New to this project? Start with the copy-paste Quick Start.

This Linux kernel module enables the SM speed-select override before the stock NVIDIA driver starts. RTX 4090 (10de:2684, AD102) support has demonstrated approximately 2× FP16 and BF16 tensor throughput with FP32 accumulation:

Mode	Before	After
FP16 → FP32 accumulate	165.6 TF/s	331.6 TF/s
BF16 → FP32 accumulate	165.5 TF/s	330.6 TF/s

The override survived PCIe Function Level Reset (FLR), stock NVIDIA 615.78.08 initialization and CUDA benchmarks. INT8, dp4a and scalar FP32 stayed near their baseline rates. The original transaction preceded the repository cleanup; see the technical report and original evidence.

An independent owner run of revision 8855e7b also completed the RTX 4090 transaction and NVIDIA driver handoff on Linux 7.2.11-arch1-1. Matched stock/unlocked measurements with NVIDIA 615.78.08 and the same 450 W cap found FP16/FP32 instruction throughput 160.4 → 320.4 TF/s and BF16/FP32 159.7 → 320.2 TF/s. All exercised numerical checks passed. Real cuBLAS GEMM median throughput varied by shape: FP16 gained 12–34%, BF16 gained 16–20% at two shapes and was essentially unchanged at 4096³. These are one matched run pair with the desktop active, not whole-application speedups or an extended stability test. See the matched evidence and limits and testing workflow.

A modified 48 GB RTX 4090 (10de:2684, AD102) on Ubuntu 26.04 with Linux 7.0.0-38-generic and NVIDIA 610.57.04 open kernel modules also completed the transaction and driver handoff on 2026-10-25: fmla32 went from 1 to 0, FP16/FP32 instruction throughput 164.9 → 328.9 TF/s, BF16/FP32 164.8 → 328.7 TF/s, with INT8, dp4a and scalar unchanged and no Xid. A quantized mixture-of-experts serving workload on the same card showed no gain (8K prefill 0.983×), which is consistent with quantized matmuls not consuming the FP32-accumulate tensor path. That is one card, one cold boot and one workload. See the modified-card evidence and limits. Further matched owner reproductions used NVIDIA 610.57.04 and 595.91.07.

A separate cold test of the Ada port on one RTX 4070 Ti SUPER (10de:2705) with NVIDIA 580.178.04 measured FP16/FP32 tensor issue rate at 95.0 → 189.9 TF/s and BF16/FP32 at 95.0 → 189.7 TF/s under the same 240 W power cap. INT8 stayed at 379.5 TOP/s. The stock driver loaded, the NVIDIA RM getter reported full FMLA32 rate, deterministic FP16/BF16 GEMM outputs matched exactly, and owner mining recovered with fresh accepted shares. A second matched instruction benchmark found TF32/FP32 47.49 → 94.51 TF/s (1.99×), with its FP16/FP32 control also near 2×. Unsigned INT4/INT32 stayed at approximately 754.8 TOP/s in both states. These are register-resident issue rates, not application GEMM or inference speedups. A matched, verified E4M3 FP8/FP32 CUTLASS GEMM measured 59.326 TF/s stock → 59.495 TF/s unlocked on this card, a 0.3% difference. That kernel reaches only about 17% of this card's FP8 peak, so it cannot resolve an issue-rate change. Other measurements do put FP8 in the affected group: an 8192³ E4M3 GEMM on a Windows RTX 4080 SUPER measured 199.4 → 310.4 TF/s (+55.7%) and 199.4 → 298.7 TF/s (+49.8%) in two runs (see goldlionren/gpu-unlock-4080s; its author flags FP8 variance and keeps FP8 out of that repo's own pass/fail gate), and register-resident E4M3 probes read 603.1 TF/s unlocked on the modified 48 GB RTX 4090 (see that evidence directory) and 165.45 → 323.95 TF/s (1.96×) on a Windows RTX 4070 Ti SUPER (see A-XIANGQAQ/rtx4070tis-ad103-tensor-unlock). The 4090 figure is 91% of its 660.6 TF/s full FP8 rate; NVIDIA rates a stock RTX 4090 at 330.3 TF/s for FP8 with FP32 accumulate and 660.6 with FP16 accumulate. So FP8 does move at the issue-rate level; a matched AD103 pair is 1.96×. At a fixed power limit the application GEMM stays below that: an unlocked RTX 4090 peaked at 346 W against 313 W stock and spent both MMA and GEMM passes under SW Power Cap (see the 610.57.04 evidence). This is one card; the unlock succeeded on two cold boots and two ordinary warm reboots, the last from a persistent EFI default. Throughput was measured on the first cold boot. See the 2705 canary report and its sanitized raw benchmark output.

The AD102 payload and its debug-booter analysis come from TOTORONG/AD103_win_unlocker. This fork integrates that published payload into Zagitta/gpu-unlock's Linux module, firmware generation and boot packaging. The payload is essential to the result.

Experimental Windows contribution — Latence.app

A separate MSI RTX 4070 Ti AD104 Windows UEFI reference port records one successful manual boot with NVIDIA 616.56 and matched cuBLAS FP16/FP32 +80.1%, BF16/FP32 +76.4% measurements. It is tied to one captured MSI Z490 profile; it does not add AD104 admission to the Linux module. See the report and limitations. No firmware/Windows/ROM binaries or automatic boot installer are distributed. Contribution by Latence.app; original upstream credits and licences are retained.

Supported targets

The module selects the AD102 firmware and payload automatically for RTX 4090 10de:2684. The original GeForce/CMP GA102/GA104 device list and GA102 payload are retained, along with the newer GA103/GA106 device list. Upstream reported RTX 3090 success; this branch does not revalidate every legacy board. GA107 RTX 3050 6GB (10de:2584) uses the same GA102 Booter path and was unlocked on one card (2584 report).

Experimental Ada targets

Each ID below needs its own opt-in flag. Rows marked exact must also be named by their exact PCI address in gpu_unlock.targets=, alone or in a comma-separated list, and targets=all never selects them. All of them use the RTX 4090's pinned AD102-family firmware and payload. Results come from one or a few cards per ID; other boards, memory configurations and laptop models remain unverified. tools/print_boot_option.py prints the exact option for any card listed here.

PCI ID	Card	Die	Opt-in	Tested	Details
2685	RTX 4090 D	AD102	rtx4090d_experimental	eight cards in one host	#7
2689	RTX 4070 Ti SUPER	AD102	rtx4070tis_ad102_experimental, exact	one card, 30-minute soak	#19
2702	RTX 4080 SUPER	AD103	ad103_experimental, exact	one card, matched mma and cuBLAS, alongside a 2704 in one host	matched evidence, #53
2704	RTX 4080	AD103	rtx4080_experimental, exact	one modified 32 GiB card; a second card in matched evidence	report
2705	RTX 4070 Ti SUPER	AD103	ad103_experimental, exact	one card, matched stock/unlocked	report
2709	RTX 4070	AD103	ad103_experimental, exact	one card, matched	report
2757	RTX 4090 Laptop GPU	AD103M	ad103m_experimental, exact	Clevo: unlocked, no throughput change; Lenovo: 2× on two cold boots	report, Lenovo
2782	RTX 4070 Ti	AD104	ad104_experimental, exact	one card in a VFIO guest, matched; one bare-metal owner report	report, #15
2783	RTX 4070 SUPER	AD104	ad104_experimental, exact	one card, matched	report
2786	RTX 4070	AD104	ad104_experimental, exact	one card; three owner reports	report, #11
2803	RTX 4060 Ti	AD106	ad106_experimental, exact	one 16 GB card, matched; two more boards in #34	report, #34
2805	RTX 4060 Ti 16GB	AD106	ad106_experimental, exact	one card, no matched throughput run	report
2882	RTX 4060	AD107	ad107_experimental, exact	one card, matched cuBLAS; cold boot and warm reboot	report
28a0	RTX 4060 Laptop GPU	AD107	ad107_experimental, exact	one card, unlocked state only	report
28e0	RTX 4060 Laptop GPU	AD107	ad107_experimental, exact	one owner report	#10
28e1	RTX 4050 Laptop GPU	AD107	ad107_experimental, exact	one laptop, matched	report

Laptops differ in VBIOS ROM exposure, embedded-controller power rails and FLR handling, so a result on one model does not carry over to another. Other Ada IDs, professional RTX A-series and data-centre parts are not accepted. Shared firmware across Ada chips does not establish support for untested cards.

The RTX 4090 path runs the signed firmware, checks the payload's completion marker and required permissions, writes SS0 0x88888888 and SS1 0x8, then performs FLR before the unmodified NVIDIA driver binds. It requires the expected effective speed-select readout before and after FLR. No patched NVIDIA driver or UEFI application is needed. The payload uses instruction addresses derived from published decrypted debug-booter analysis; the production booter is authenticated and decrypted inside the GPU.

Build and install

Clone the GPL-2.0-only source and build the packages locally so the pinned NVIDIA firmware is downloaded and generated on your own machine:

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

make install detects the supported distribution family. Its explicit targets are:

make install-arch   # makepkg -sif
make install-deb    # apt-get build-dep, dpkg-buildpackage, apt-get install
make install-rpm    # dnf builddep, rpmbuild, dnf install

make arch, make deb and make rpm build without installing. Packages are built from the committed HEAD; Debian and RPM outputs go to build/deb and build/rpm. Kernel headers and a working DKMS toolchain are required.

Builds fetch the pinned NVIDIA 610.43.03 source and binary inputs over HTTPS and verify their SHA-256 hashes. make inputs prepares the cache; make firmware generates assets from cached inputs; make assets does both. The default cache is $HOME/.cache/gpu-unlock and can be overridden with CACHE=, INPUTS= and ASSETS=. A clean-chroot Debian build needs GPU_UNLOCK_INPUTS pointing to the input cache because dpkg-buildpackage does not fetch inputs itself.

The source archive produced by make dist, the Debian source package and the .nosrc.rpm contain no NVIDIA archives or generated firmware. See firmware/README.md for the input and asset layout. Prebuilt firmware packages are not distributed by this repository.

Enable at boot

The package installs the module and firmware into the initramfs so the unlock runs before NVIDIA binds the card. It remains inactive unless the kernel command line opts in:

gpu_unlock.targets=0000:03:00.0
gpu_unlock.targets=0000:03:00.0,0000:04:00.0
gpu_unlock.targets=all
gpu_unlock.targets=0000:01:00.0 gpu_unlock.ad103_experimental=1

These PCI addresses are examples. Use the full PCI names printed by lspci -D. all selects every accepted unbound card, including legacy GA102/GA104 targets; an explicit address is useful when selecting one RTX 4090. Do not mix all with addresses. Every experimental Ada target requires its own flag from the experimental target table, and the exact ones must be named by their exact PCI address; all never selects them. Invalid entries are logged and ignored.

The packages support these initramfs generators:

• mkinitcpio: /etc/mkinitcpio.conf.d/zz-gpu-unlock.conf puts gpu_unlock immediately before nvidia in MODULES, or last if NVIDIA is absent. Images are rebuilt on installation. A custom configuration selected with -c or a preset _config does not read drop-ins; add gpu_unlock before nvidia in that configuration's MODULES yourself.
• initramfs-tools: /usr/share/initramfs-tools/modules.d/gpu-unlock.conf.
• dracut: /usr/lib/dracut/dracut.conf.d/gpu-unlock.conf.

There is no softdep nvidia pre: rule because kmod uses only the first matching soft dependency line, which can conflict with other packages.

Power the machine off and start it again for the first attempt. The module needs a card whose firmware state is fresh before NVIDIA loads. A warm reboot is not guaranteed to provide that state: it did on the tested RTX 4070 Ti SUPER, where the pre-firmware WPR2 readback returned to zero and the unlock succeeded again without an AC cut. Other boards may retain WPR2; the module refuses to rerun the firmware sequence in that state. It also skips cards that are already bound or powered down. If board firmware left bus mastering enabled, the module clears it and verifies that it stayed cleared; the later WPR2 check still rejects an already initialized card. Loading the module manually after NVIDIA has bound the card cannot unlock it.

After validating a cold boot and a one-time warm reboot on the exact card, a persistent bootloader entry can select the same early initramfs and exact gpu_unlock.* options on every boot. A one-shot entry by itself does not persist. Keep an unmodified stock entry available, and verify gpu_unlock: unlocked, the rate readback and GPU health after each boot. The tested canary uses a kernel-version-specific image; regenerate and revalidate that image before removing its pinned kernel during an upgrade.

Avoid nvidia-smi, CUDA, NVML and nvidia-modprobe before a manual cold-card attempt: they can load NVIDIA by name. Ordinary modprobe blacklist rules do not block explicit named loads. The packaged initramfs flow avoids the need for manual firmware staging and research parameters.

The original display-card handoff is retained: the firmware framebuffer is removed just before FLR, and the display stays blank until nvidia-drm takes over. Display handoff and multi-card operation were not part of the RTX 4090 validation.

Check and disable

Inspect the kernel log after boot:

journalctl -k -b | grep gpu_unlock
lspci -nnk -s 0000:03:00.0  # replace the example address

A loaded module or successful modprobe return alone does not prove an unlock: look for the selected card's gpu_unlock: unlocked message and the payload, selector and post-FLR checks. The report describes independent register and performance confirmation. The optional read-only sampler lives in diagnostics/peek; the CUDA instruction benchmark is in diagnostics/bench. Never load a second BAR0 sampler while gpu_unlock is performing its firmware transaction.

If the module reports that firmware did not stop and its DMA buffers are being kept, power off before trying again. The module intentionally retains buffers that firmware may still use. Do not force-unload it in this state.

Remove gpu_unlock.targets= from the kernel command line to disable future attempts. It can be removed for one boot in the bootloader editor. Removing the packages removes the module and initramfs integration; power off to return to a fresh hardware state.

Development and credits

make test runs the host-side suite without accessing a GPU. Missing generated firmware is reported as an explicit skip of the full transaction tests. After make assets, use make test-full for a strict check that requires both the pinned source and all generated assets. Override ASSETS= and NVIDIA_SOURCE= for inputs built elsewhere. These tests mock firmware execution and device behavior; they do not replace a hardware repeat. See testing and interpreting results for the validation levels, matched benchmark commands and recorded hardware evidence.

The kernel verifies each loaded asset against its pinned SHA-256 before claiming the card or starting firmware DMA, including files staged into an initramfs. The kernel's SHA-256 library (CONFIG_CRYPTO_LIB_SHA256) must be available. make -C kernel/gpu_unlock builds against the running kernel's headers; use LLVM=1 when those headers support a Clang build.

The Linux module and packaging derive from Zagitta/gpu-unlock. The AD102 four-PLM payload derives from TOTORONG/AD103_win_unlocker, including its published payload generator and debug-booter analysis. The GA102 V67 chain derives from WildFlash1st/cmp90hx-unlock. The earlier CMP unlock work, NVIDIA's published loader code, Nouveau and envytools also underpin the loader and investigation.

The kernel module, packaging and payload integration are GPL-2.0-only. NVIDIA firmware remains under its own terms. See firmware/NOTICE.md and licenses/ for provenance and preserved license notices.