Scale-out
Connecting many nodes into a looser cluster fabric (InfiniBand or Ethernet) to scale beyond one coherent domain.
Current numbers
1.0.3, released 2026-07-16UEC scale-out specification; contract the implemented profile
~18× GB200 / ~9× GB300NVLink 5 per-GPU scale-up (~900 GB/s per direction) against one per-GPU scale-out NIC — ~18× on GB200/CX-7/400G, ~9× on GB300/CX-8/800G; keep collectives in scale-up
up to 1.6 Tb/sper-GPU scale-out bandwidth on the Rubin platform — two 800 Gb/s ConnectX-9 SuperNICs per GPU, not one 1.6T NIC; double the ConnectX-8 generation; BlueField-4 (integrates a CX-9) at 800 Gb/s