Benchmarks / Diffusion /Entry / BF16
Diffusion · Entry · BF16
A 4B-class model at BF16 on 12 GB. The universal baseline: every supported NVIDIA GPU from an RTX 3060 12 GB upward runs this identically, so it is the one leaderboard that spans the whole hardware range.
Best Seconds / step
—s/step
Ranked results
0
0 distinct systems
GPU models
0
Minimum VRAM
12 GB
~14 min run
Best result per GPU
Seconds / step · s/stepNo results yet for this view.
Show the values behind this chart (0 rows)
| GPU | GPUs | Samples | Best (s/step) | Mean (s/step) |
|---|
Leaderboard
Including unranked results →No results match these filters yet.
What is pinned
Everything below is fixed by the profile and checked on every submission. A result whose runtime flags differ from these is published, but never ranked.
Model
- Repository
- stabilityai/stable-diffusion-xl-base-1.0
- Revision
- pending freeze
- Precision
- bf16
- Parameters
- 4 B
- Licence
- CreativeML-Open-RAIL-M++
Runtime
- Engine
- diffusers 0.39.0
- Harness
- diffusion-adapter
- torch_dtype
- float16
- enable_attention_slicing
- false
- enable_vae_slicing
- false
- enable_model_cpu_offload
- false
- torch_compile
- false
Workloads
- t2i_1024
- width=1024 height=1024 steps=30 scheduler=DPMSolverMultistep guidance_scale=5
Ranking
- Ranked on
- Seconds / step (s/step, lower is better)
- Gate
- Error rate ≤ 0
- Gate
- CPU offload ≤ 0
- Also shown
- Images / min, Per image p50, Per image p95, Energy / image, Pipeline load
Design notes
- Release candidate. The runtime and harness versions are pinned; the model revision is not frozen until the bake-off confirms fit on the tier's minimum-VRAM reference system, so this leaderboard is provisional.
- Ranked on seconds per denoising step: it is resolution- and step-count-normalised, so it stays comparable if a future profile changes step count. Images per minute is derived from it and is the headline shown to readers.
- Diffusion is compute-bound where the LLM profiles are bandwidth-bound. This is deliberate: it is the counterweight that stops the site being one memory-bandwidth chart in three costumes.
- CPU offload is disabled and gated. A run that offloads is published but never ranked, because it measures the host as much as the GPU.