NVIDIA reference read target for computer-vision training on datasets over about 30 TB; no single per-GPU floor: B300 Standard is ~0.14-0.16 GB/s/GPU and Enhanced ~0.43-0.49, depending on the reference architecture
~4 GB/s/GPUrecommendation
| Value kind | recommendation — Recommendations are design guidance, not measured facts; validate them against the project boundary conditions. |
|---|---|
| Scope | Published reference-architecture guidance, not a measured requirement. The 4 GB/s per GPU computer-vision line appears in NVIDIA's H200, B300 and Vera Rubin reference architectures. Keep shared-tier and single-client figures on separate axes: NVIDIA's per-scalable-unit (shared-tier) read tiers work out to ~0.06–0.49 GB/s per GPU (H200 Good to B300 and Vera Rubin Enhanced), while its single-client H200 guidance runs 0.5–5 GB/s per GPU. NVIDIA says the tiers assume a mix of workloads. Derive the target from the loader's measured consumption rate and use these as comparison cases. |
| As of | 2025-11 |
| Source | NVIDIA DGX SuperPOD B300 RA, Storage Architecture (4 GB/s per GPU of read performance for automotive and computer-vision workloads with datasets over 30 TB); the H100 RA gives the same figure |
| Review | checking…review by 2027-03-01 · standard cadence |
| Recorded changes | last 2026-10-16 · 2 revisions tracked |
| Claim id | nvidia-reference-read-target-for-vision |
Where the guide uses it
← Full numbers register — every date-stamped figure in the guide, with revision history.