XNews

gpu-unlock Open-Source Project Unlocks Full Tensor Core Speed on GeForce RTX 20/30/40 Cards

gpu-unlock Open-Source Project Unlocks Full Tensor Core Speed on GeForce RTX 20/30/40 Cards

The open-source gpu-unlock project removes NVIDIA's tensor core cap on GeForce RTX 20/30/40 series cards at boot, before the official driver loads, delivering up to 2× FP16/BF16/TF32 tensor throughput. An RTX 3060's matrix compute rises from 27 to 51 TFLOPS, with ComfyUI image generation 30–55% faster and LoRA fine-tuning 25–50% faster, without firmware flashing or driver replacement.

Event version 1.

Key insights

Tensor cores capped at half speed

NVIDIA ships GeForce cards with tensor cores capped to half speed for FP16, BF16 and TF32 math with FP32 accumulation. The same silicon sold as an RTX A-series or professional card runs at full rate.

Unlock occurs before driver load

gpu-unlock removes the cap at boot, before the stock NVIDIA driver loads. No patched driver or VBIOS flash is required, and nothing needs to be redone when NVIDIA releases a new driver.

RTX 3060 compute nearly doubles

An RTX 3060's matrix compute rises from 27 to 51 TFLOPS. Results are bit-identical before and after unlocking, only faster.

Real-task speedups reported

ComfyUI image generation runs 30% to 55% faster, LoRA fine-tuning 25% to 50% faster, and LLM prefill improves 25–55%.

Supported GPU generations

Ampere RTX 30 series support is stable with 100+ user reports, covering RTX 3050 through 3090 Ti plus laptop GPUs. Ada Lovelace RTX 40 series support is experimental, covering RTX 4060 through 4090 D plus laptop GPUs.

Source posts