Modders have discovered a way to unlock hidden AI performance on NVIDIA GeForce RTX 20, 30 and 40 series graphics cards. The open-source GPU Unlock project, created by developer Zagitta, removes software restrictions on Tensor Core instruction issue rates without needing a VBIOS flash or hardware modifications.
What happened
Consumer GeForce cards have artificial limits on Tensor Core operations with FP32 accumulation compared to their professional equivalents. The Linux kernel module of GPU Unlock changes configuration registers before the NVIDIA driver loads, unlocking up to double the throughput in precisions like BF16 and FP16.
Reported results include: RTX 3090 Ti rising from 87.5 to 157.4 TFLOPS in BF16; RTX 4090 doubling from 165.6 to 331.6 TFLOPS in FP16; and RTX 2080 Ti reaching 111 TFLOPS. Real gains in AI workloads like LLM prefill range from 14% to 36%, and in image generation in ComfyUI up to 60%. More info on VideoCardz.
Limitations and cautions
The unlock is temporary (clears on power off) and currently works mainly on Linux. It does not improve gaming performance, as it focuses on AI compute. It may void warranty and requires administrator privileges. Sources: VideoCardz, TechPowerUp and Zagitta’s GitHub repository.
Why it matters
This gives new life to older GPUs for those running local AI models, lowering the barrier to entry without buying new hardware. The modding community continues to expand support.
Image credit: Reproduction / VideoCardz
By GeekikiBot
