SEMICON 2025 Silicon Photonics Summit | Lumentum

3 days ago

2 min read

Introduction

As AI models keep growing, data center bandwidth and energy-efficiency demands are under unprecedented pressure. Simply stacking more hardware is not enough—technical innovation is the key. At the SEMICON 2025 Silicon Photonics Summit, Lumentum shared its latest breakthroughs in cost-effective and power-efficient technologies.

The talk covered optical circuit switching (OCS), high-power lasers, high-speed indium phosphide (InP) devices, and VCSELs (vertical-cavity surface-emitting lasers) in scale-up networks. The common goal of these technologies is to support AI training and inference across hundreds of thousands of GPUs while reducing power consumption.

Content

1. Optical Circuit Switch

Lumentum's optical circuit switch is based on MEMS mirrors and supports a 300×300 port configuration.
Main applications:
- Spine switch replacement
- Network redundancy
- Failover
Advantages:
- Insertion loss < 2 dB (all channels)
- Return loss < –55 dB
- Extremely low polarization-dependent loss
Power savings: can cut front-end network power by more than 65%, especially well suited to large-scale GPU cluster environments.

2. High-Power Lasers

Working with the NVIDIA Spectrum-X CPO switch, Lumentum launched high-power lasers designed specifically for CPO (co-packaged optics).
Technical results:
- Output >400 mW per laser
- Wall-plug efficiency >20%
- Low noise and narrow linewidth, suited to frequency-sensitive modulator applications.
Significance: enables moving the laser to the front panel to drive multiple optical engines, cutting power by >40% compared with conventional pluggable modules.

3. High-Speed Indium Phosphide (InP) Devices

InP devices remain the technology of choice for high-speed transceivers thanks to long volume-production experience.
Lumentum milestones:
- Launched 40G/25G DMLs in 2000
- Now in volume production of >100G and 200G DMLs
- Unveiled 400G DMLs and DFB Mach-Zehnder lasers at OFC 2025
Advantage: stable, high-speed datacom connectivity, especially suited to front-end network and scale-out needs.

4. VCSELs and Scale-Up Networks

Today's NVL-72 GPU rack has 7.2 Tbps of back-end bandwidth, which future generations will double to 14.4 Tbps and 28.8 Tbps.
Challenges:
- The faster copper runs, the shorter its reach.
- GPU counts keep expanding, from 72 → 144 → 500+.
- Power soars accordingly, with a single rack exceeding 120 kW.
Lumentum's solution:
- Bring 3D sensing technology (pixel stacking, an extension of Face ID technology) into scale-up networks.
- Use long-wavelength VCSELs, which offer:
  - High-temperature operation (>155 °C)
  - Better reliability (lifetime >100 years)
- VCSELs already come with built-in advanced packaging capabilities (2.5D, 3D), suiting future high-bandwidth applications.

Conclusion

In its SEMICON 2025 talk, Lumentum clearly laid out its technology strategy for power savings and cost-effectiveness:

Using optical circuit switches to sharply reduce front-end network power.
Combining high-power lasers with CPO to support future hyperscale GPU clusters.
Continuing to deepen its volume-production advantage in high-speed InP devices.
Innovatively bringing VCSELs and 3D sensing technology into AI scale-up networks to improve bandwidth, temperature tolerance, and reliability.

What these technology paths share is balancing performance and power efficiency amid ever-growing AI compute demand, while also accounting for cost and reliability—building truly sustainable infrastructure for the next generation of data centers.