Plainbrush / Benchmarks / NVIDIA GeForce RTX 3060 12 GB

RTX 3060 12 GB: stock vs. tensor-core unlock

NVIDIA's GeForce RTX 20, 30 and 40 cards run one kind of AI math (16-bit inputs with 32-bit accumulation) at half the rate of NVIDIA's workstation cards built on the same chip. An open-source tool released in October 2026 lifts that limit until the next full power-off, without modifying the card's firmware. We measured what it changes for real image generation on the card Plainbrush runs on today.

SDXL image, stock
20.6 s
SDXL image, unlocked
13.9 s
Difference
49% faster
JobStockUnlockedDifference
SDXL 1.0
1024 × 1024 image, 30 steps
20.6 s13.9 s49% faster
Stable Diffusion 1.5 (Realistic Vision 6)
512 × 768 image, 30 steps
5.3 s3.9 s36% faster

Time to generate one image; lower is better.

What it means for users

  • Faster image and video generation on a card people already own, with no change to the results: the images are the same, only the time changes.
  • Community reports elsewhere show about 80% more peak AI throughput on paper and up to 59% faster image generation 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.
  • Games and 3D path-traced rendering do not use this kind of math and are unaffected.
  • The effect lasts until the computer is fully powered off; it does not permanently change the card.

How we measure

Every card runs the same two ComfyUI image jobs with identical settings: same model, prompt, size, sampler (Euler, CFG 6) and fixed random seeds, one image at a time. Each job runs once to warm up and then three times; we report the average generation time. Repeat runs on the same card agree within about 1%, so differences larger than that are real.

The tensor-core unlock is a third-party, open-source tool. It is not made or endorsed by NVIDIA, and Plainbrush does not ship, install or require it. We test it so users can see what their own hardware can do.

Test system: NVIDIA GeForce RTX 3060 12 GB, PyTorch 2.11 (CUDA 12.8), ComfyUI. Updated 7 November 2026.