GPU Unlock for RTX 4090 and RTX 40 series

Q: Can I use targets=all for an RTX 4070 Ti SUPER?

No. The experimental cards are never selected by all. They need their exact PCI address and their experimental flag.

GPU Unlock Ada is a fork of GPU Unlock by JustAResearcher that brings the same full-rate tensor unlock to the RTX 4090, with experimental support for the RTX 4070 Ti SUPER and RTX 4060 Ti 16GB. It is Linux-only, like the original.

GPU Unlock Ada 1.1.1, GPL-2.0-only.

GPU Unlock Ada on GitHub Official quick start

RTX 4090 results

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

An independent owner run at the same 450 W cap measured FP16 160.4 → 320.4 TF/s and BF16 159.7 → 320.2 TF/s in instruction throughput. In real cuBLAS GEMM the gain was smaller and depended on matrix shape: FP16 +12–34%, BF16 +16–20% at two shapes and about unchanged at 4096³. INT8, dp4a and scalar FP32 stayed near their normal rates. The project stresses that these are single runs, not whole-application speedups or a long stability test.

Which RTX 40 cards are supported

Chip	Card	PCI ID	Status	Boot option
AD102	RTX 4090	`10de:2684`	Supported; about 2× FP16/BF16 tensor issue rate measured	gpu_unlock.targets=<address>
AD102	RTX 4090 D	`10de:2685`	Experimental	add gpu_unlock.rtx4090d_experimental=1
AD103	RTX 4070 Ti SUPER	`10de:2705`	Experimental, validated on one card	add gpu_unlock.ad103_experimental=1
AD106	RTX 4060 Ti 16GB	`10de:2805`	Experimental, unlocked on one card, no throughput run yet	add gpu_unlock.ad106_experimental=1

The RTX 4070 Ti SUPER measured 95.0 → 189.9 TF/s FP16 and TF32 47.49 → 94.51 TF/s on one card at 240 W. The RTX 4060 Ti 16GB switched to full rate on one card but has no throughput run yet. Professional RTX A-series and data-centre parts are not accepted.

How to install GPU Unlock Ada

Find your card’s address and ID: lspci -Dnn | grep -i nvidia. Check it is 10de:2684 (4090), 10de:2705 (4070 Ti SUPER) or 10de:2805 (4060 Ti 16GB).
Install Git, your kernel headers, a C toolchain, DKMS, Python 3 and zstd, then build:
```
git clone https://github.com/JustAResearcher/gpu-unlock-ada.git
cd gpu-unlock-ada
make install
```
The build downloads SHA-pinned NVIDIA 610.43.03 inputs and generates the firmware locally. It can take several minutes.
Print the exact boot option for your card with python3 tools/print_boot_option.py, or write it yourself:
```
gpu_unlock.targets=0000:01:00.0                                  # RTX 4090
gpu_unlock.targets=0000:01:00.0 gpu_unlock.ad103_experimental=1  # RTX 4070 Ti SUPER
gpu_unlock.targets=0000:01:00.0 gpu_unlock.ad106_experimental=1  # RTX 4060 Ti 16GB
```
Use your own address, not the example. Add it with the same bootloader steps as in the main install guide, and keep a normal boot entry available.
Shut down, cut AC power, wait about 30 seconds, then start the machine.
Check the result:
```
journalctl -k -b --no-pager | grep -E 'gpu_unlock: unlocked|gpu_unlock: unlock failed|post-FLR SEC2'
nvidia-smi
```
Success means your card’s gpu_unlock: unlocked line, a healthy nvidia-smi and a workload check. A loaded module alone is not proof.

Avoid running nvidia-smi, CUDA or NVML before a manual cold-card attempt: they can load the NVIDIA driver first. Test one ordinary reboot before making the boot entry permanent.

RTX 4090 questions

Does GPU Unlock work on the RTX 4090?

Yes, through the GPU Unlock Ada fork. On the RTX 4090 it measured about twice the FP16 and BF16 tensor throughput with FP32 accumulation.

Does it work on other RTX 40 cards?

The RTX 4070 Ti SUPER and RTX 4060 Ti 16GB are experimental, each validated on one card, and need their own flag. The RTX 4090 D is experimental too. Other Ada cards are not enabled.

Can I use targets=all for an RTX 4070 Ti SUPER?

No. The experimental cards are never selected by all. They need their exact PCI address and their experimental flag.

Does FP8 get faster on the 4070 Ti SUPER?

One matched FP8 GEMM test was effectively unchanged, 59.326 → 59.495 TF/s. Other FP8 kernels are untested.

Why do I need a full power-off?

The module needs the card’s firmware in a fresh state before the NVIDIA driver loads. A warm reboot does not always give that state, so the first attempt should start from a full power-off.

Back to the GPU Unlock guide or the full GPU list.