Worked examples
Ten complete builds, generated by the engine rather than described. Five GPU counts, each done twice: once on nodes, once on NVL72 racks.
Each page below is a full bill of materials, produced by running the sizing engine at build time. The numbers are the tool's own output, not a summary of it, and every quantity shows the calculation behind it.
The pairing is the useful part. At each GPU count there is a node-scale build and a rack-scale NVL72 build, so the rounding, the switch counts and the power draw are directly comparable. NVL72 cannot be bought in units smaller than a 72-GPU rack, which is visible immediately at the 64-GPU target.
| Build | GPUs | Nodes | Compute switches | Tiers | Racks | kW |
|---|---|---|---|---|---|---|
| 64 GPUs — DGX H100, InfiniBand | 64 | 8 | 1 | 1 | 3 | 86 |
| 64 GPUs — GB300 NVL72 | 72 | 1 | 1 | 1 | 2 | 129 |
| 128 GPUs — DGX B200, Cisco Nexus | 128 | 16 | 1 | 1 | 7 | 239 |
| 128 GPUs — GB200 NVL72 | 144 | 2 | 8 | 2 | 3 | 253 |
| 256 GPUs — one DGX H100 SuperPOD SU | 256 | 32 | 12 | 2 | 9 | 342 |
| 256 GPUs — GB300 NVL72 | 288 | 4 | 6 | 2 | 5 | 524 |
| 512 GPUs — DGX B300, Spectrum-X twin-planar | 512 | 64 | 24 | 2 | 34 | 866 |
| 512 GPUs — GB200 NVL72 | 576 | 8 | 27 | 2 | 10 | 999 |
| 1,024 GPUs — four DGX H100 SuperPOD SUs | 1,024 | 128 | 48 | 2 | 34 | 1,359 |
| 1,024 GPUs — GB300 NVL72 | 1,080 | 15 | 27 | 2 | 19 | 1,981 |
Related
- DGX SuperPOD switch counts — the published figures the 256 and 1,024 GPU builds reproduce.
- GB200 and GB300 NVL72 networking
- Fat-tree sizing and tier boundaries