Skip to content

Latest commit

Β 

History

122 Commits

Folders and files

Repository files navigation

Universal GPU Unlock (One-Click UEFI Tool)

High-Performance Tensor Core Rate-Limiter Bypass for NVIDIA GPUs

GitHub Release Release Version Downloads Stars Safety Support on Ko-fi


Buy Me a Coffee at ko-fi.com

An automated, universal 1-click UEFI pre-boot unlocker for NVIDIA RTX 40 (Ada), RTX 30 (Ampere) & RTX 20 (Turing) series.
Unlocks full enterprise-grade Tensor Core compute across Windows 10/11, PyTorch, CUDA, and WSL2.


⚑ 60-Second Quick Start

Important

BIOS Prerequisite: Ensure CSM (Compatibility Support Module) is set to Disabled (pure UEFI mode) in your motherboard BIOS. Modern GeForce RTX 30 & 40 GPUs require native UEFI to avoid legacy VGA conflicts (Code 43) and to unlock Resizable BAR.

  1. Download: Get the latest release package: Universal_GPU_Unlock_v1.8.24.zip.
  2. Provision: Extract the ZIP, insert a USB flash drive (FAT32), and run Setup.bat (formats & copies all required UEFI files automatically).
  3. Boot & Unlock: Restart your PC, open your motherboard boot menu (F8 / F11 / F12), and boot from the USB drive:
    • [U] Standard Unlock: RAM-only, temporary unlock (reverts cleanly on complete power off).
    • [I] Install to NVRAM: Registers the UEFI DXE driver into motherboard NVRAM and internal ESP for automatic, stickless boot on every cold start.
    • (Tip: The boot menu features a 30-second countdown. Press [Space] at any time to pause and read in peace).
  4. Verify: In Windows, run Check-GPU-Unlock.bat to verify your unlocked Tensor Core throughput (>200 TFLOPS on RTX 4090).
    • (RTX 20 / Turing users: Run Toggle-Turing-GSP.bat once as Administrator to bind NVIDIA drivers to GSP firmware mode).
    • (Optional: Launch Universal-Control-Center.cmd for detailed hardware telemetry, SMI sensor readbacks, and diagnostics).

πŸš€ Supported Hardware

Generation Architecture Supported Graphics Cards Device IDs
RTX 40 Series Ada Lovelace RTX 4090, RTX 4090D, RTX 4080, RTX 4080 SUPER, RTX 4070 Ti SUPER 0x2684, 0x2685, 0x2702, 0x2703, 0x2704, 0x2705
RTX 30 Series Ampere RTX 3090, RTX 3090 Ti, RTX 3080 Ti, RTX 3080 (10/12GB), RTX 3070 Ti, RTX 3070, RTX 3060 Ti, RTX 3060, CMP 90HX 0x2204, 0x2208, 0x2206, 0x220a, 0x2216, 0x2207, 0x2482, 0x2484, 0x2487, 0x2488, 0x2489, 0x2503, 0x2504, 0x2507, 0x2508, 0x2520, 0x2560
RTX 20 Series Turing (TU102, in testing: #15) RTX 2080 Ti, TITAN RTX, RTX 2080 / SUPER, RTX 2070 / SUPER, RTX 2060 / SUPER, CMP 50HX, Quadro RTX 0x1E04, 0x1E07, 0x1E02, 0x1E81, 0x1E82, 0x1E87, 0x1E84, 0x1F02, 0x1F07, 0x1F06, 0x1F47, 0x1F03, 0x1F08, 0x1F12, 0x1E09, 0x1EB0, 0x1EB1, 0x1E30, 0x1E35
  • Any Vendor: Compatible with Founders Edition, ASUS, MSI, Gigabyte, ZOTAC, PNY, Palit, Gainward, etc.
  • Any Platform: Intel (Z690, Z790, Z890, B760, Z390) and AMD (AM4/AM5, X570, B550, B650, X670E, X870E).
  • Multi-Drive: Prefers bootmgfw.efi and follows your firmware's boot order when several Windows volumes exist.
  • Multi-GPU & Headless Compute: Unlocks primary or dedicated compute GPUs. Discrete cards can operate completely headless as dedicated AI/compute accelerators with all monitors connected to integrated graphics (e.g. AMD Ryzen Raphael iGPU or Intel UHD) β€” resolving #14.
  • ASUS & AMI Console Topology Verified: Multi-display console detachment verified on ASUS PRIME Z690-A at 127.61 TFLOPS (100% unrestricted compute) on EVGA GeForce RTX 3080 Ti β€” resolving #16.
  • ASRock AM4 & CSM Resolution Verified: Budget AM4 POST discovery and CSM legacy video conflict resolved on ASRock A520M-HVS at 106.05 TFLOPS (+95.7% boost, unrestricted compute) on Gainward GeForce RTX 3080 Phoenix β€” refs #17.
  • Ampere Verified, Turing in Testing: GA106 (RTX 3060) and GA102 (RTX 3080, RTX 3080 Ti, RTX 3090) are confirmed in the community table below. GA104 is supported but not yet confirmed in an issue. On TU102 (RTX 2080 Ti) the unlock transaction completes, and the throughput gain is still being confirmed in #15.
  • Stickless Internal NVMe-ESP Persistence ([I] Mode): Installs AdaUnlockDxe.efi and the firmware assets to your internal EFI System Partition (\EFI\Drivers\) and registers it in the UEFI DriverOrder, so the USB stick is no longer needed. If the copy to the ESP fails, [I] registers the driver on the USB drive instead, which then has to stay connected. If an unlock fails, the driver disables itself for the next boot, and [I] enables it again.
πŸ” Community-Verified Hardware & Live Benchmarks (Click to expand)

The following GPU models, system motherboards, and compute readbacks have been independently verified on physical hardware by community users:

Verified GPU Architecture & Silicon Motherboard & Platform Real-World Throughput Verification Issue
Palit GeForce RTX 3060 12GB Ampere (GA106 LHR) MSI MPG Z390 GAMING PLUS 46.08 TFLOPS (~45% ComfyUI speedup) #11
Gainward GeForce RTX 3080 Phoenix 10GB Ampere (GA102) ASRock A520M-HVS (AMD AM4) 106.05 TFLOPS (+95.7% boost, Unrestricted Compute) #17
EVGA GeForce RTX 3080 Ti 12GB Ampere (GA102) ASUS PRIME Z690-A 127.61 TFLOPS (100% Unrestricted Compute) #16
EVGA GeForce RTX 3080 Ti 12GB Ampere (GA102) MSI X570-A PRO 119.0 TFLOPS (+84.5% boost, Unrestricted Compute) #12
NVIDIA GeForce RTX 3090 24GB Ampere (GA102) Gigabyte TRX40 AORUS MASTER (AMD Threadripper 3960X) 117.24 TFLOPS (100% Unrestricted Tensor Compute) #13
ZOTAC GeForce RTX 4080 Trinity OC Ada Lovelace (AD103) GIGABYTE X870 AORUS ELITE WIFI7 (AMD 9800X3D) 174.29 TFLOPS (Full Unrestricted GEMM) #9
NVIDIA GeForce RTX 4090 Ada Lovelace (AD102) MSI MPG X670E Carbon WIFI (AMD Ryzen 7 7800X3D iGPU) 255.0 TFLOPS (+54.5% boost, Headless Compute) #14
NVIDIA GeForce RTX 4090 Ada Lovelace (AD102) Intel Z790 Platform (Multi-Drive ESP) 255.4 TFLOPS (Peak Compute) #8
NVIDIA GeForce RTX 4090 Ada Lovelace (AD102) Intel Z690 Platform (Hybrid iGPU & Multi-Display) 214.0+ TFLOPS (Sustained FP16 GEMM) #5
ASUS WS C621E SAGE Workstation Multi-Socket Server ASPEED AST2500 BMC / IPMI Console Multi-Console Isolation Verified #2

[!TIP] All benchmarks above were measured directly using Check-GPU-Unlock.bat under Windows. If your GPU and motherboard combination is verified and working, feel free to share your benchmark in Discussions or in an issue!


πŸ“Š Real-World Performance Gains

The unlock removes artificial firmware restrictions on the Tensor Cores (GEMM matrix engines). FP32 shader and memory throughput stay at stock. Measured on an RTX 4090:

Workload / Model Factory Stock Unlocked UEFI Real-World Improvement Operational Impact
FLUX.1-dev AI Sampling 1.20 it/s (25.0s) 1.97 it/s (15.2s) πŸš€ +64.2% Faster High-res image generation in ~24s instead of ~34s
MiniMax-H3 Video DiT 15.20 s/it (304s) 11.15 s/it (223s) πŸš€ +36.3% Faster Saves 81 seconds per 5-second video render
FP16 Tensor GEMM ($4096^2$) 132.1 TFLOPS 208.9 TFLOPS πŸš€ +58.1% Boost Maximum compute throughput within 72MB L2 cache
BF16 Tensor GEMM ($8192^2$) 149.3 TFLOPS 167.4 TFLOPS πŸš€ +12.1% Boost Sustained large-matrix training acceleration
FP32 CUDA Shaders 51.3 TFLOPS 48.7 TFLOPS βœ… Stock Parity Standard raster rendering remains at factory stock
VRAM Copy Bandwidth 733.4 GB/s 707.2 GB/s βœ… Stock Parity Physical GDDR6X bus operation unchanged

πŸ›‘οΈ Safety & Anti-Cheat

  • Volatile (RAM-only): The unlock lives in GPU registers and is gone after a full power-off. Without the [I] driver installed, the next boot is stock. On the test bench, the card always booted stock again after a power-off.
  • Official NVIDIA Drivers: Runs seamlessly with standard, unmodified NVIDIA Game Ready and Studio drivers (nvlddmkm.sys).
  • Anti-Cheat: No driver patches, kernel hooks or runtime memory injection. With the image or certificate enrolled in db, UEFISecureBootEnabled stays 1. Individual anti-cheat engines (Riot Vanguard, Easy Anti-Cheat, BattlEye, Ricochet) were not tested systematically.
  • S3 Sleep Protection: Includes Toggle-Sleep-Button.bat in the root directory to hide the "Sleep" button from the Windows Start menu, preventing accidental loss of volatile RAM state.
  • 44-Cycle Testbench: 44 supervised runs on my primary RTX 4090 platform. The last 18 (RUN0027 – RUN0044) all passed. The earlier runs were prototyping.
  • Operational Deep-Dive: Read the comprehensive telemetry and persistence whitepaper:
    πŸ‘‰ Operational Architecture, Persistence & Secure Boot Telemetry

πŸ” Click to view Deep Technical Architecture, GEMM Scaling & Secure Boot Details

Why is the boost largest at $4096 \times 4096$ (+58.1%)?

The Ada Lovelace AD102 architecture features a massive 72 MB L2 cache. A $4096 \times 4096$ FP16 matrix (~32 MB per operand) fits entirely inside this ultra-fast cache (>3.5 TB/s cache bandwidth), so execution is purely Compute-Boundβ€”allowing the unlocked Tensor Cores to reach their full unrestricted 208+ TFLOPS potential. At $16384 \times 16384$ (>1.5 GB memory footprint), workloads become Memory-Bandwidth Bound by external GDDR6X VRAM. Because real-world AI inference & training models slice tensors into $1024$ to $4096$ blocks, they operate right in this maximum speedup sweet spot!

Secure Boot Key Enrollment

The included EFI binaries are pre-signed with my custom Authenticode key. To boot with Secure Boot enabled:

  1. Enter your BIOS / UEFI setup menu.
  2. Navigate to Security / Secure Boot / Key Management.
  3. Choose Append / Enroll DB Key, browse to your USB drive, and select AdaUnlock_SecureBoot.cer.
  4. Save and reboot. Alternatively, you can temporarily set Secure Boot to Disabled or Custom.

πŸ§ͺ Build & Tests

Requires the Visual Studio 2022 C++ Build Tools (x64).

Command What it does
src\Build-Experimental-MSVC.ps1 Builds BOOTX64.EFI and overwrites bin\BOOTX64.EFI. Signs it only if the maintainer certificate is installed.
src\Build-Dxe-Driver.ps1 Builds AdaUnlockDxe.efi and overwrites bin\AdaUnlockDxe.efi, with the same signing rule.
src\Build-Gpu-Topology-Diagnostic.ps1 Builds the read-only topology diagnostic EFI.
tests\Run-MsvcHostTests.ps1 Runs the MSVC host tests for driver_nvram.c ([I]/[R]) and the Windows volume selection in boot_io.c.

The gcc/clang host tests for topology, inventory and discovery are in tests/. The CI workflow (.github/workflows/gpu-topology-diagnostic.yml) runs on every push and pull request that touches src/ or tests/. It runs all host tests, builds the diagnostic EFI, BOOTX64.EFI and AdaUnlockDxe.efi, and uploads the unsigned builds as workflow artifacts. All tests mock the firmware services; none of them runs on a GPU.


πŸ“œ Changelog & Version History

All version releases, bug fixes, and device expansions are maintained in the separate changelog file:
πŸ‘‰ View Full CHANGELOG.md (or browse GitHub Releases).


πŸ› οΈ Repository Structure

β”œβ”€β”€ bin/
β”‚   β”œβ”€β”€ BOOTX64.EFI              <- Pre-compiled, Authenticode-signed universal UEFI bootloader
β”‚   └── AdaUnlockDxe.efi         <- Pre-compiled, Authenticode-signed early DXE driver
β”œβ”€β”€ AD102-ASSETS/                <- Ada microcode payloads (4090, 4080, 4070TiS)
β”œβ”€β”€ GA102-ASSETS/                <- Ampere microcode payloads (3090, 3080, 3070, 3060)
β”œβ”€β”€ TU102-ASSETS/                <- Turing microcode staging payloads (2080Ti, 2080, 2070, 2060)
β”œβ”€β”€ src/                         <- Full C source code for the universal UEFI application
β”œβ”€β”€ tests/                       <- Host tests with mocked UEFI services (no GPU required)
β”œβ”€β”€ Setup.bat                    <- The 1-Click Provisioner Launcher (Primary tool)
β”œβ”€β”€ Setup-AdaUnlock.ps1          <- Automated PowerShell deployment engine
β”œβ”€β”€ Check-GPU-Unlock.bat         <- Standalone Windows verification benchmark (>200 TFLOPS)
β”œβ”€β”€ Universal-Control-Center.cmd <- Optional Diagnostic & Telemetry Dashboard (WPF GUI)
β”œβ”€β”€ Toggle-Sleep-Button.bat      <- Non-invasive tool to hide/restore Start Menu Sleep button (S3 protection)
└── CHANGELOG.md                 <- Complete release and version history

πŸ™ Acknowledgements & Credits

  • 🌟 geodnb / Ablation Lab (Special Dedication): Massive thanks for pioneering research, hardware state analysis, and technical insights into GPU cold states and register architecture that catalyzed this entire universal Windows project. Explore their research over at Ablation Lab on GitHub and subscribe to their channel on YouTube (@AblationLab)!
  • Zagitta & WildFlash1st: For pioneering the Linux gpu-unlock research and cmp90hx-unlock rejoining methodologies.
  • Ahsan Umair (@ahsanumair): For contributing the initial Windows desktop WPF GUI layout and XAML groundwork adapted into the optional diagnostic telemetry companion (Universal-Control-Center.cmd).
  • TOTORONG & JustAResearcher: For early proof-of-concept Windows UEFI implementations.
  • Igor Wallossek (Igor'sLAB): For in-depth technical analysis, power verification, and community discussion.

β˜• Support the Project

If Universal GPU Unlock accelerated your AI workflows, saved you render time, or helped unlock your graphics card's full silicon potential, consider buying me a coffee to support continuous updates and hardware testing:

Buy Me a Coffee at ko-fi.com

Releases

Packages

Contributors

Languages