Unlock Full Tensor Core Performance on Your NVIDIA GPU
NVIDIA sells two kinds of graphics cards built on the same silicon: a GeForce card for gamers and a professional RTX A-series card for workstations. Both use identical tensor core hardware, but the GeForce version leaves the factory with a speed selector set to half rate for FP16, BF16 and TF32 math with FP32 accumulation. Those are exactly the precisions used by nearly all AI training, fine-tuning, image generation and unquantized inference. The gpu-unlock project removes that limitation at boot time, before the stock NVIDIA driver ever touches the device. No patched driver, no VBIOS flash, and nothing to redo whenever NVIDIA ships a new driver release.
Why Consumer GeForce Cards Run at Half Rate
The restriction is not a hardware defect and it is not something the driver decides. Each GPU carries a hidden speed selector in its firmware and VBIOS ROM. On consumer parts, that selector is programmed to half rate, which means the tensor cores still execute every instruction correctly, they simply issue at a reduced cadence for a defined set of matrix math modes. Professional cards carry the same silicon with the selector set to full rate. This is product segmentation implemented in firmware, and it is the reason a GeForce card can feel disproportionately slow compared with a workstation card of the same generation.
How gpu-unlock Removes the Cap
The module runs from the initramfs ramdisk during early boot. It reads the firmware parameters, flips the speed selector from half rate to full rate, then triggers a PCIe Function Level Reset so the device reinitialises with the new setting applied. Only after that handover does the official NVIDIA driver bind to the card. Because the change happens before driver initialisation, it survives driver updates. There is no binary patch to reapply, no signed module to rebuild with each kernel release beyond a normal recompile, and no modified VBIOS sitting on the card. If you boot without the module, the GPU simply behaves like a stock card.
Which GPUs Are Supported
Development and stability testing focus on the RTX 30 series, specifically the GA102, GA103, GA104 and GA106 chips. Experimental support also covers selected RTX 20 series and RTX 40 series desktop and laptop GPUs. Cards that are already uncapped, such as the professional RTX A-series, gain nothing because there is no cap to remove. The table below summarises the current situation.
| Generation | Typical chips | Status |
|---|---|---|
| RTX 30 series | GA102, GA103, GA104, GA106 | Fully supported, primary target |
| RTX 20 series | TU106 and similar | Experimental support |
| RTX 40 series | Desktop and laptop models | Experimental support |
| Professional RTX A-series | Same silicon, full rate by default | No change needed |
Measured Performance Gains
The gains show up where you would expect them, in dense matrix multiplication. On an RTX 3090 Ti, FP16 throughput rises from roughly 87 to more than 118 TFLOPS, BF16 climbs from about 87 to 157 TFLOPS, and TF32 moves from around 42 to over 82 TFLOPS. Across the supported consumer cards, dense GEMM throughput improves by roughly 25 to 80 percent depending on the precision and the model shape. In practice that means faster prefill for large language models, shorter fine-tuning runs, quicker image generation steps and more headroom when you keep weights in BF16 or FP16 instead of quantising to INT8.
| Precision with FP32 accumulation | Stock TFLOPS | Unlocked TFLOPS | Gain |
|---|---|---|---|
| FP16 | 87 | 118 | about 35 percent |
| BF16 | 87 | 157 | about 80 percent |
| TF32 | 42 | 82 | about 95 percent |
Where You Will See No Difference
The unlock targets tensor matrix core operations only. It does nothing for conventional rasterised gaming workloads, scalar FP32 instructions, INT8 quantised inference or video decoding. Those paths are limited by memory bandwidth, cache behaviour or fixed function hardware, not by the tensor core cadence selector. If your pipeline is already fully quantised to INT8, expect no change at all. The same applies to workloads that are dominated by data movement rather than arithmetic intensity.
Boot Configuration Notes
Sequence matters. The module must load before the NVIDIA driver claims the device, which in practice means adding it to the initramfs and ensuring the correct ordering against the proprietary driver's own early load hooks. Distributions that build the initramfs with dracut, mkinitcpio or update-initramfs all handle this differently, so follow the project's own instructions for your distribution. Secure Boot setups need the module signed with an enrolled key, exactly as you would sign any other out-of-tree kernel module. Verify the result by checking the reported tensor throughput rather than assuming success.
Is It Worth Using
If you train, fine-tune or generate images on a consumer GeForce card in BF16, FP16 or TF32, the unlock is one of the cheapest performance wins available, because it costs nothing at runtime and adds no per-driver maintenance. If you mostly game, decode video or run INT8 quantised inference, skip it. Read the repository documentation, confirm your chip is covered, and benchmark before and after so you know exactly what you gained.
Let's work together
Do you need more info, help with your project, or to develop an idea?
Whether it's an easy question, a quick doubt, or just a 5-minute chat, send me a message—it costs nothing and I'm always ready to help. I love discussing a problem to understand it, getting creative with solutions, and focusing on simple, reliable, and straightforward ideas that we can actuate quickly.
Contact me →