New tool unlocks up to 75% more AI compute on RTX 20, 30, and 40 GPUs

Unlocking hidden AI power

A new open-source project called NV UEFI Unlocker is making it possible for users to access restricted compute power on NVIDIA graphics cards. The tool specifically targets the Tensor Cores, which are the specialized parts of the chip designed to handle AI and machine learning tasks. By using this tool, owners of RTX 20, 30, and 40 series GPUs can see significant jumps in performance.

As Wccftech reported, early tests show that these performance gains can reach as high as 75% in certain scenarios. For example, an RTX 2070 saw its FP32-accumulation matrix multiplication jump from 36 TFLOPS to roughly 60–63 TFLOPS. These numbers represent a major leap for users trying to run heavy AI models on older hardware.

The project was developed by a GitHub user named physicsdolphin to bring functionality to Windows that was previously mostly limited to Linux. While previous projects showed similar results, they often required complex kernel modules. This new tool makes the process more accessible to a wider range of regular computer users across different operating systems.

How the firmware modification works

The NV UEFI Unlocker works by intervening at a very early stage of the computer's startup process. It uses a UEFI-based approach, which means it modifies how the GPU behaves before the NVIDIA graphics driver even begins to load. This is different from traditional methods that might involve permanently changing the card's internal software.

Because it operates at the firmware level, the tool does not require the user to flash the VBIOS, which is the permanent firmware on the graphics card. This makes the process much safer for the hardware because the changes are not permanent. The unlock is temporary and will automatically disappear as soon as the computer is completely powered off.

To use the tool, a user must create a bootable USB drive containing the necessary UEFI application and firmware components. The user then boots the computer from this USB drive, allowing the tool to execute its instructions. Once the modifications are applied, the system automatically hands control back to the main operating system for normal use.

Compatibility and hardware limits

The tool is designed to work across three major generations of NVIDIA hardware, including the RTX 20, 30, and 40 series. However, the developers note that the project is still experimental and success is not guaranteed for every single setup. Some specific models may still encounter issues during the unlocking process or when trying to run benchmarks.

Users with older GTX 16 series cards might find that support is currently experimental, while the upcoming RTX 50 series is not supported at all. For those using RTX 20 series cards, there is an additional requirement to enable a specific GSP firmware mode. These requirements mean that the tool is best suited for enthusiasts who are comfortable navigating system settings.

The project also requires certain security configurations to be adjusted to work correctly. Specifically, Secure Boot must either be turned off or configured to trust the specific EFI application being used. Without these adjustments, the system may block the tool from making the necessary changes during the boot sequence.

Real-world vs. theoretical gains

While the numbers in synthetic benchmarks are impressive, it is important to understand what these gains mean for daily use. The tool significantly improves FP16 and FP32 compute throughput, which are the math operations used by AI models. For instance, an RTX 4060 Laptop GPU showed a 65% improvement in FP16 compute throughput, moving from 29.89 to 49.33 TFLOPS.

However, higher theoretical throughput does not always result in a perfectly linear increase in real-world speed. The actual performance of an AI model often depends on other factors like memory bandwidth and power limitations. A card might be able to do more math per second, but it might still be slowed down if it cannot move data fast enough to keep up.

It is also worth noting that these modifications do not appear to help with standard gaming. There is currently no evidence that the unlock improves gaming performance or the execution times for DLSS, which is NVIDIA's AI upscaling technology. The benefits are strictly focused on AI inference and other heavy compute workloads that utilize the Tensor Cores.

What to expect next

The release of NV UEFI Unlocker provides a new path for owners of older NVIDIA hardware to stay relevant in the AI era. It allows people to extract more value from their existing investments without having to purchase new, more expensive GPUs. This is particularly useful for hobbyists who want to run local image generation or large language models.

Because the project is open-source and experimental, the community will likely continue to refine it. Users on platforms like Reddit have already begun discussing the results, which may lead to better documentation and broader compatibility. As more people test the tool, the community will gain a clearer picture of which specific models work best and which ones face hurdles.

For now, the tool remains a specialized option for those who understand the risks of experimental firmware modifications. While it offers a significant boost for AI enthusiasts, it is not a