Cloud provider Lambda is using Nvidia co-packaged optical switches to connect more than 10,000 GPUs in an AI training cluster. The deployment links GB300 NVL72 racks and moves the optical conversion closer to the switch chip, shortening the electrical path and removing retimer components.
Lambda’s tests found network latency fell by about 4% at rack scale and by as much as 10% at pod scale while maintaining the throughput of conventional pluggable optics. Each switch has 144 ports running at 800 gigabits per second. The company measured a median optics temperature of 36 degrees Celsius, compared with 51 degrees for traditional pluggable modules.
Nvidia says its photonics design can deliver up to five times better power efficiency and up to 15 times higher reliability than pluggable optics, while using four times fewer lasers. Those are vendor claims, and the announcement does not provide an independent benchmark or a full cost comparison.
The system places 18 removable light sources outside each switch package, allowing technicians to replace a failed source in the field. For large clusters, where one broken network link can leave thousands of GPUs waiting, lower power use and easier repairs could translate into more useful compute time.