Over the past couple of weeks, upstream researchers showed that consumer GeForce RTX 40, 30, and 20 series cards have their FP32-accumulate Tensor Core instructions (FP16 -> FP32, BF16 -> FP32, TF32 -> FP32, and FP8 E4M3 -> FP32) capped at half speed (1/2) in firmware, and that this cap can be lifted in volatile registers before the OS driver loads.
When I tried setting this up on my Windows 11 PC with a Palit RTX 4070 Ti (12 GB, 300 W), the setup process was rough. Existing repos were either Linux-only or tied to a single specific board profile, so running it on Windows meant manually pulling your card IDs, VRAM offsets, and EFI partition GUID, editing C headers by hand, and compiling BOOTX64.EFI yourself. Worse, if a card had a non-standard VBIOS layout or BitLocker enabled, early UEFI loaders could leave you on a black screen or shut down on boot.
To fix this, I combined and refined the open-source research (TOTORONG, Zagitta, elias-gaeros, JustAResearcher, KohakuBlueleaf, impacifist, and Jon Pry) into one free cross-platform tool (GPL-2.0-only) with a native GUI localized in English ([EN]), Russian ([RU]), Chinese ([中文]), and Hindi ([हिन्दी]):
https://github.com/MindAllik/GPU-Tensor-Unlock-Universal
What it does on Windows 10/11 (Windows_Launch_UI.bat)
- Automatically detects your card's PCI IDs, VRAM window, and Windows boot partition and builds the bootable FAT32 USB stick without editing any C code.
- Checks BitLocker and Secure Boot upfront so your PC doesn't shut down on boot.
- Before disabling the display driver in UEFI, it dry-runs the VBIOS patch in RAM (
PromCheckviaBAR0 + 0x300000). If your card's VBIOS structure is non-standard, it aborts safely and boots straight into Windows without a black screen. - Includes a 7-kernel PTX JIT benchmark (with runtime-randomized non-zero matrices) and a one-click button to raise the Power Limit to your card's VBIOS maximum via NVML.
What it does on Linux (Linux_Launch_UI.sh or make install)
- Covers RTX 40 (
Ada), RTX 30 (Ampere), and RTX 20 (Turing) series in a single kernel module (gpu_unlock.ko), installs via DKMS, and updatesinitramfsautomatically. - Can unlock secondary headless GPUs live without a reboot.
- Includes the same GUI benchmark and direct driver status check via
/dev/nvidiactl.
Freshly Measured Numbers on Palit RTX 4070 Ti (10de:2782, Windows 11 UEFI, 300.0 W on both runs)
| Kernel / Instruction | Gate | Stock (300W) | Unlocked (300W) | Gain |
|---|---|---|---|---|
FP16 -> FP16 (Control) |
FMLA16 |
174.96 TF/s | 175.99 TF/s | 1.01x (+0.6%) |
FP16 -> FP32 |
FMLA32 |
87.36 TF/s | 175.43 TF/s | 2.01x (+100.8%) |
BF16 -> FP32 |
FMLA32 |
87.20 TF/s | 174.87 TF/s | 2.01x (+100.5%) |
TF32 -> FP32 |
FMLA32 |
43.68 TF/s | 87.77 TF/s | 2.01x (+100.9%) |
FP8 E4M3 -> FP32 |
FMLA32 |
174.29 TF/s | 350.66 TF/s | 2.01x (+101.2%) |
INT8 -> INT32 (Control) |
IMLA1 |
348.95 TOP/s | 350.43 TOP/s | 1.00x (+0.4%) |
Scalar FP32 (Control) |
FFMA |
42.85 TF/s | 42.84 TF/s | 1.00x (-0.0%) |
The VBIOS flash chip is never overwritten; the patch lives in volatile GPU registers and resets completely after a full power-off.
If you test it on other desktop RTX 40, 30, or 20 series cards, please share your card model and EFI\UNLOCK\REPORT.TXT (or dmesg on Linux) in GitHub Issues even if the pre-check auto-aborted (:
(screenshot to show the GUI, without unlocking)
Top comments (0)