NV UEFI Unlocker aims to remove GeForce RTX tensor performance limits

Chinese modder physicsdolphin has released NV UEFI Unlocker with Windows support. The tool aims to remove the tensor computing limit on GeForce RTX 20, RTX 30 and RTX 40 series graphics cards. According to reports about the development, which came to light on 30 October, it also works on Linux and runs at the UEFI stage, before the NVIDIA driver loads. The project aims to deliver higher compute performance for AI workloads.

What limit does the tool aim to change?

According to TechPowerUp, FP16, BF16 and TF32 tensor operations on GeForce cards run at half speed with FP32 accumulation, while the same silicon in professional cards runs at full speed. NV UEFI Unlocker aims to remove this limit before the driver starts. The process does not require rewriting the VBIOS.

Playground.ru describes the tool as open source and says the change resets when the computer is fully shut down. The performance gain is therefore not a permanent change to the card. The tool is launched from a USB drive prepared with the UEFI application and required components; Secure Boot must be disabled or the EFI application set as trusted. For RTX 20 series cards, the relevant firmware mode for GSP must also be enabled beforehand, according to the report.

What gains have early tests reported?

In initial results reported by Playground.ru, the RTX 2070’s FP32 matrix multiplication performance rises from 36 TFLOPS to a range of 60–63 TFLOPS. According to the source, depending on the test conditions, that represents an increase of approximately 67–75%. FP16 performance on a mobile RTX 4060 is also reported to rise from 29.89 TFLOPS to 49.33 TFLOPS.

These figures should not be presented as results verified by the tool’s developer or an independent lab and reproduced across all systems. Playground.ru stresses that results may vary by model and computer configuration, and that a theoretical increase in compute power does not guarantee the same speedup in real-world applications. Memory bandwidth and power limits can also affect the results.

Compatibility and usage considerations

The tool targets the RTX 20, RTX 30 and RTX 40 series, though the source notes that compatibility issues remain for some models. GTX 16 support is experimental, while the RTX 50 series is not supported. So compatibility should not be assumed for every card based on series name alone.

Playground.ru reports that the earlier GPU Unlock tool nearly doubled tensor throughput in some calculations and delivered an increase of around 60% in AI and image-generation tests. However, these results should not be treated as an indication of the new tool’s performance across all scenarios. The sources also note that there is no convincing evidence of improved gaming performance or faster DLSS.

Impact on architecture and visualization workflows

Architecture firms and visualization artists running AI-based image generation or similar tensor computations on local graphics cards may want to explore whether they can get more out of compatible older-generation cards. However, this does not mean every render or design application will run faster; the tests cited by the sources focus on specific types of computation.

Before adding the tool to a workflow, check graphics card compatibility, the operating system in use and the Secure Boot requirements. Since the tool is experimental, a cautious approach would be to test it on a separate system before using it on workstations critical to production, and to measure actual runtimes in the applications used in the workflow. The project is reported to be open source, but the sources do not provide details about its license terms.