scale-up vs scale-out per-GPU bandwidth by generation, per direction
~18x (GB200, 400G) / ~9x (GB300, 800G) / ~7.5x (Vera Rubin, 1.6T)derived
| Value kind | derived — Derived values are guide calculations; their result is only as fixed as the stated inputs, scope, and method. |
|---|---|
| Scope | Peak interface-rate ratios (per GPU, one direction), not measured collective throughput. NVIDIA's Vera Rubin NVL72 product page lists NVLink 6 at 3.0 TB/s per GPU and 216 TB/s per rack; 3.6 TB/s and 260 TB/s were the figures NVIDIA announced in January 2026. |
| As of | 2026-10-05 |
| Source | NVIDIA NVLink, GB300 NVL72 and Vera Rubin NVL72 product pages; NVIDIA Technical Blog (2026-01-05) for ConnectX-9 · Blackwell platform feature list: fifth-generation NVLink and networking specifications |
| Review | checking…review by 2027-04-05 · standard cadence |
| Recorded changes | last 2026-10-05 · 2 revisions tracked |
| Claim id | scale-up-nvlink5-gpu-vs-scale-out-per-nic |
Where the guide uses it
← Full numbers register — every date-stamped figure in the guide, with revision history.