Cards tested
Current results
NVIDIA GeForce RTX 3060 12 GB
Stock SDXL: 20.6 s · Unlocked: 13.9 s (49% faster)
Stock SD 1.5: 5.3 s · Unlocked: 3.9 s (36% faster)
Other cards
Not measured yet. We own one test GPU. If you run ComfyUI on a different card and want to run the same benchmark, write to us; results go here with your setup.
See the Roadmap for the cards we would like to test next.
Methodology
How we test
Identical jobsSame models, prompts, sizes, sampler (Euler, CFG 6), seeds. One image at a time.
Warm-up + 3 runsFirst run discarded. Average of 3 reported. Repeat runs on same card agree within ~1%.
Same machineCPU, RAM, driver, PyTorch, CUDA identical. Only the GPU changes.
Stock vs unlockedTensor-core unlock tool applied, then re-tested. No firmware changes. Effect lasts until full power-off.
Published rawRaw CSV + environment details on each card page. Independent QA verifies reproducibility.
What the unlock changes
Tensor-core unlock
- NVIDIA GeForce RTX 20 / 30 / 40 consumer cards run BF16/TF32 tensor math at half the rate of workstation cards on the same silicon.
- An open-source tool (October 2026) lifts that limit until the next full power-off. No firmware modification.
- Plainbrush does not ship, install, or require this tool. We test it so users can see what their own hardware can do.
- Games and 3D path-traced rendering do not use this math path and are unaffected.
- Community reports show ~80% more peak AI throughput on paper, up to 59% faster image gen on larger cards. Mid-range cards like the 3060 can be held back by their power limit, which is why we measure the real gain ourselves.
Independent QA
Verification gate
Every benchmark passes
- Zero console errors across all pages
- Reproducible within 1% on the same hardware
- Focus trap and 390px overflow gates pass
- Theme system, command palette, star clear, and accessibility verified
- Manifest integrity: hashes and sizes match actual files
QA verdict: PASS (as of 7 Nov 2026)