Lambda Boosts AI Factory Resiliency and Efficiency With NVIDIA Photonics

What happened
Lambda has deployed NVIDIA Photonics co-packaged optics switches to connect more than 10,000 GPUs in its AI training racks. This integration is designed to improve the efficiency and reliability of the network that links these GPUs. The new switches lower network latency and reduce power consumption compared to previous technologies. The deployment is part of Lambda’s efforts to support large-scale AI model training in its cloud infrastructure.
Key details
- The deployment connects over 10,000 GPUs using NVIDIA GB300 NVL72 racks.
- NVIDIA Photonics switches each have 144 ports operating at 800Gb/s.
- Network latency dropped by about 4% at rack scale and up to 10% at pod scale in Lambda's tests.
- Median optics temperature is 36°C with liquid cooling, compared to 51°C for older pluggable optics.
Why it matters
Cloud providers and organizations training large AI models can now operate GPU clusters with lower energy costs and fewer network disruptions. This reduces the risk of stalled training jobs and allows more power to be allocated to computing instead of networking. Technicians also benefit from easier maintenance, as failed light sources can be replaced without swapping out entire modules, minimizing downtime.
