The Windows route: an EFI payload that runs before Windows boots
The three projects documented on this site are Linux kernel modules. Windows-GPU-UNLOCK is a different implementation: it installs its own EFI System Partition and performs the register work in a payload that hands off to Windows Boot Manager. It also publishes no stock-vs-unlocked pair, so this page reports the recorded profiles and the boot routes rather than a gain number.
This is not the Linux module and does not replace it
Nothing here applies to a Linux install, and the kernel module does not apply to Windows. The two share the underlying idea — write the rate-select register before the NVIDIA driver binds — but the boot path, the tooling and the failure modes are different. If you are on Linux, use the card list and the benchmarks on this site.
What the release is, at the time of writing
The latest tag is v0.3.1, a source-only pre-release. Read the scope list before assuming a card is covered.
- Later main-branch hardening added host and build checks, not a new physical boot or recovery test.
- The latest tag, v0.3.1, is a source-only pre-release; the tree carries changes beyond it, including the USB route and the X870 profile.
- Firmware assets and partition-specific EFI binaries are not bundled — you build locally with your own partition GUIDs.
- A generated local profile does not add another verified machine profile.
Recorded and reported hardware
A profile covers one card in one specific system. The same GPU model on a different motherboard is not covered by that profile, and only the first row was recorded by the maintainer under their own test.
| Card | Chip | Status | Route |
|---|---|---|---|
| RTX 3090 FE 24 GB | GA102 | Recorded by the maintainer | Internal headless — default profile (ASUS TUF B650M) |
| RTX 3070 Ti 8 GB | GA104 | Hardware success reported | Internal headless (ASUS TUF B660M-E D4) |
| RTX 3090 Ti 24 GB | GA102 | Hardware success reported | Headless USB only (Gigabyte X870 Aorus Elite WiFi7) |
| RTX 3060 12 GB | GA104 | Hardware success reported | Headless USB only (ASUS Prime B660-Plus D4) |
| RTX 3080 | GA102 | Upstream target only | No dedicated profile in this port |
Four boot routes, and they are not interchangeable
Which route you take depends on whether the display stays on the discrete card and whether you can use a USB stick. Each has its own guide in the repository.
| Route | Display and input | Guide |
|---|---|---|
| Internal headless | Monitors on the RTX card, display temporarily removed, no T/U/Enter input | Open guide |
| Headless USB | Profile-specific wrapper and child staged on a FAT32 USB stick | Open guide |
| Internal AMD-display manual / auto | Motherboard or iGPU display preserved; the manual variant needs U | Open guide |
| Legacy USB manual | AMD/iGPU display profile; U once, then Boot Override to Windows without a reset | Open guide |
Requirements before you start
- Exact supported hardware from a recorded profile — a matching GPU model on another board is not covered.
- A separate GPT FAT32 EFI System Partition of at least 1 GiB. An ordinary FAT32 data partition does not qualify.
- The existing Windows ESP and recovery partition preserved; the tools never create, shrink or format partitions.
- Secure Boot disabled, a fully decrypted Windows C: with BitLocker protection off, and a connected keyboard.
- Python 3.9+, clang/lld-link 20.1.8, the pinned repository dependencies and five firmware assets obtained locally.
The order that matters
1. Prepare a second ESP, not a data partition
The port boots from its own EFI System Partition and leaves Windows' ESP and recovery partition untouched. It does not create, shrink, delete or format anything: you prepare a GPT FAT32 partition of at least 1 GiB and give it a letter.
2. Configure and build from an elevated PowerShell
Both scripts run from the repository root in an administrator PowerShell. Configuration and build stage no boot entry by themselves.
.\internal-boot\Configure.ps1 -UnlockDriveLetter U `
-WindowsEspGuid 'YOUR-WINDOWS-ESP-GUID'
.research\headless\build-unlock.ps1 `
-Clang 'C:\TOOLS\LLVM20\bin\clang.exe' `
-LldLink 'C:\TOOLS\LLVM20\bin\lld-link.exe'3. Register a separate F8 test entry first
Registering adds its own test entry and keeps your existing boot priority. Automatic boot is a separate, explicit action and stays off until the F8 test has actually worked.
.\internal-boot\Headless.ps1 -Assets '.\assets' -Register4. Run one F8 test and wait
Save your work and fully power off, then pick the test entry. Do not press T, U or Enter. The screen goes black while the RTX GOP is removed; the child attempts once and the wrapper starts Windows directly.
5. Only then enable automatic boot
After the F8 test succeeds and the display recovers, automatic boot can be enabled. That action places headless first, Windows Boot Manager second, and disables Windows Fast Startup. Registering the test entry alone does nothing like this.
.\internal-boot\Headless.ps1 -EnableAutomatic -ConfirmedSuccess6. Recovery is Windows-first and independent of the build
If the handoff fails or Windows does not return, power off fully and choose F8 → Windows Boot Manager, or run the restore switch. Windows-first recovery works without saved state.
.\internal-boot\Headless.ps1 -RestoreWindowsFirstWhat the recorded run measured
The original B650M / RTX 3090 FE run reported these CUDA 12.6 / cuBLAS 12.6 numbers at 4096 x 4096, as a median of three runs. These are absolute figures from an unlocked card — there is no stock side published for them, so they are not a gain.
| Path | Value |
|---|---|
FP32 pedantic | 22.85 TFLOPS |
FP16 Tensor / FP32 accumulate | 124.97 TFLOPS |
BF16 Tensor / FP32 accumulate | 128.93 TFLOPS |
What the recorded run left behind
The successful child reported this state after the unlock, which is the evidence that the card was handed to the driver cleanly.
- SS0 read back as 0x88888888
- SS1 read back as 0x8
- Bus-master enable disabled
- Zero remaining mappings
- Cleanup completed and one FLR performed
Licence
GPL-2.0-only for the port and the reused core; EDK2 and separately attributed components keep their own licences. Based on rough2515-hub/rtx3080-unlock-public and Zagitta/gpu-unlock. NVIDIA firmware, PROM/BIOS images, drivers and CUDA DLLs are not redistributed in the source release.