RTX 3090 family gets experimental Tensor Core unlock
An experimental Linux kernel module claims major gains in selected RTX 3090 Tensor Core workloads, but the results are not a general gaming upgrade.
Nvidia’s GeForce RTX 3090 family has become the focus of an experimental open-source project that claims to unlock more Tensor Core throughput on compatible Linux systems. The project, gpu-unlock, changes the instruction-issue rate for FP16 and BF16 operations with FP32 accumulation; it does not require a modified Nvidia driver or a permanently flashed VBIOS, according to its developer. 12
What the project changes
The project loads a Linux kernel module early in the boot process, before the Nvidia driver binds to the card. It changes speed-selection registers, performs a function-level reset and then hands the GPU to the unmodified driver. The repository says the technique targets Nvidia’s RTX 30-series cards, including GA102, GA103, GA104 and GA106 GPUs. 2
The headline numbers are substantial, but they are not a universal gaming upgrade. A community report cited by Notebookcheck recorded an RTX 3090 reaching 129.6 TFLOPS in BF16 matrix multiplication, up from 77.5 TFLOPS, while TF32 performance rose from 40.2 to 69.8 TFLOPS. A separate RTX 3090 Ti test in the project’s repository recorded BF16 performance rising from 87.50 to 157.41 TFLOPS. 12
That distinction matters. Other testing reported only a 16 per cent improvement in long-context local-LLM processing, with token-generation speeds barely changing. The project is also limited primarily to Linux, while its Windows support remains experimental. There is no independent evidence that the unlock improves ordinary game frame rates, so the results should be treated as workload-specific experiments rather than a free performance upgrade for every RTX 3090 owner. 12
Nvidia’s official product page confirms that the RTX 3090 and 3090 Ti are Ampere cards with 24GB of GDDR6X memory and third-generation Tensor Cores, the hardware feature most relevant to the project’s claims. Nvidia’s page is used here for product context, not as confirmation of the community benchmark results. 3
Our opinion
This is an intriguing piece of hardware archaeology, especially for owners using an RTX 3090 as a local AI or rendering accelerator. It is not a reason to alter a primary machine casually: the project changes low-level boot behaviour, remains experimental and can affect system stability. The sensible takeaway is that ageing Ampere hardware may have more specialised compute headroom than Nvidia’s consumer segmentation exposed, but the burden of testing and the risk of breaking a working Linux installation remain with the user.