DGX SuperPOD switch counts
The published figures, and what a scalable unit does and does not settle.
NVIDIA's DGX SuperPOD reference architectures define a scalable unit (SU) as the block a cluster grows in. For H100 and B200 class systems an SU is 32 DGX nodes, or 256 GPUs, and one SU's compute fabric is 8 leaf and 4 spine switches.
| Scalable units | Nodes | GPUs | Leaf | Spine |
|---|---|---|---|---|
| 1 | 32 | 256 | 8 | 4 |
| 4 | 128 | 1,024 | 32 | 16 |
Those two rows are what this tool reproduces in its test suite. A sizing engine that cannot match a published reference architecture should not be trusted on a design that has no published reference.
Why 8 leaves for 8 rails
Per rail, a 32-node SU needs leaf capacity for 32 node links. One 64-port leaf covers that with room to spare, so each rail gets one leaf and the SU gets eight, one per rail, exactly as rail-optimized design requires. Scale to four SUs and the node count per rail goes to 128, which needs 32 downlink ports × 4, or four leaves per rail. Eight rails × four leaves gives 32.
B300 Spectrum-X is a different shape
The Ethernet reference for DGX B300 is not the same design with different badges. It uses 64 nodes per scalable unit and 2×400GbE per GPU across two independent planes, so 16 NICs per node instead of 8, and it reaches a 512-node SuperPOD without a third tier. Size it as a single-plane 16-rail fabric and you get a valid topology that is not the reference design.
What the reference does not settle
Published architectures cover the compute fabric in detail and are much thinner on everything else. Storage, in-band management and out-of-band management still have to be sized. Out-of-band scales with everything: every node BMC, every switch in every other fabric, every PDU. It is routinely the largest port count in the cluster and the line most often missing from a quote.
Related
- Fat-tree sizing and tier boundaries
- Server specifications — GPU, NIC and power figures per platform.
- NVIDIA DGX SuperPOD documentation