Double Your GPU Speed: RTX 3000 Cards Had 50% Locked Away
Six years ago, NVIDIA shipped the RTX 3090 with BF16 tensor math running at half speed. Not because the hardware couldn't go faster—because a kernel setting kept it throttled. On Tuesday, developer Zagitta released gpu-unlock, a Linux kernel module that flips the switch before NVIDIA's driver loads. The card wakes up.
People who train LoRAs, generate images, or run local music models are scrambling to test it. On the same hardware, same driver, same power limit, users measured BF16 tensor throughput jump from 82 to 155 TFLOPS—almost 1.9× faster. Byte-for-byte outputs match the stock run, confirming it's real math, not different computation.
Why NVIDIA locked it
The RTX 3090 is a consumer card. The RTX A100 is a workstation card with a workstation price tag. Half-speed BF16 on the consumer side meant the workstation card looked worth the money. After six years of LoRA training running through gimped math, even this small unlock stings.
What actually gets faster
Raw tensor peaks hit nearly 2× on synthetic benchmarks. Real-world numbers are quieter but real. On Qwen models in inference, prefill improved 10–12 percent. LLM decode (the part people wait for) barely budges because it's bottlenecked by memory bandwidth, not compute. But image generation, speech synthesis, and any workload heavy on BF16 matrix multiplies sees the lift. One user noted LoRA training on a 3090 took exactly the path that was cut.
Independent tests on Windows using a UEFI port of the same method show similar 1.8–2× gains on BF16 and FP16 tensor math with FP32 accumulation.
How to use it
Linux: gpu-unlock by Zagitta loads before the driver and sets FMLA32 to full. Windows users have ports available from projects like impacifist and bjrasmussen-jpg. The module does not change clocks, power limits, or permanently alter the card. Cold boot required.
A few caveats: it's unofficial, stability testing is ongoing, and NVIDIA has made no comment. Community reports suggest it works most reliably on 30-series cards. Some users report clean installs work better than upgrades. If you train or generate images on a 3090 or 3080, the math is worth ten minutes to test.
What doesn't change
FP32 performance stays the same. Gaming stays the same. The cards still draw 300W. But for creators working in ComfyUI, training frameworks, or local audio/music models that rely on BF16, a card you bought six years ago just became 10–20 percent faster at its real job.