Lightricks/LTX-Video · text-to-video
| Condition · Value | |
|---|---|
| GPU | RTX 3090 24GB |
| Power cap (W) | 420 |
| Resolution | 768x448 |
| Frames | 49 |
| FPS | 24 |
| Steps | 30 |
| Seed | 42 |
| Precision profile | bf16 (AUTO, resident) |
| Warm run | no |
| Metric · Value | |
|---|---|
| Wall time, median (s) | 18.5 |
| Wall time, min (s) | 17.4 |
| Wall time, max (s) | 20.4 |
| Wall time, stdev (s) | 0.97 |
| Peak VRAM (GB) | 9.55 |
| Peak VRAM spread (GB) | 0.0 |
| Peak host RAM (GB) | 18.7 |
| Peak host RAM (GB) | 18.7 |
| Load time, median (s) | 7.4 |
| Observed peak power draw (W) | 418.78 |
| Repeat count | 5 |
| Render stability | STABLE (0.94, motion 0.9) |
VERIFIED commercial-condition measurement (INC-002 closed): physical RTX 3090 GPU1 at a CONFIRMED 420W cap (Windows '\GPU1_300W' enforcer disabled; GPU1 measured drawing 418.78W @ 100% util mid-render, which a 300W cap makes impossible). 5 runs: wall 18.5s median (range 17.4-20.4, stdev 0.97s), peak VRAM constant 9.55GB (0.000 spread), load ~7.4s. Supersedes the earlier unverified 18.7s figure — the true verified-420W number is 18.5s. Still internal (0 external reproductions) — no comparative/SOTA claim until reproduced on >=2 outside environments per public_gate.
Run the same fixed configuration on your card and open a result PR — see Run it yourself. Raw run manifests are attached to the internal evidence record and are released alongside external reproduction tooling.