early ramp. Lumentum demonstrated 400G/lane technology at OFC, which will be key to single-mode 3.2T links.
The efficiency of optical circuit switching (OCS): to handle the relatively "deterministic" traffic patterns of AI training, MEMS-based passive optical switches can replace conventional packet switches at the spine layer, saving up to 65% of power while delivering ultra-low latency and protocol transparency.
CPO and external laser sources (ELS): once pluggable power and laser count become reliability concerns, co-packaged optics (CPO) becomes inevitable. Lumentum favors ELSFP-style external lasers, which move the heat source away from the ASIC and keep the serviceability of a pluggable.
VIAVI Solutions: Tackling the Production Challenge of Going "Fiber-Wide"
Matthew Adams, Product Director at VIAVI, introduced a key concept: "Fiber-Wide". As CPO and multi-lane transmission spread, the number of optical paths is growing geometrically.
A 100x jump in test density: test-equipment density has grown 100-fold over the past 30 years; VIAVI can now fit nearly 200 independent power-meter channels in a single 3U chassis to meet high-throughput production needs.
Metrology challenges with new fiber types: silicon photonics has revived polarization-maintaining (PM) fiber, mainly for "in-box" interconnects. On top of that, accurately identifying individual cores in multicore fiber during test, and splice loss in hollow-core fiber, are real manufacturing challenges.
Expert Panel: Hidden Challenges and Technical Divides in the AI Era
In a panel discussion moderated by Fibre Systems editor Keely Portway, the experts debated "hidden bottlenecks" and the "2030 outlook":
Where Each Company Stands
Genuine Optics (David Huff): emphasized the cost of reliability. In a 512K-GPU cluster, he noted, optical failures causing 10% downtime would cost as much as $3 million per day. Silicon photonics FIT (failures in time) therefore has to be pushed below 1.
AFL (Ted Lichoulas): focused on connector damage risk. With fiber counts exploding, conventional cleaning and maintenance become extremely difficult. He is bullish on expanded-beam technology in the data center to reduce the impact of dust on high-density connectors.
Lumentum (Christian Urricariet): predicted that coherent technology will start moving inside the data center around 2030 to solve signal-integrity issues beyond 400G/lane.
Key Consensus and Disagreements
Consensus: bandwidth density at the ASIC edge has hit its limit, and conventional pluggables will face severe physical space constraints beyond 1.6T. CPO is no longer a question of if, but when.
Disagreement: on the lifespan of pluggables, Christian believes their flexible supply-chain model means they will coexist with CPO for a long time, while David Huff predicts that, for extreme power management, DSPs may be widely removed after 2030 in favor of linear-drive architectures.
Simple Tech Trend's Take: The "Determinism" Revolution in AI Networking
The session sent a clear signal: optics is evolving from a simple "transmission component" into the core of "system design". The industry is at a turning point, moving from traditional Ethernet architectures toward a dedicated AI fabric.
Reliability is money: data centers used to tolerate link flaps, but in synchronous parallel AI training, a single link glitch leaves the whole cluster waiting. This will push equipment vendors from "low-price strategies" toward "carrier-grade reliability".
The CPO serviceability paradox: the industry agrees on CPO's performance, but the risk of "replacing the whole board when it fails" remains a pain point for hyperscalers. That is exactly why Lumentum insists the external laser source (ELS) must be pluggable.
Looking ahead to 2030, we expect a "hybrid interconnect" data center: detachable CPO for short reach, coherent optics for longer reach, and OCS passive optical switching dominating the middle tiers to maximize energy efficiency.
Comments