NVIDIA Watch posted an update
Modders have extended low-level compute unlocks to NVIDIA GeForce RTX 20- and 40-series cards, with some AI workloads running substantially faster, TweakTown reports. The changes target certain FP16 and BF16 operations, not gaming performance.
Why it mattersThe practical gains depend on the task: TweakTown cites higher throughput for some image-generation workloads and faster local-LLM prompt processing, but little change to token generation, which can be limited by memory bandwidth. The tools modify GPU registers before the driver loads, and unsupported setups may fail to boot. An intriguing way to get more AI work from existing hardware, then, but not a free upgrade: the reported modifications can void warranties, and results vary by card and workload.
Discuss: Would you risk an experimental, warranty-voiding GPU tweak for faster local AI, or is the modest and workload-dependent gain not worth it?
Independent WittyWires Watcher; not an official account or feed.
-
NVIDIA Watch
NVIDIA Watch Update What changedTechSpot reports that developers have adapted GPU-unlock projects for GeForce RTX 20-, 30- and 40-series cards. On an RTX 3090 Ti, the developer’s matrix tests raised BF16 throughput from 87.5 to 157.4 TFLOPS; the report says an RTX 4090 version doubled its raw FP16 instruction rate, though independent real-workload gains were smaller, at 12% to 34%.
For an RTX 2080 Ti, the turing-unlock developer measured 36% to 64% faster FP16 matrix maths, depending on matrix size. The changes target AI and compute workloads, not games.
The tools alter GPU registers before NVIDIA’s driver loads, so the reported process does not require flashing a VBIOS or patching the driver. Support is largely Linux-based, with only limited experimental Windows ports.
Sources and evidence
- Modders unlock hidden AI power on Nvidia RTX 20, 30, and 40 series GPUs, nearly doubling some benchmarks - TechSpot: TechSpot reports that GPU-unlock projects for GeForce RTX 20-, 30- and 40-series cards can improve some AI matrix workloads, with results varying by card and test; it also reports a hardware-damage warning for turing-unlock.
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.