HGX baseboard · NVLink fabric
Why 5 + 4 + 4 + 5 and not 4 × 4
Each GPU has 18 NVLinks. Every switch chip has a fixed number of ports. Pick a layout and see which ones fit. Tap a GPU to trace its links, then tap another to pick its peer.
Line thickness is the number of links between that GPU and that switch. Each switch's bar is its port budget.
GPU-facing ports
External ports
Free
Over budget
How a transfer between two GPUs uses the planes
The board layouts come from NVIDIA's Fabric Manager user guide : on HGX H100 each GPU sends 5, 4, 4 and 5 links to four NVSwitches, and on HGX B200 it sends 9 to each of two. The H100-side port numbers come from Alexander Ishii and Ryan Wells's Hot Chips 34 talk on NVIDIA's NVLink Network switch , which gives 64 ports per switch and shows DGX H100 wiring 20, 16, 16 and 20 of them out to 18 OSFP cages. That's 72 external links against 144 GPU-facing ones, hence the 2:1 taper. B200's switch has 72 ports and uses all of them. NVIDIA hasn't said the external ports are the reason H100 needs four switches; that part is just reading the numbers. Per-link speed is 25 GB/s per direction on H100 and 50 GB/s on B200.