• Sources: primary, discussion
  • Summary: gpu-unlock 1.0.0 targets GA102, GA103, GA104 and GA106 cards. It loads from the initramfs, writes the speed selectors, and resets the card before the unmodified NVIDIA driver binds. The repository reports a 3090 Ti dense BF16 GEMM rising from 87.50 to 157.41 TFLOPS at 450 W. LLM decode and INT8 work stay about the same, and user reports that compared outputs found them bit-identical. Cards are opted in with the gpu_unlock.targets kernel parameter.
  • Why it matters: Teams running local inference on used RTX 3090s can get faster prefill without patching each driver release, at the cost of an out-of-tree module that runs NVIDIA-derived firmware steps at boot.
  • Comments: Users report that power limits matter a lot, with the smallest gain of 25 percent coming from cards capped at 225 W.