Three days and 3,704 stars. The project gpu-unlock-ada was born on the night of Friday, October 23, and by this Monday it has accumulated 489 forks, fourteen open issues and more than 1,100 new stars in 24 hours. What has driven the attention is not a new graphics card, but a Linux kernel module that makes a 2022 RTX 4090 perform twice as fast in half of its tensor operations.
The mechanism has little magic and a lot of firmware. The module loads from the initramfs and acts before the NVIDIA driver takes control of the card, runs a signed chain that opens internal permissions and writes the chip's speed selector. With that, the card issues tensor core instructions in FP16 and BF16 at twice the rate when they accumulate in FP32: 165.6 to 331.6 TF/s in the first mode and 165.5 to 330.6 TF/s in the second.
The figure matters because the RTX 4090's tensor cores are today the engine behind much of the local AI that runs at home, and because doubling the issue rate is not the same as doubling the speed of a model. The repository itself says so without embellishment: these are instruction rates measured with data in registers, not speedups of a full application. What is on the table is an open, reproducible experiment with more questions than certainties.
Unlocking more power for your RXT 4090
The module does not patch the driver or touch the UEFI. It takes advantage of a gap at boot, with the card cold and no driver loaded, to launch the signed firmware that NVIDIA already ships inside the card and make it execute four privileged writes.
It then sets the internal selectors, forces a PCIe function reset and lets the NVIDIA driver initialize the card as on any other day. The missing piece was not the Linux loader, inherited from Zagitta/gpu-unlock, but the map of the interior: the AD102 payload comes from analysis of the debug booter that TOTORONG published for the AD103 family and that here is translated to the addresses of the larger chip.
The repository does not distribute NVIDIA firmware either. It generates the nine files it needs on the user's machine from inputs pinned by SHA-256 of version 610.43.03, and it is published under GPL-2.0 with the module written in C, just like the project from which it inherits the loader.
The numbers the repository puts on the table
The first measurement, from October 23, uses the same benchmark cores, the same launch geometry and the same compilation flags in both states, with a 450 W cap and 128 active SMs. The contrast is better seen in a table than in prose, because what matters is that only part of the numerical matrix moves.
| Instruction mode |
Before |
Unlocked |
| FP16 with FP32 accumulation |
165.6 TF/s |
331.6 TF/s |
| BF16 with FP32 accumulation |
165.5 TF/s |
330.6 TF/s |
| INT8 with INT32 accumulation |
661.6 TOP/s |
659.5 TOP/s |
| dp4a |
164.5 TOP/s |
164.1 TOP/s |
| FP32 scalar (ffma) |
79.1 TF/s |
79.1 TF/s |
An independent reproduction, on revision `8855e7b` and Arch kernel 7.2.11 with driver 615.78.08, repeated the result: 160.4 to 320.4 TF/s in FP16 and 159.7 to 320.2 in BF16. There the first real brake appears. With cuBLAS and its CUDA requirements on RISC-Vthe same CUDA stack used by PyTorch and almost everything else, the median gain for a large GEMM was 12 to 34% in FP16, 16 to 20% in BF16 in two shapes and nil at 4096³.
On October 25, a modified 48 GB card with Ubuntu 26.04 and driver 610.57.04 was added. The `fmla32` register went from 1 to 0, the rate rose from 164.9 to 328.9 TF/s in FP16 and from 164.8 to 328.7 in BF16, with no Xid errors in the journal. And a quantized mixture-of-experts service improved nothing, 0.983 times in 8K prefill, which is consistent with quantized matmuls not going through the FP32 accumulation path.
The nuance that separates a 2x from a 10%
On the same day as the star peak, October 26, a user posted in the original repository a report with a factory-stock ASUS RTX 4090 that dismantles any triumphalist reading.
The unlock worked and the kernel logged `gpu_unlock: unlocked`, but the numbers were far more modest: an FP16 GEMM went from 162.1 to 178.8 TF/s, a 10%, BF16 from 160.2 to 188.1, a 17%, and TF32 from 84.4 to 131.8, a 56% that looks nothing like the promised doubling. Processing a 14,600-token prompt with a quantized 27-billion-parameter model barely moved, 2.6%, and generating 256 tokens fell 7%.
That contrast does not invalidate the project, but it does narrow what it promises. Two cards of the same model, with the same power and the same driver, give results ranging from 10% to 100% depending on the workload, the shape of the GEMM and the accumulation mode the library chooses. The author of that report admits he has not investigated whether the blame lies with torch's accumulation path, the clocks or the 450 W limit. No one has the answer yet.
The RTX 4070 Ti SUPER, the second test bench
The project's other leg is experimental and is labeled as such. An RTX 4070 Ti SUPER with identifier `10de:2705`, 66 SM and a 240 W cap went from 95.0 to 189.9 TF/s in FP16 and from 95.0 to 189.7 in BF16. INT8 stayed at 379.5 TOP/s, a figure identical to the stock one, and TF32 nearly doubled its rate to 94.51 TF/s.
Unsigned INT4 held at 754.8 TOP/s in both states, and an E4M3 GEMM with CUTLASS measured 59.326 TF/s stock and 59.495 unlocked, a 0.3% difference that does not distinguish an instruction issue rate.
Each Ada identifier requires its own enable flag, and those marked as exact must be named by their full PCI address: `targets=all` never selects them. The repository collects results from one card or a few per model, warns that laptop variants depend on the VBIOS and rejects the professional RTX A cards and data center cards.
What does whoever tries it gain and what do they risk?
The gain is concrete for anyone who already has an Ada and does fp16 or bf16 with FP32 accumulation: small training, fine-tuning, inference without quantization. The price is too, and it does not appear in any performance table.
- The module is open source, but it lives outside the kernel tree and ships unsigned, so Secure Boot has to be off.
- The project documents how to roll back: if boot fails, the parameter must be removed from the kernel command line before recovering the card.
- The published evidence covers one or a few cards per identifier, with short stability tests and no representative training workloads.
- The warranty, NVIDIA's terms of use and any mining use remain entirely on the side of whoever runs the module.
Consumer hacking goes where local AI is
The pattern repeats every generation. A few years ago the community unlocked CMP mining cards to expand their memory; in 2026 the target is the tensor cores, because the money and the curiosity have moved to local AI, a market where capable hardware still costs an arm and a leg, as the article on local AI on your PC for 4,000 euros. That a three-day-old kernel module gathers 3,700 stars says less about the code than about the appetite behind it.
The hole is not in a graphics driver but in how the card's firmware boots, and closing it completely would mean reviewing that boot chain. Until that happens, what remains is a fascinating and fragile tool, with a good part of its results still unreproduced in other hands. We will have to wait for the coming weeks to know whether the 2x holds up across a broad fleet of cards or remains the anecdote of a few test benches.