justaresearcher/gpu-unlock-ada provides a Linux kernel module and a firmware generation tool that change the tensor operation path of NVIDIA graphics cards. In v1.1.1 the project published tensor operation issue rate measurements together with the artifacts and hashes needed for reproduction. Under the same power limit condition, the TF32 issue rate was reported as 1.99x and CUTLASS FP8 GEMM as 1.95x. These are measurements of instruction issue rate and do not mean application performance improves.
The project's approach covers the graphics card firmware path and the Linux kernel module together. The README links single-device measurements and raw records for the RTX 4090 and the experimental RTX 4070 Ti SUPER. That makes it possible to follow the published figures and reproduction materials to look at the tensor operation path of a specific card. The implementation language is C and the license is GPL-2.0.
The README, however, distinguishes that a rise in some instruction throughput does not necessarily carry over to real application performance at the same scale. Practical GEMM acceleration has not been demonstrated. The Windows-related port is also treated as a separate experiment, so it should not be confused with the Linux module's supported targets. What the current material backs is issue rate measurement on a limited set of cards, and gains on real workloads still need follow-up verification.