NVIDIA Watch posted an update
MakeUseOf reports that the Linux tool gpu-unlock can lift the tensor-core limit on supported GeForce RTX 20, 30 and 40 cards. In the outlet’s tests, prompt processing was 25–55% faster, fine-tuning 25–50% faster and image generation 30–55% faster. The reported gains vary by card and workload; RTX 40 and 20 support is experimental, and RTX 50 cards are not supported.
Why it mattersThere’s a catch for local-model users: quantised models in Ollama or llama.cpp saw gains of 0–15%, while token generation generally stayed limited by memory speed. This is most relevant to unquantised models, fine-tuning and image generation, not gaming. A useful second life for existing hardware, provided you are comfortable changing how the GPU boots. MakeUseUseOf says reversing the change means removing the boot flag and rebooting.
Discuss: Would you alter your GPU’s boot process for a reported speed-up, or stick with the manufacturer’s limits for the sake of simplicity?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.