An open-source UEFI tool called nv-uefi-unlocker can remove Tensor Core throughput restrictions on supported NVIDIA GeForce RTX 20, RTX 30 and RTX 40 series graphics cards running Windows. Developed by GitHub user physicsdolphin, the modification operates before the NVIDIA graphics driver loads, extending an earlier Linux-focused approach to Windows systems without requiring a graphics card VBIOS flash.
The project targets restrictions affecting Tensor Core calculations, particularly FP16, BF16 and TF32 workloads involving FP32 accumulation. These operations are important for AI inference, machine learning and other GPU-accelerated computing tasks. NVIDIA’s GeForce hardware can operate at lower Tensor throughput than comparable professional configurations, and the modification attempts to remove the relevant hardware-level restrictions.
The Windows implementation follows earlier Linux modifications that demonstrated higher Tensor Core performance on several RTX generations. Unlike the original Linux approach, nv-uefi-unlocker performs the modification during the UEFI boot process, allowing the adjusted GPU settings to remain active when Windows starts.
The tool uses NVIDIA firmware components and a specially constructed payload to modify protected GPU registers before handing control to the operating system. Its developer describes a process involving firmware initialization, register modifications, verification and GPU reset before Windows Boot Manager takes over.
Because the changes are applied in memory, users do not need to flash a modified VBIOS or permanently alter the graphics card’s firmware. However, the unlock is temporary and must be applied again after a complete power-off. The project’s standard procedure uses a prepared USB drive to launch the UEFI utility before Windows boots.
Early benchmark results suggest substantial improvements in supported Tensor workloads. One reported RTX 4070 Ti Super test increased FP16 throughput from 57.00 TFLOPS to 150.93 TFLOPS, approximately 2.65 times the original result. Another FP16 test using a different matrix size increased from 55.50 TFLOPS to 127.28 TFLOPS, representing a 2.29-fold improvement.
BF16 results from the same reported configuration also improved significantly, rising from 57.51 TFLOPS to 127.32 TFLOPS in one test. These figures come from community testing rather than independent verification across a broad selection of retail graphics cards, so results should not be assumed to apply universally.
The developer’s published testing provides additional evidence of improvements on other hardware. An RTX 4060 Laptop GPU increased FP16 matrix multiplication throughput from 29.89 TFLOPS to 49.33 TFLOPS, approximately 1.65 times its original performance. RTX 20-series testing has also demonstrated gains, although compatibility and operating requirements differ between GPU architectures.
The project’s compatibility documentation covers Turing-based RTX 20-series, Ampere-based RTX 30-series and Ada Lovelace-based RTX 40-series GPUs. Its October 28 updates expanded device-ID coverage across all three architectures, with successful RTX 30-series testing on configurations including the RTX 3060 Ti, RTX 3070 and RTX 3080.
However, expanded device-ID support does not mean every graphics card and motherboard combination has been independently verified. The project’s GitHub issue tracker includes successful configurations alongside reports of boot problems, display initialization failures and Windows device errors. RTX 20-series cards also require additional GSP firmware-mode preparation. NVIDIA’s newer Blackwell-based RTX 50-series cards are outside the supported hardware range.
Higher Tensor throughput can also increase power consumption. One community report involving the RTX 4070 Ti Super recorded board power rising from approximately 227 W to nearly 292 W during testing. The developer’s documentation warns that additional computational activity can increase power draw, temperatures and cooling requirements, particularly when combined with other GPU modifications.
This differs from RTX GPU power and voltage modifications, which primarily adjust operating limits rather than directly changing Tensor instruction throughput. The UEFI unlocker’s performance improvements come from altering restrictions on the execution of supported compute operations, not simply increasing clock speeds.
There are also potential compatibility implications for Windows gaming systems. The project requires Secure Boot to be disabled or an appropriate self-signed certificate to be enrolled in the system firmware. Disabling Secure Boot may prevent games with anti-cheat systems that require it from launching, making the modification unsuitable for some gaming PCs.
Users should also distinguish AI compute improvements from gaming performance. The reported benchmark gains concern specific Tensor workloads and do not demonstrate higher frame rates, improved rasterization performance or faster ray tracing in games. Any benefit to individual AI-assisted gaming features would require separate application-level testing.
Nv-uefi-unlocker remains an experimental community project, with ongoing development addressing hardware compatibility and firmware initialization issues. Its Windows support represents a notable expansion of the earlier Linux-based Tensor unlock, but the available results primarily establish improvements in selected compute benchmarks rather than general-purpose gaming performance.
