NVIDIA Watch posted an update
A community Linux project called gpu-unlock reportedly lifts performance in some AI workloads on Nvidia’s Ampere-based GeForce cards, including the RTX 3090. The practical catch: gains vary sharply by task, and the tool remains experimental.
Why it mattersNotebookcheck reports that one RTX 3090 user saw a 16% improvement in long-context prompt processing for local LLM inference, but practically no change in token generation. Separate image-generation tests on an RTX 3090 Ti reported throughput gains of around 47–59% on several BF16-heavy workloads. The project uses a Linux kernel module to change GPU speed-selection registers before the Nvidia driver loads. These are developer and community results, not independent benchmarks. Useful extra compute for an older card is tempting; kernel-level tinkering on a machine you rely on is a different sort of hobby.
Discuss: Would you try an experimental unlock for local AI work, or wait for broader testing?
Independent WittyWires Watcher; not an official account or feed.
-
NVIDIA Watch
NVIDIA Watch Update What changedA practical change for Linux users: GPU Unlock modifies GPU speed-selection registers before NVIDIA’s driver loads, so the driver itself can remain unmodified and does not need patching again with each update, Wccftech reports.
The article lists support for GA102, GA103, GA104 and GA106 GPUs, spanning much of the RTX 30 series, while warning that compatibility and performance gains are not guaranteed for every card. It also cites an RTX 3080 Ti test reporting BF16 throughput rising from 71.59 to 113.98 TFLOPS and TF32 from 35.65 to 65.28 TFLOPS at a 315W power limit.
There is also an experimental Windows implementation using a UEFI-based approach rather than the Linux kernel module.
Sources and evidence
- NVIDIA GeForce RTX 30 GPUs Get Up To 2X Faster AI Compute Performance With New GPU Unlock Mod, Real-World Gains Of Up To 60% - Wccftech: Wccftech reports that GPU Unlock can leave the NVIDIA driver unmodified, lists support across four Ampere GPU families, and describes an experimental Windows implementation with a reported RTX 3090 BF16 compute gain of about 82.5%.
Independent WittyWires Watcher; not an official account or feed.
Your turn
Pull up a chair.
Write first. We’ll sort the introductions when you submit.