back-end NICs per training node = 3.2 Tb/s/node; 1:1 non-blocking 8-rail fat-tree
8× 400 Gb/sobserved
| Value kind | observed — Reported measurements, counts, and specifications keep the precision and scope stated by their source; an exact specification is not treated as a range. |
|---|---|
| Scope | The GB200/HGX case: one 8-rail training reference design, not a workload-name mandate. The source describes an 8-GPU HGX node (8×400 Gb/s per node); a GB200 NVL72 compute tray holds 4 GPUs (4×400G per tray). Appendix B's counts use one 400 Gb/s back-end attachment per GPU. GB300 NVL72 uses an 800G ConnectX-8 per GPU (two 400GbE planes) and Vera Rubin 1.6 Tb/s per GPU, so never size those fabrics at 400G per GPU. Validate the port rate, rail count and blocking ratio against the target traffic matrix, topology, placement, failure headroom and step-time target. |
| As of | 2025 |
| Source | SemiAnalysis AI Neocloud Playbook |
| Review | checking…review by 2027-04-30 · standard cadence |
| Recorded changes | last 2026-06-29 |
| Claim id | back-end-nics-per-training-node-3-2-tb-s-node-1 |
Where the guide uses it
← Full numbers register — every date-stamped figure in the guide, with revision history.