Nvidia RTX GPUs see AI performance boosts via open-source tool

An open-source Linux tool named “gpu-unlock,” developed by GitHub user Zagitta, allows Nvidia GeForce RTX 20, 30, and 40 series graphics cards to bypass artificial instruction-rate limits on Tensor Cores. By exploiting a firmware vulnerability in the SEC2 Booter, the utility modifies low-level GPU speed-selection registers during the system boot process before official proprietary drivers initialize.

This modification enables higher AI math throughput for FP16 and BF16 calculations without requiring VBIOS modifications or patched drivers. Reported performance gains are significant: the RTX 4090 can see nearly double the theoretical Tensor speed, while the RTX 3060 has shown improvements between 78% and 99%. These boosts are particularly effective for local Large Language Models (LLMs) and AI image generation.

While the tool provides a substantial performance lift for AI workloads, it does not impact gaming performance or token generation speeds significantly. The process requires deep system intervention via the Linux initramfs, making it a specialized solution rather than a tool for general consumers. The ability to unlock this performance suggests that consumer GeForce hardware may have had latent AI capabilities restricted to differentiate them from more expensive professional-grade GPUs.