Lambda Boosts AI Factory Resiliency and Efficiency With NVIDIA Photonics
Lambda is deploying NVIDIA Photonics co‑packaged optics (CPO) switches to interconnect over 10,000 GPUs across NVIDIA GB300 NVL72 racks, achieving up to a 10% latency reduction at pod scale and up to 5× better power efficiency with 15× higher reliability versus conventional pluggable optics. Each switch provides 144 × 800 Gb/s ports, uses liquid cooling to keep median optics temperature at 36 °C (versus 51 °C for pluggable optics), and features 18 removable external light sources for field‑replaceable lasers. The co‑packaged design shortens electrical paths from inches to millimeters, eliminating retimer chips and cutting power draw while improving signal integrity. Engineers can leverage this architecture to build more resilient, power‑efficient AI training clusters that minimize downtime and energy waste.
⚡ Key Takeaways
- Network latency dropped ~4% at rack scale and up to ~10% at pod scale in Lambda’s tests.
- Switches use NVIDIA Photonics CPO with 144 ports at 800 Gb/s each.
- Power efficiency improves up to 5× and reliability up to 15× compared to conventional pluggable optics.
- Each switch includes 18 removable external light‑source modules for field replacement.
- Co‑packaged optics reduce the electrical path to millimeters, removing retimer chips and lowering component count.
The reduced latency, power draw, and dramatically higher MTBI enable larger AI training jobs to run with fewer interruptions and lower operational cost, directly benefiting production AI workloads that depend on massive GPU clusters.
✅ Practical Steps
- Deploy NVIDIA Photonics CPO switches (144 × 800 Gb/s) to interconnect NVIDIA GB300 NVL72 racks in your AI cluster.
- Implement liquid cooling for the switch packages to maintain optics temperatures around 36 °C.
- Use the removable external light‑source modules for field maintenance instead of replacing entire pluggable optics.
Want the full story? Read the original article.
Read on NVIDIA Blog ↗