NVIDIA GPU Unlock

weingeist · 24 October 2026 21:18

nvidia 40xx

nvidia 30xx

Hat das schon jemand ausprobiert ?
Geht das nur wenn man nvidia-dkms installiert hat ?
Bei mir findet er am Ende das Modul nicht.

Wollie · 25 October 2026 06:04

Was genau hast Du denn ausprobiert?

weingeist · 25 October 2026 12:13

diese Anleitung hier:

github.com/JustAResearcher/gpu-unlock-ada

QUICKSTART.md

main

# 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](rtx4090-register-report.md).
RTX 4060 Ti 16GB `10de:2805` is experimental and has been unlocked on **one**
card ([2805 report](docs/ada-2805-report.md)). 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

```sh
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

This file has been truncated. show original

da wird dann dkms installiert, aber am Ende findet er dkms header nicht. Ich denke dadurch klappt das nicht.

Wollie · 25 October 2026 13:28

Sorry, zum Helfen ist mir das zu unkonkret…

weingeist · 25 October 2026 15:35

ich habe das jetzt verworfen. Der lädt ja die firmware 610. Wir sind bei 615. Wenn ich das richtig sehe müsste man ja bei jedem Update patchen. Da ich auf unstable bin würde das ausufern.

fumum_vendidi · 25 October 2026 15:47

Wozu willst du das denn überhaupt anwenden? Lokale KI?
Für Gaming bringt das nichts.

weingeist · 31 October 2026 12:46

Minimax H3 Video Generation in ComfyUI. Da sollte das doch passen.

github.com/JustAResearcher/gpu-unlock-ada

Compatibility report: RTX 4080 SUPER (AD103, 10de:2702) on Linux (Maxsun, driver 615.78.08) — matched stock/unlocked + end-to-end ComfyUI A/B

opened 07:15PM - 26 Oct 26 UTC

ccchenyulin

### Result **Unlocked, and FP16/BF16 got faster — matched stock/unlocked on the… same card, plus an end-to-end application measurement.** This is the first **Linux kernel-module** report for `10de:2702` (existing 2702 reports are Windows/UEFI: #53 modified 32 GB, #70 ASUS reference). Board is a different AIB (Maxsun), driver is `615.78.08`, kernel `7.2.11-arch1-1`. ### Card NVIDIA GeForce RTX 4080 SUPER, AD103 rev a1, 80 SM, 16 GB, 340 W cap (max 360 W). It is the only GPU and it drives the display. ### PCI IDs ```text 0000:01:00.0 VGA compatible controller [0300]: NVIDIA Corporation AD103 [GeForce RTX 4080 SUPER] [10de:2702] (rev a1) Subsystem: GUANGZHOU MAXSUN INFORMATION TECHNOLOGY CO., LTD. Device [6688:2702] ``` ### VBIOS version `95.03.44.40.68` ### Is the card modified? Stock. No VBIOS flash, no memory mod, no overclock; the 340 W board limit was left alone. ### Distribution Arch Linux ### Kernel `7.2.11-arch1-1` (mkinitcpio 42.2, `systemd` hook) ### NVIDIA driver `615.78.08`, open kernel modules. Forced upgrade not needed: the unlock ran before this driver bound and the driver initialized normally afterwards. ### Initramfs generator mkinitcpio. The package's drop-in puts `gpu_unlock` into `MODULES`; `nvidia.ko` is not in the initramfs on this host, so `gpu_unlock` always runs first (`lsinitcpio | grep -c nvidia.ko` → 0). ### gpu-unlock version and install method `JustAResearcher/gpu-unlock-ada` at `d28c485` (the commit that admitted `0x2702`), built locally with `makepkg` → `gpu-unlock-dkms 1.1.1-1` + `gpu-unlock-firmware 1.1.1-1`, DKMS-built against the running kernel. All nine firmware assets generated locally from the pinned 610.43.03 inputs; every SHA-256 matched the manifest and the published values (AD102 payload `44c36b2d…`, AD102 booter `6ee05b20…`, GSP `572373620…`). ### Kernel command line ```text gpu_unlock.targets=0000:01:00.0 gpu_unlock.ad103_experimental=1 ``` Persisted in GRUB for normal boots. For A/B I added a second GRUB entry with the same kernel/initramfs but `module_blacklist=gpu_unlock` instead of the two options, so no reboots require hand-editing. ### Setup Single GPU, display card, Denial/Wayland desktop. Two cold boots (full power-off) after enabling, both green. Around one hour of unlocked runtime total across the two boots, including 20 s of sustained GEMM and 12 workflow runs. **No 8 h/24 h soak.** ### Kernel log (trimmed, host and paths redacted) ```text <host> kernel: pci 0000:01:00.0: entry: plm=0xffffff8f ss0=0x6246065 ss1=0x4 wpr2_hi=0x0 <host> kernel: pci 0000:01:00.0: AD102 permission before-firmware: FEAT[0x823804]=0xffffff8f before=0xffffff8f expected=0xffffffff required=1 valid=1 pci=0x270210de chip=0x193000a1 <host> kernel: pci 0000:01:00.0: FWSEC WPR2: low=0x3ff6000 high=0x3ff6e00 expected_page=0x3ff600 end=0x3ff700000 <host> kernel: pci 0000:01:00.0: AD102 payload result: boot=0 halt=0 marker=0x67b00001 expected=0x67b00001 marker_ok=1 required_permissions=1 cpuctl=0x10 halted=1 debug=0xda550000 ss0=0x6246065 ss1=0x4 <host> kernel: pci 0000:01:00.0: AD102 permission after-payload: WPR_CFG[0x1fa7cc]=0xfffff0ff required=1 valid=1 <host> kernel: pci 0000:01:00.0: AD102 permission after-payload: WPR[0x1fa7c4]=0xffffffff required=1 valid=1 <host> kernel: pci 0000:01:00.0: AD102 permission after-payload: FEAT[0x823804]=0xffffffff required=1 valid=1 <host> kernel: pci 0000:01:00.0: AD102 permission after-payload: FBPA[0x9a0148]=0xffffff0f required=0 valid=1 <host> kernel: pci 0000:01:00.0: selector result: SS0=0x88888888 expected=0x88888888 matched=1 <host> kernel: pci 0000:01:00.0: selector result: SS1=0x8 expected=0x8 matched=1 <host> kernel: pci 0000:01:00.0: AD102 selectors-applied: readout1=0x0 fields=0 0 0 0 0 0 0 0 0 valid=1 full=1 <host> kernel: pci 0000:01:00.0: PCIe FLR begin: BME=0 DMA retained, chip=0x193000a1 <host> kernel: pci 0000:01:00.0: PCIe FLR returned: result=0 <host> kernel: pci 0000:01:00.0: post-FLR SEC2 reset released, GSP halted, BME=0 pending=0 WPR2=1ffffe00:00000000 SS0=88888888 SS1=00000008 <host> kernel: pci 0000:01:00.0: AD102 post-FLR: readout1=0x0 fields=0 0 0 0 0 0 0 0 0 valid=1 full=1 <host> kernel: pci 0000:01:00.0: gpu_unlock: unlocked ``` Module load → `unlocked` was about 0.9 s. No Xid, NVRM, SEC2 or GSP timeout in either boot. **Second cold boot entered with different cold selectors** — `ss0=0x0 ss1=0x0` instead of `0x6246065`/`0x4` — and unlocked identically (`SS0=0x88888888`, `SS1=0x8`, `readout1=0x0 full=1`). That matches the "cold selector values are fuse-strapping noise, not a precondition" note in #24. ### RM getter (read-only) | Field | Stock | Unlocked | | --- | ---: | ---: | | FMLA32 | 1 (half) | **0 (full)** | | IMLA2 | 1 (half) | **0 (full)** | | IMLA4 | 1 (half) | **0 (full)** | | IMLA0, FMLA16, DP, FFMA, IMLA1, IMLA3 | 0 | 0 | Note: on this card with driver `615.78.08` the flipped state reads **`0`**, not `3`. The "3 on 615.x" note in #22/#34 does not reproduce here, so the module's own post-FLR `full=1` plus a matched benchmark remain the rate evidence. ### Instruction throughput (same cubin, 340 W) `diagnostics/bench/tensor_modes_issue.cu`, cubin SHA-256 `33aa6db385aa9eb8a1f73bb2f74891aea84794326b7c40992bc642f93a3c614d` (identical to the cubin used in #19): | Mode | Stock | Unlocked | Ratio | Mismatches | | --- | ---: | ---: | ---: | ---: | | FP16 / FP32 accumulate | 105.76 TF/s | **203.07** | **1.92×** | 0 / 270,336 | | TF32 / FP32 accumulate | 53.06 | **109.61** | **2.07×** | 0 / 270,336 | | unsigned INT4 / INT32 | 844.3 TOP/s | 869.1 | 1.03× | 0 / 270,336 | ### cuBLAS GEMM (torch 2.9.0+cu130, median of 10, 3 s warmup before each point) | Shape | Mode | Stock | Unlocked | Ratio | | --- | --- | ---: | ---: | ---: | | 4096³ | FP16 / FP32 acc | 96.3 | **181.0** | **1.88×** | | 4096³ | BF16 / FP32 acc | 96.2 | **189.6** | **1.97×** | | 8192³ | FP16 / FP32 acc | 111.0 | **175.9** | **1.58×** | | 8192³ | BF16 / FP32 acc | 111.1 | **184.7** | **1.66×** | | 4096³ / 8192³ | FP32 (control) | 37.3 / 37.0 | 35.8 / 36.1 | 0.96× (flat) | The flat FP32 control is what makes the rest credible — it shows the gain is on the tensor path, not clocks. ### Sustained load (20 s, 8192³ FP16) | | Stock | Unlocked | | --- | ---: | ---: | | Throughput | 105.3 TF/s | **170.9 TF/s (1.62×)** | | Board power (median / peak) | 316 W / 319 W | **339 W / 341 W (at the 340 W cap)** | | SM clock (median) | 2640 MHz | **2415 MHz** | | Temperature (peak) | 63 °C | 66 °C | That is the power wall, and it explains why 8192³ gains less than 4096³: the card now sits at its own cap. ### End-to-end application measurement (the part I have not seen reported anywhere) ComfyUI (mkinitcpio-installed module, `--use-sage-attention --cache-none --fast-disk`), one workflow: SDXL/Illustrious UNET (`waiNSFWIllustrious_v160`), VAE `sdxl-vae-fp16-fix`, **1024×1360**, **27 steps**, `euler_ancestral`, CFG 4.5, scheduler `normal`. Prompt source is a fixed 27-tag set, so every run does identical work. Timing is the server-side `execution_start` → `execution_success` interval from `/history`, i.e. it excludes browser/queue overhead. One warmup run (model load) before each series, then six measured runs: | State | Six runs (s) | Median | | --- | --- | ---: | | stock (`module_blacklist=gpu_unlock`) | 7.79 / 7.85 / 7.91 / 7.98 / 8.44 / 8.45 | **7.94** | | unlocked, cold boot 1 | 5.91 / 5.98 / 6.04 / 6.08 / 6.12 / 6.70 | **6.06** | | unlocked, cold boot 2 (after persisting) | 6.00 / 6.21 / 6.26 / 6.30 / 6.42 / 6.55 | **6.28** | **1.31× / 1.26× end-to-end, 21–24 % less wall time, 1.7–1.9 s per image.** Per step: 294 ms → 224 ms. Solving `C + M = 7.94` and `0.5C + M = 6.06` gives C ≈ 3.8 s (≈47 % of the runtime on the unlocked path), M ≈ 4.2 s (VAE decode, weight movement, non-tensor work). SageAttention was active in all three series. ### Anything else - **Method note for anyone benchmarking this**: short GEMM probes without warmup are misleading. My first unlocked cuBLAS run read 46.6 TF/s (FP16 4096³) — *slower than stock* — with the FP32 control also down to 15.9; the GPU was still at 210–345 MHz idle clocks. Adding a 3 s full load before each measurement and keeping an FP32 control fixed it. The control moving in lockstep with the target is the tell. - **Not verified**: other 2702 boards (this is one Maxsun card), other driver branches, another kernel, long-duration load, other workflows (FLUX / video), and suspend/resume (fork has no resume re-unlock; this machine has not suspended in 90 days). - Happy to provide sanitized transaction logs, benchmark JSON and the exact scripts on request.

Wenn ich das richtig sehe ist die Treiberversion egal.
Er installiert es aber ein wenig anders als im Quickstart.

makepkg → gpu-unlock-dkms 1.1.1-1 + gpu-unlock-firmware 1.1.1-1

Ich habe mit make jetzt keine Ahnung.
Weiß wer wie er das meint ?

Es gibt jetzt im AUR 2 Pakete die das erledigen.

Diese habe ich installiert und die grub Parameter hinzugefügt.
Aber irgendwie funktioniert es nicht.

[2026-10-31T13:10:51+0100] [ALPM] installed gpu-unlock-firmware (2.0.0-1)
[2026-10-31T13:10:51+0100] [ALPM] installed gpu-unlock-dkms (2.0.0-1)
[2026-10-31T13:10:51+0100] [ALPM-SCRIPTLET] ==> gpu_unlock now loads ahead of nvidia in the initramfs but does nothing until
[2026-10-31T13:10:51+0100] [ALPM-SCRIPTLET] ==> you opt in on the kernel command line:
[2026-10-31T13:10:51+0100] [ALPM-SCRIPTLET] ==>   gpu_unlock.targets=all              every supported unbound card
[2026-10-31T13:10:51+0100] [ALPM-SCRIPTLET] ==>   gpu_unlock.targets=0000:01:00.0     only these PCI addresses (comma-separated)
[2026-10-31T13:10:51+0100] [ALPM-SCRIPTLET] ==> Remove the option in your bootloader to boot without the unlock.
[2026-10-31T13:10:51+0100] [ALPM-SCRIPTLET] ==> Laptops: also set `options nvidia NVreg_DynamicPowerManagement=0x00`,
[2026-10-31T13:10:51+0100] [ALPM-SCRIPTLET] ==> or the unlock is lost when the GPU powers down.
[2026-10-31T13:10:51+0100] [ALPM-SCRIPTLET] ==> Rebuild the initramfs after install, upgrade and removal:
[2026-10-31T13:10:51+0100] [ALPM-SCRIPTLET] ==>   mkinitcpio -P   (dracut users: dracut --regenerate-all --force)
[2026-10-31T13:10:51+0100] [ALPM-SCRIPTLET] ==> See /usr/share/doc/gpu-unlock/README.md.
[2026-10-31T13:10:51+0100] [ALPM] transaction completed
[2026-10-31T13:10:51+0100] [ALPM] running '35-systemd-update.hook'...
[2026-10-31T13:10:51+0100] [ALPM] running '70-dkms-install.hook'...
[2026-10-31T13:10:51+0100] [ALPM-SCRIPTLET] ==> ERROR: Missing rootfs-pkgs.txt kernel headers for module gpu-unlock/2.0.0.
[2026-10-31T13:10:51+0100] [ALPM-SCRIPTLET] ==> ERROR: Missing boot kernel headers for module gpu-unlock/2.0.0.
[2026-10-31T13:10:51+0100] [ALPM-SCRIPTLET] ==> ERROR: Missing tmp kernel headers for module gpu-unlock/2.0.0.
[2026-10-31T13:10:51+0100] [ALPM-SCRIPTLET] ==> ERROR: Missing lost+found kernel headers for module gpu-unlock/2.0.0.
[2026-10-31T13:10:51+0100] [ALPM-SCRIPTLET] ==> ERROR: Missing run kernel headers for module gpu-unlock/2.0.0.
[2026-10-31T13:10:51+0100] [ALPM-SCRIPTLET] ==> ERROR: Missing sbin kernel headers for module gpu-unlock/2.0.0.
[2026-10-31T13:10:51+0100] [ALPM-SCRIPTLET] ==> ERROR: Missing root kernel headers for module gpu-unlock/2.0.0.
[2026-10-31T13:10:51+0100] [ALPM-SCRIPTLET] ==> ERROR: Missing mnt kernel headers for module gpu-unlock/2.0.0.
[2026-10-31T13:10:51+0100] [ALPM-SCRIPTLET] ==> ERROR: Missing etc kernel headers for module gpu-unlock/2.0.0.
[2026-10-31T13:10:51+0100] [ALPM-SCRIPTLET] ==> ERROR: Missing desktopfs-pkgs.txt kernel headers for module gpu-unlock/2.0.0.
[2026-10-31T13:10:51+0100] [ALPM-SCRIPTLET] ==> ERROR: Missing bin kernel headers for module gpu-unlock/2.0.0.
[2026-10-31T13:10:51+0100] [ALPM-SCRIPTLET] ==> ERROR: Missing var kernel headers for module gpu-unlock/2.0.0.
[2026-10-31T13:10:51+0100] [ALPM-SCRIPTLET] ==> ERROR: Missing lib kernel headers for module gpu-unlock/2.0.0.
[2026-10-31T13:10:51+0100] [ALPM-SCRIPTLET] ==> ERROR: Missing dev kernel headers for module gpu-unlock/2.0.0.
[2026-10-31T13:10:51+0100] [ALPM-SCRIPTLET] ==> ERROR: Missing home kernel headers for module gpu-unlock/2.0.0.
[2026-10-31T13:10:51+0100] [ALPM-SCRIPTLET] ==> ERROR: Missing usr kernel headers for module gpu-unlock/2.0.0.
[2026-10-31T13:10:51+0100] [ALPM-SCRIPTLET] ==> ERROR: Missing srv kernel headers for module gpu-unlock/2.0.0.
[2026-10-31T13:10:51+0100] [ALPM-SCRIPTLET] ==> ERROR: Missing sys kernel headers for module gpu-unlock/2.0.0.
[2026-10-31T13:10:51+0100] [ALPM-SCRIPTLET] ==> ERROR: Missing lib64 kernel headers for module gpu-unlock/2.0.0.
[2026-10-31T13:10:51+0100] [ALPM-SCRIPTLET] ==> ERROR: Missing opt kernel headers for module gpu-unlock/2.0.0.
[2026-10-31T13:10:51+0100] [ALPM-SCRIPTLET] ==> ERROR: Missing proc kernel headers for module gpu-unlock/2.0.0.
[2026-10-31T13:10:51+0100] [ALPM] running '90-mkinitcpio-install.hook'...
[2026-10-31T13:10:51+0100] [ALPM-SCRIPTLET] ==> Building image from preset: /etc/mkinitcpio.d/linux618.preset: 'default'
[2026-10-31T13:10:51+0100] [ALPM-SCRIPTLET] ==> Using default configuration file: '/etc/mkinitcpio.conf'
[2026-10-31T13:10:51+0100] [ALPM-SCRIPTLET]   -> -k /boot/vmlinuz-6.18-x86_64 -g /boot/initramfs-6.18-x86_64.img
[2026-10-31T13:10:51+0100] [ALPM-SCRIPTLET] ==> Using drop-in configuration file: 'zz-gpu-unlock.conf'
[2026-10-31T13:10:51+0100] [ALPM-SCRIPTLET] ==> Starting build: '6.18.60-1-MANJARO'
[2026-10-31T13:10:51+0100] [ALPM-SCRIPTLET]   -> Running build hook: [systemd]
[2026-10-31T13:10:53+0100] [ALPM-SCRIPTLET]   -> Running build hook: [autodetect]
[2026-10-31T13:10:53+0100] [ALPM-SCRIPTLET]   -> Running build hook: [microcode]
[2026-10-31T13:10:54+0100] [ALPM-SCRIPTLET]   -> Running build hook: [kms]
[2026-10-31T13:10:58+0100] [ALPM-SCRIPTLET]   -> Running build hook: [modconf]
[2026-10-31T13:10:58+0100] [ALPM-SCRIPTLET]   -> Running build hook: [block]
[2026-10-31T13:10:58+0100] [ALPM-SCRIPTLET]   -> Running build hook: [keyboard]
[2026-10-31T13:10:58+0100] [ALPM-SCRIPTLET]   -> Running build hook: [sd-vconsole]
[2026-10-31T13:10:58+0100] [ALPM-SCRIPTLET]   -> Running build hook: [filesystems]
[2026-10-31T13:10:58+0100] [ALPM-SCRIPTLET]   -> Running build hook: [fsck]
[2026-10-31T13:10:58+0100] [ALPM-SCRIPTLET] ==> Generating module dependencies
[2026-10-31T13:10:59+0100] [ALPM-SCRIPTLET] ==> Creating zstd-compressed initcpio image: '/boot/initramfs-6.18-x86_64.img'
[2026-10-31T13:10:59+0100] [ALPM-SCRIPTLET]   -> Early uncompressed CPIO image generation successful
[2026-10-31T13:10:59+0100] [ALPM-SCRIPTLET] ==> Initcpio image generation successful
[2026-10-31T13:10:59+0100] [ALPM-SCRIPTLET] ==> Running post hooks
[2026-10-31T13:10:59+0100] [ALPM-SCRIPTLET]   -> Running post hook: [sbctl]
[2026-10-31T13:10:59+0100] [ALPM-SCRIPTLET] Secureboot key directory doesn't exist, not signing!
[2026-10-31T13:10:59+0100] [ALPM-SCRIPTLET] ==> Post processing done
[2026-10-31T13:10:59+0100] [ALPM-SCRIPTLET] ==> Building image from preset: /etc/mkinitcpio.d/linux73.preset: 'default'
[2026-10-31T13:10:59+0100] [ALPM-SCRIPTLET] ==> Using default configuration file: '/etc/mkinitcpio.conf'
[2026-10-31T13:10:59+0100] [ALPM-SCRIPTLET]   -> -k /boot/vmlinuz-7.3-x86_64 -g /boot/initramfs-7.3-x86_64.img
[2026-10-31T13:10:59+0100] [ALPM-SCRIPTLET] ==> Using drop-in configuration file: 'zz-gpu-unlock.conf'
[2026-10-31T13:10:59+0100] [ALPM-SCRIPTLET] ==> Starting build: '7.3.1-1-MANJARO'
[2026-10-31T13:10:59+0100] [ALPM-SCRIPTLET]   -> Running build hook: [systemd]
[2026-10-31T13:11:00+0100] [ALPM-SCRIPTLET]   -> Running build hook: [autodetect]
[2026-10-31T13:11:00+0100] [ALPM-SCRIPTLET]   -> Running build hook: [microcode]
[2026-10-31T13:11:00+0100] [ALPM-SCRIPTLET]   -> Running build hook: [kms]
[2026-10-31T13:11:04+0100] [ALPM-SCRIPTLET]   -> Running build hook: [modconf]
[2026-10-31T13:11:04+0100] [ALPM-SCRIPTLET]   -> Running build hook: [block]
[2026-10-31T13:11:05+0100] [ALPM-SCRIPTLET]   -> Running build hook: [keyboard]
[2026-10-31T13:11:05+0100] [ALPM-SCRIPTLET]   -> Running build hook: [sd-vconsole]
[2026-10-31T13:11:05+0100] [ALPM-SCRIPTLET]   -> Running build hook: [filesystems]
[2026-10-31T13:11:05+0100] [ALPM-SCRIPTLET]   -> Running build hook: [fsck]
[2026-10-31T13:11:05+0100] [ALPM-SCRIPTLET] ==> Generating module dependencies
[2026-10-31T13:11:05+0100] [ALPM-SCRIPTLET] ==> Creating zstd-compressed initcpio image: '/boot/initramfs-7.3-x86_64.img'
[2026-10-31T13:11:05+0100] [ALPM-SCRIPTLET]   -> Early uncompressed CPIO image generation successful
[2026-10-31T13:11:05+0100] [ALPM-SCRIPTLET] ==> Initcpio image generation successful
[2026-10-31T13:11:05+0100] [ALPM-SCRIPTLET] ==> Running post hooks
[2026-10-31T13:11:05+0100] [ALPM-SCRIPTLET]   -> Running post hook: [sbctl]
[2026-10-31T13:11:05+0100] [ALPM-SCRIPTLET] Secureboot key directory doesn't exist, not signing!
[2026-10-31T13:11:05+0100] [ALPM-SCRIPTLET] ==> Post processing done

Weiß wer was da schief geht ?

PS: Update

Es haben die entsprechenden linux-header gefehlt !
Bei mir waren es linux618-header und linux73-header.
Anschliessend mkinitcpio -P nicht vergessen (so wie ich)
und siehe da:

kt 31 14:14:08 manjaro kernel: Kernel command line: BOOT_IMAGE=/boot/vmlinuz-7.3-x86_64 root=UUID=5d4a4f03-c2b9-40b6-8d24-00897eadfd31 rw mitigations=off split_lock_detect=off gpu_unlock.targets=0000:01:00.0 gpu_unlock.ad103_experimental=1 quiet loglevel=3 rd.udev.log_level=3
Okt 31 14:14:08 manjaro kernel: gpu_unlock: loading out-of-tree module taints kernel.
Okt 31 14:14:08 manjaro kernel: gpu_unlock: module verification failed: signature and/or required key missing - tainting kernel
Okt 31 14:14:08 manjaro kernel: gpu_unlock: unknown parameter 'ad103_experimental' ignored
Okt 31 14:14:08 manjaro kernel: pci 0000:01:00.0: gpu_unlock: stage=claim device=0x2705 family=AD102/AD103/AD104/AD106/AD107 cmd=0x3 original=0x3 bar_len=0x1000000 region=1 enabled=0 master=0
Okt 31 14:14:08 manjaro kernel: pci 0000:01:00.0: gpu_unlock: stage=layout fb=0x3ff800000 vga=0x3ff700000 frts=0x100000 wpr=0x3f4c00000..0x3ff700000 boot=0x3ff5f7000 gsp=0x3fa5a0000 heap=0x3f4d00000
Okt 31 14:14:08 manjaro kernel: pci 0000:01:00.0: gpu_unlock: stage=dma booter=56832 payload=0xfa00 pages=20606 meta=0x33eff000 sig=0x1d410000 stock=0x1d420000 boot=0x1d440000
Okt 31 14:14:08 manjaro kernel: pci 0000:01:00.0: gpu_unlock: stage=gfw plm=0xffffff8f ss0=0x57436035 ss1=0x0 wpr2_hi=0x0
Okt 31 14:14:08 manjaro kernel: pci 0000:01:00.0: gpu_unlock: stage=fwsec version=3 fuse=0x1 ver=1 scratch=0x0 wpr2_raw=0x3ff6000:0x3ff6e00 frts=0x3ff6000:0x3ff6e00
Okt 31 14:14:08 manjaro kernel: pci 0000:01:00.0: gpu_unlock: stage=publish meta=0x33eff000 sig=0x1d410000 size=0xfa00 radix=0x1d421000
Okt 31 14:14:08 manjaro kernel: pci 0000:01:00.0: gpu_unlock: pass=1/4 attempt=1 target=0x1fa7cc value=0xfffff0ff before=0x4cb8f after=0xfffff0ff wpr2=0x3f4c000:0x3ff6e00 reset_plm=0xffffffff mbox=0x0:0x67b00001 ret=0
Okt 31 14:14:08 manjaro kernel: pci 0000:01:00.0: gpu_unlock: pass=2/4 attempt=1 target=0x9a0148 value=0xffffffff before=0x4cb8f after=0xffffff0f wpr2=0x3f4c000:0x3ff6e00 reset_plm=0xffffffff mbox=0x0:0x67b00001 ret=0
Okt 31 14:14:08 manjaro kernel: pci 0000:01:00.0: gpu_unlock: pass=3/4 attempt=1 target=0x1fa7c4 value=0xffffffff before=0x4cb8f after=0xffffffff wpr2=0x3f4c000:0x3ff6e00 reset_plm=0xffffffff mbox=0x0:0x67b00001 ret=0
Okt 31 14:14:08 manjaro kernel: pci 0000:01:00.0: gpu_unlock: pass=4/4 attempt=1 target=0x823804 value=0xffffffff before=0x4cb8f after=0xffffffff wpr2=0x3f4c000:0x3ff6e00 reset_plm=0xffffffff mbox=0x0:0x67b00001 ret=0
Okt 31 14:14:08 manjaro kernel: pci 0000:01:00.0: gpu_unlock: stage=host ss0=0x88888888 ss1=0x8 plm_feat=0xffffffff plm_wpr_cfg=0xfffff0ff
Okt 31 14:14:09 manjaro kernel: pci 0000:01:00.0: gpu_unlock: stage=flr post_wpr2=0x1ffffe00:0x0 ss0=0x88888888 ss1=0x8 plm=0xffffffff cmd=0x3 devsta=0x0 chip=0x193000a1
Okt 31 14:14:09 manjaro kernel: pci 0000:01:00.0: gpu_unlock: stage=handoff ret=0 state=4 retained=0 pages=0/20606 cleanup=0
Okt 31 14:14:09 manjaro kernel: pci 0000:01:00.0: gpu_unlock: unlocked
Okt 31 14:14:09 manjaro systemd-modules-load[348]: Inserted module 'gpu_unlock'

Mod edit: Aufeinanderfolgende Beiträge zusammengeführt.