GPU-unlock: FP/BF16 instruction issue rate on Nvidia 3000's cards

GPU-unlock is a Linux kernel module that raises the FP/BF16 instruction issue rate of Nvidia GeForce RTX 3000 cards based on GA102, GA103, GA104, and GA106 GPUs to levels associated with professional models. It combines an existing CMP-unlocker technique with early execution during boot: the module sets hardware speed selectors and resets each eligible GPU before the standard Nvidia driver binds. Because it does not patch the Nvidia driver, users do not need to repeat driver-specific patching after updates. The change doubles FP/BF16 FMA issue speed into FP32 operations in supported cases. In a reported RTX 3090 Ti test, dense GEMM throughput rose from 87.5 to 157.4 TFLOPS for BF16 and from 42.6 to 82.6 TFLOPS for TF32.

User reports show successful unlocks on several RTX 3060, 3070, 3080, and 3090 models, with BF16 GEMM gains varying roughly from 25% to 84%; one RTX 3080 Ti variant completed the procedure but remained at half rate, and some models lack reports. Gains depend on the card’s power limit and workload: matrix-heavy tasks such as LLM prefill, diffusion, and FlashAttention can improve, while memory-bound LLM decoding and INT8 work generally do not. Reported output comparisons were bit-identical. Installation uses Linux packages and an initramfs module, with unlocking explicitly enabled through a kernel command-line option and optionally limited to selected PCI addresses. The project documents boot, resume, and compatibility behavior; hibernation is not covered.