NVIDIA BlogT1·Infra
72 pts
Lambda has deployed NVIDIA Photonics co-packaged optics (CPO) switches to interconnect NVIDIA GB300 NVL72 racks spanning more than 10,000 GPUs, forming a large-scale AI factory network. Tests show roughly 4% lower rack-scale latency and up to about 10% lower pod-scale latency, with throughput matching conventional pluggable optics while cutting power draw and extending mean time between interruptions. It marks one of the first production-scale CPO deployments at the 10,000-GPU level, offering a reusable engineering blueprint for next-generation AI cluster interconnect.