Zagitta/gpu-unlock
UNLOCKING GEFORCE TENSOR CORES
Linux kernel module that removes NVIDIA's artificial limit on GeForce RTX 20/30/40 tensor core speed. Doubles FP16/BF16/TF32 throughput without VBIOS flashing or driver patches.
// KEY FEATURES
- Doubles FP16/BF16/TF32 tensor throughput
- Works on RTX 20/30/40 series via kernel module at boot
- Gives +25–55% speedup for LLM prefill and image generation
- No driver patches, VBIOS flashing, or reinstall on driver updates
// DEEP DIVE: HOW IT WORKS
gpu-unlock loads as a kernel module before the proprietary NVIDIA driver initializes and programmatically removes the limit NVIDIA imposes on consumer GeForce: tensor cores run at half speed for FP16/BF16/TF32 with FP32 accumulation. The unlock is applied once at system start and persists until reboot.
Unlike patched drivers or VBIOS flashing, the solution does not touch NVIDIA files and requires no repeat actions after driver updates. Supported architectures: Turing (RTX 20), Ampere (RTX 30 — stable), and Ada Lovelace (RTX 40 — experimental). Exact PCI IDs are listed in kernel/gpu_unlock/recipes.c.
Linux only; Windows not supported. On laptops the unlock resets if the discrete GPU is fully powered down — the NVreg_DynamicPowerManagement=0x00 parameter helps but results are not guaranteed. No gain on memory-bound workloads (LLM decode), quantized models (INT4/INT8/FP8), gaming, or FP32 CUDA. RTX 50 series not supported.
// USE CASES
Best for: For ML engineers and enthusiasts running LLMs and generation on GeForce under Linux
- Accelerate LLM prefill in vLLM or Transformers by 25–55%
- Reduce LoRA and full fine-tuning time in BF16 by 25–50%
- Boost SDXL and DiT image generation speed in ComfyUI by 30–55%
// QUICK START
git clone https://github.com/Zagitta/gpu-unlock
cd gpu-unlock
make installCommands are taken from the repository README — check the current version on GitHub before running.
New repositories every 30 minutes
REDDYX AI scans GitHub 24/7 and ships the best AI/ML/Web3 projects to Telegram.
Join on Telegram