stepfun-ai/Step-Video-T2V (30B) · text-to-video
| Condition · Value | |
|---|---|
| GPU | RTX 3090 24GB |
| Power cap (W) | 420 |
| Resolution | 992x544 |
| Frames | 51 |
| Steps | 30 |
| Seed | 42 |
| Execution engine | DiffSynth tiled-VAE, resident/tiered |
| Metric · Value | |
|---|---|
| Wall time, median (s) | 1730 |
| Wall time (s) | 1730.3 |
| Peak VRAM (GB) | 12.48 |
| Peak host RAM (GB) | 96.0 |
| Peak host RAM (GB) | 96.0 |
| Load time (s) | 151.6 |
| Repeat count | 1 |
| Render stability | STABLE (1.01, motion 4.0) |
★30B flagship on a single 24GB 3090, no OOM: peak VRAM only 12.48GB (weight-tiering) BUT peak host RAM 96.0GB. wall 1730.3s (~29min) single run. Demonstrates the core capacity claim (30B video DiT on 24GB). ★buyer spec: 24GB GPU + ~96GB host RAM. 1 run; note 'optimus lib skipped (OSError)'.
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.