A GitHub project named gpu-unlock-ada, published on 28 October 2026 according to its GitHub page, offers a Linux kernel module that its maintainer says roughly doubles FP16 and BF16 tensor throughput on the NVIDIA GeForce RTX 4090. The figures come from the project's own runs and linked owner reports, so they are claims to weigh rather than settled results.
What does the module change on the card?
The module enables a speed-select override before NVIDIA's stock driver starts, according to the project's README. It runs a firmware sequence and a payload, performs a function-level reset, and only then lets the unmodified NVIDIA driver attach to the card. The project says the AD102 payload comes from the TOTORONG/AD103_win_unlocker project and that the result depends on it.
The main measured change is instruction throughput. In one matched run on an RTX 4090 with NVIDIA driver 615.78.08 and a 450 W cap, a Linux owner recorded FP16/FP32 instruction throughput rising from 160.4 to 320.4 TF/s, and BF16/FP32 from 159.7 to 320.2 TF/s. A separate modified 48 GB RTX 4090 on Ubuntu 26.04 showed a similar jump, from 164.9 to 328.9 TF/s for FP16/FP32.
Do the gains reach real workloads?
Only partly, according to the source. Real cuBLAS matrix multiplication median throughput varied by shape. FP16 gained between 12 and 34 percent, and BF16 gained 16 to 20 percent at two shapes but was essentially unchanged at 4096 cubed. INT8, dp4a and scalar FP32 stayed near their baseline rates in the reports.
The clearest negative result comes from the modified 48 GB card. A quantized mixture-of-experts serving workload showed no gain, with 8K prefill at 0.983 times baseline. The project suggests quantized matrix multiplications do not use the FP32-accumulate tensor path, but that is its own explanation and the source does not test it further.
FP8 results are mixed. On a separate RTX 4070 Ti SUPER, a CUTLASS E4M3 GEMM moved by only 0.3 percent, which the project says is because that kernel reaches about 17 percent of the card's FP8 peak. Other reports the project cites show larger FP8 gains on other cards, and the project itself flags that variance.
Who it is for and where it falls short
The source describes a narrow audience: Linux users with a supported NVIDIA card who can power the machine fully off before the first attempt. The project's requirements and limits include the following.
- Linux only. The Windows reference port from Latence.app is described as one manual boot on one MSI RTX 4070 Ti AD104 with one motherboard profile, and the project says it does not add support to the Linux module.
- A full power-off before the first attempt. The source says a warm reboot is not guaranteed to work, and loading the module after NVIDIA has bound the card does not unlock it.
- Kernel headers and a working DKMS toolchain, plus firmware the user builds locally from pinned NVIDIA inputs. The project says it does not distribute prebuilt firmware.
- Experimental support for the RTX 4070 Ti SUPER, where each device ID needs its own opt-in. Laptops, other memory configurations and professional or data-centre cards are not validated, and the project says some are not accepted.
Maturity is the main concern. The results come from one or a few cards per device ID. The project describes one matched run pair with the desktop active as not an extended stability test. The source does not say how the unlock behaves over long periods or under sustained load.
The source also does not address warranty coverage or NVIDIA's position on modifying firmware. Readers who own a card should check those points before they proceed.
How should readers weigh the project's evidence?
The repository presents its numbers as the maker's and owners' results, and it links to a technical report and raw benchmark output. The project says its first transaction came before a repository cleanup. ToolHub Daily has not run or inspected the module, so the figures here are reported as the project states them.
Bottom line: The project publishes detailed benchmark data for an unofficial RTX 4090 unlock, but the evidence covers few cards and mostly register-level issue rates. Treat the speedups as unproven for your own workload, and wait for independent, repeated tests before changing firmware on a card you rely on.
Source and verification: The information comes from the GitHub repository JustAResearcher/gpu-unlock-ada, which lists a publication date of 28 October 2026. Repository figures are from the GitHub API as of 30 October 2026: 4,543 stars, 580 forks, GPL-2.0 licence, C as the main language, latest release v1.1.1 dated 24 October 2026, and last commit pushed 26 October 2026. Benchmark figures are the project's and its linked owner reports, not tests by ToolHub Daily.
Related: AirDrop on Linux: open-source omdrop-owl targets non-Apple Wi-Fi cards · New Free Driver Brings NVIDIA RTX Cards to macOS Sequoia Intel Macs
Comments (0)