NVIDIA BlogT1·72 pts
- Lambda interconnected over 10,000 GB300 NVL72 GPUs using NVIDIA CPO switches in a large-scale production deployment
- Rack-scale latency fell about 4% and pod-scale up to about 10%, with throughput on par with pluggable optics
- CPO cuts components in the signal path, lowering power draw and extending mean time between interruptions