Keep the lowest precision that clears the quality contract — conditional result
Select FP8: 89.4% at 1,200 tokens/s and 45 ms. NVFP4 fails quality; 8,900 correct items is its flip threshold.derived
| Value kind | derived — Derived values are guide calculations; their result is only as fixed as the stated inputs, scope, and method. |
|---|---|
| Scope | Baseline accuracy = 9,000/10,000 = 90.0%. FP8 = 89.4%, a 0.6-percentage-point drop; NVFP4 = 88.8%, a 1.2-point drop. Both synthetic runs clear throughput and latency, but only FP8 clears 89.0%. NVFP4 must reach at least 8,900 correct items on this fixed test, an increase of 20, while retaining the other gates. This threshold concerns percentage points, not relative percent accuracy. |
| As of | 2026-09 |
| Source | NVIDIA, Introducing NVFP4 — method reference; guide-authored hypothetical scenario September 8, 2026. No supplier quote or test measurement. |
| Derivation | Baseline accuracy = 9,000/10,000 = 90.0%. FP8 = 89.4%, a 0.6-percentage-point drop; NVFP4 = 88.8%, a 1.2-point drop. Both synthetic runs clear throughput and latency, but only FP8 clears 89.0%. NVFP4 must reach at least 8,900 correct items on this fixed test, an increase of 20, while retaining the other gates. This threshold concerns percentage points, not relative percent accuracy. |
| Review | checking…review by 2027-03-08 · standard cadence |
| Recorded changes | last 2026-09-16 |
| Claim id | guide2-t29-7-10-result |
Where the guide uses it
← Full numbers register — every date-stamped figure in the guide, with revision history.