black-forest-labs/FLUX.1-dev · text-to-image
| Condition · Value | |
|---|---|
| GPU | RTX 3090 24GB |
| Power cap (W) | 420 |
| Resolution | 1024x1024 |
| Steps | 24 |
| Seed | 42 |
| Precision profile | fp8 weight-only (AUTO, resident) |
| Warm run | yes |
| Metric · Value | |
|---|---|
| Wall time, median (s) | 49.25 |
| Wall time, min (s) | 48.8 |
| Wall time, max (s) | 50.5 |
| Wall time, stdev (s) | 0.64 |
| Peak VRAM (GB) | 12.78 |
| Peak VRAM spread (GB) | 0.0 |
| Peak host RAM (GB) | 45.2 |
| Peak host RAM (GB) | 45.2 |
| Load time, median (s) | 15.2 |
| Repeat count | 4 |
VERIFIED commercial-condition FLUX.1-dev benchmark (INC-002 closed): physical RTX 3090 GPU1 at CONFIRMED 420W (enforcer disabled), fp8 weight-only resident. HONEST pooled set = ALL 4 warm runs across the two verified batches (main-batch r2=50.5s + flux3 x1/x2/x3=48.8/49.4/49.1s): wall 49.25s median, full observed warm range 48.8-50.5s, stdev 0.64s (pop) / 0.74s (sample), peak VRAM constant 12.78GB, load ~15s. This SUPERSEDES an earlier note that reported only the 3 flux3 runs (median 49.1, max 49.4) — that under-stated the true warm max (50.5s) and stdev; corrected here per the claim-integrity audit. Excluded: ONE cold first-load run (83.6s wall + 89.7s first-load) as warmup — disclosed, not silently dropped. Verified 420W is faster than the earlier unverified 50.9s figure, as expected once the 300W enforcer no longer capped the card. Still internal (0 external).
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.