Two open-source projects, GPU-unlock and GPU-unlock-ADA, have been released. They remove a limit that Nvidia applies in software to matrix compute speed on its RTX 30 (GA102, GA103, GA104, GA106 chips) and RTX 40 gaming cards.
Tensor cores rely on exactly this kind of math when training and running neural networks. The limit is widely seen as a way to keep gaming cards from competing with professional accelerators in AI workloads.
How it works
The unlocker loads before the graphics driver and does not modify the driver or firmware. It builds on earlier research into unlocking CMP mining cards. The main version targets Linux, and an experimental Windows variant that runs through UEFI is in progress.
What it delivers
According to the authors, tensor core throughput in some modes such as FP32 and BF16 rises by up to 60%, depending on the card and workload. Local image and video generation gets almost 60% faster. The Windows port has so far been tested only on an RTX 3090, where FP16 went from 70.6 to 125.9 teraflops.
Limitations
Video memory bandwidth stays the same and no extra CUDA cores are unlocked. That is why LLM text generation, which is bound by memory, barely speeds up. No effect on gaming performance is claimed either.
Commentary
For anyone running neural networks locally this is an interesting experiment, but not a ready tool for a work machine. The module bypasses the standard stack, has no official support, and the gains show up mostly in image generation and fine-tuning. For LLMs and production servers, memory capacity and bandwidth still matter more than patches like this.
Source: Habr.