NVLink 5 per-GPU one-way scale-up bandwidth versus one 1:1 per-GPU scale-out NIC: ~18× on GB200/CX-7/400G and ~9× on GB300/CX-8/800G
~18× GB200 / ~9× GB300derived
| Value kind | derived — Derived values are guide calculations; their result is only as fixed as the stated inputs, scope, and method. |
|---|---|
| Scope | Nominal directional endpoint ratios: 900 GB/s divided by 50 GB/s for the specified GB200 profile; divided by 100 GB/s for the specified GB300 profile. Not collective or application speed ratios. |
| As of | 2026 |
| Source | Guide, endpoint-ratio arithmetic from NVIDIA NCX reference architecture: NVLink 5 and named GB200/CX-7 or GB300/CX-8 profiles. · NCP GB200/GB300 Components table, scale-out NIC paths: four 400 Gb/s ConnectX-7 per four GB200 GPUs; four 800 Gb/s ConnectX-8 per four GB300 GPUs — measured against the separate scale-up NVLink domain, where Enterprise RA gives up to 1,800 GB/s per GPU. |
| Derivation | 900/50 = 18; 900/100 = 9. Keep platform-specific scale-out profiles distinct. |
| Review | checking…review by 2026-09-30 · standard cadence |
| Recorded changes | last 2026-09-16 · 3 revisions tracked |
| Claim id | nvlink5-per-gpu-bw-1-8-tb-s-bidirectional-900 |
Where the guide uses it
- 1.2 Training Data Centers: Synchronous, Dense, Checkpointable
- 10.2 Topology-Aware & Rack-Scale Scheduling
← Full numbers register — every date-stamped figure in the guide, with revision history.