Skip to content

Latest commit

 

History

History
158 lines (121 loc) · 6.71 KB

QUICKSTART.md

File metadata and controls

158 lines (121 loc) · 6.71 KB

GPU Unlock: quick start

This guide is for a Linux machine you control. The RTX 4070 Ti SUPER path is experimental and has been tested on one 10de:2705 card. RTX 4090 10de:2684 has a separate single-card report. RTX 4060 Ti 16GB 10de:2805 is experimental and has been unlocked on one card (2805 report). Other Ada IDs are not enabled. The measured instruction-rate gain is for FP16, BF16 and TF32 tensor work with FP32 accumulation. INT8 and unsigned INT4 did not speed up in the 2705 tests. One matched FP8/FP32 GEMM was effectively unchanged (59.326 → 59.495 TF/s).

1. Find your GPU's PCI address

lspci -Dnn | grep -i nvidia

Find the line for your GPU. The full address is the first field, such as 0000:01:00.0. Confirm the device ID is 10de:2705 for the experimental RTX 4070 Ti SUPER path, 10de:2805 for the experimental RTX 4060 Ti 16GB path or 10de:2684 for the RTX 4090 path. Keep this address for step 3. On a multi-GPU host, identify the exact intended card; do not use targets=all for an Ada test.

2. Install from open source

Install Git, your running kernel's headers, a C build toolchain, DKMS, Python 3, and zstd using your distribution's package manager. Then:

git clone https://github.com/JustAResearcher/gpu-unlock-ada.git
cd gpu-unlock-ada
make install

make install selects the Arch, Debian/Ubuntu, or Fedora/RHEL package path. The build downloads SHA-pinned NVIDIA 610.43.03 inputs and generates the firmware locally. The downloadable source contains no NVIDIA firmware binary. The build may take several minutes and asks for sudo when installing.

If your distribution cannot resolve a build dependency automatically, use make install-arch, make install-deb, or make install-rpm as appropriate; the README explains the generated package paths.

3. Enable one exact card

From the cloned folder, run this read-only helper to print the exact option for each eligible GPU it finds:

python3 tools/print_boot_option.py

Choose one printed line for the first cold test. The examples below show the format if you prefer to enter it manually.

For an RTX 4070 Ti SUPER 10de:2705, append these two options to your kernel command line, replacing the sample address with yours:

gpu_unlock.targets=0000:01:00.0 gpu_unlock.ad103_experimental=1

For an RTX 4060 Ti 16GB 10de:2805, use its own experimental flag instead:

gpu_unlock.targets=0000:01:00.0 gpu_unlock.ad106_experimental=1

For an RTX 4090 10de:2684, only the first option is needed:

gpu_unlock.targets=0000:01:00.0

Use the bootloader method your machine already has:

  • Ubuntu/Debian with GRUB: Edit GRUB_CMDLINE_LINUX_DEFAULT in /etc/default/grub, append the options inside its existing quotes, save, then run sudo update-grub.
  • Arch with GRUB: Make the same edit, then run sudo grub-mkconfig -o /boot/grub/grub.cfg.
  • Fedora/RHEL with GRUB: Run sudo grubby --update-kernel=ALL --args="gpu_unlock.targets=0000:01:00.0 gpu_unlock.ad103_experimental=1", substituting your PCI address and omitting the experimental flag for 4090.
  • systemd-boot: Run bootctl status to find the active entry. Add the options to that entry's options line. Make sure its initrd points to the newly generated initramfs containing gpu_unlock; some distributions copy /boot/initrd.img-* to the EFI partition separately. Keep a copy of the original entry so you can select it if needed.

Keep a normal boot entry available. Do not copy a sample PCI address. The package does nothing until these options are present.

4. Start from full power-off

Shut down cleanly. Cut AC power to the machine, wait about 30 seconds, then restore power and start it. This gives the first test a fresh card state. If the GPU has a customer or rented workload, wait until it is fully released before doing this.

5. Check the result

After the stock NVIDIA driver loads, run:

journalctl -k -b --no-pager | grep -E 'gpu_unlock: unlocked|gpu_unlock: unlock failed|post-FLR SEC2'
nvidia-smi

Success requires the exact selected card's gpu_unlock: unlocked line, healthy nvidia-smi, and a workload check. A loaded module alone is not proof. The 2705 report shows the expected marker, permission/selector readbacks, matched performance and numerical checks from the tested card. For a single-GPU host, the read-only rate getter is:

python3 tools/read_sm_rates.py --expect-uuid GPU-PASTE-YOUR-EXACT-UUID

Get the UUID from nvidia-smi --query-gpu=uuid --format=csv,noheader after boot. Do not use a UUID from another card.

For a stock/unlocked comparison, collect the stock measurements before enabling the test and use the same running kernel, driver, binaries and power cap in both states. Check uname -r and /proc/cmdline; a boot-menu title alone does not identify the kernel or prove that the test entry ran. The testing guide includes benchmark commands and the matched RTX 4090 results.

Keep it enabled across boots

First test an ordinary reboot with the same early initramfs and exact target options selected for one boot. Check for gpu_unlock: unlocked, healthy nvidia-smi, and full rate readbacks again. On the tested 4070 Ti SUPER, a warm reboot returned the card to a fresh enough state and the unlock succeeded; other boards may retain firmware state and be rejected. A loaded module alone does not prove the unlock succeeded.

After that test succeeds, leave the target options in your persistent GRUB entry. With systemd-boot, make the tested Ada entry persistent using sudo bootctl set-default YOUR-ADA-ENTRY.conf, replacing the entry ID with the one shown by bootctl status. Keep a separate stock entry in the boot menu. bootctl set-oneshot selects only the next boot and will not persist. After a kernel update, verify that the selected entry points to a rebuilt initramfs containing gpu_unlock before booting the new kernel.

Roll back

Remove the gpu_unlock.* kernel-command-line options and boot the normal entry. If the firmware transaction did not stop cleanly, power off completely before retrying; do not force-unload the module while it may retain DMA buffers. Uninstall the generated packages if you no longer want the module or firmware. If you used bootctl set-default for a separate Ada entry, run sudo bootctl set-default YOUR-STOCK-ENTRY.conf and confirm the selected default with bootctl status before rebooting.

For licensing, provenance and distribution limits, see firmware/NOTICE.md and the release notes.