Viral claim about GPU power boost

A viral post claims to double the performance of all RTX 40 graphics cards without changing the BIOS. The project, called gpu-unlock-ada, is hosted on GitHub under the name JustAResearcher. The project's initial results using an RTX 4090 were reported on October 23, but it does not improve gaming performance, only AI-related tasks.

How the module works

The module is a Linux kernel module that changes a setting on the GPU at startup, before the official Nvidia driver loads. This setting affects the Tensor Cores, which are responsible for AI calculations. These cores handle mathematical operations with high precision to avoid rounding errors. Nvidia has been limiting this feature to half power on GeForce cards, but the module restores full power.

Limitations of the module

The project's claim that it can turn an RTX 4090 into an RTX 4060 Ti is not accurate. The module works only on the RTX 4090 and, in experimental mode, the RTX 4070 Ti Super. A Windows version targeting a modified RTX 4080 with 32 GB was published separately. Performance improvements are confirmed on these specific cards, but only in controlled tests without memory access.

Real-world performance gains

On the RTX 4070 Ti Super, the performance gain for real matrix calculations in FP8 is minimal, at 0.3%. For INT8, there is no improvement on either the RTX 4070 Ti Super or the RTX 4090. On FP16 or BF16 calculations, the gains exist but decrease with larger matrix sizes. On a modified RTX 4080, the gain reaches 100% for small matrices but drops to between 60 and 71% for larger ones.

User feedback on older GPUs

Users with RTX 30 series cards report modest gains, up to 36%, when running large language models. However, there is little improvement during text generation due to memory limitations, and no gains in gaming. Some users found the performance before and after the module to be identical when tested closely.

Steps to install the module

The installation process involves identifying the GPU using a command-line tool, installing necessary software, cloning the project from GitHub, and configuring the system to load the module. Users are advised to keep a normal boot option in case of issues. The module requires the GPU to be cold (unpowered) during installation.

Risks and warnings

The module bypasses Nvidia's secure boot process, which the company has not validated. There are risks of boot failure, display issues, data corruption, or hardware damage. The developer admits that the accuracy of calculations has not been tested on real workloads for the RTX 4090. For gamers, the benefits are negligible, but for AI researchers using the RTX 4090 in Linux, the effort might be worthwhile. Users are warned to avoid unofficial mirrors and stick to the original repositories on GitHub and Hugging Face.