Kernel module doubles FP/BF16 speed on Nvidia GeForce cards

A new kernel module has been released that modifies the behavior of Nvidia GeForce graphics cards. The module runs an unlock sequence, writes specific speed selectors, and performs a function level reset before the standard NVIDIA driver binds to the hardware. This process doubles the instruction issuing speed for FP/BF16 FMA into FP32 operations on GA102, GA103, GA104, and GA106 chips. The module accepts every GeForce ID listed in its source code and unlocks cards concurrently in about one second.

User reports indicate significant performance improvements in matrix-heavy workloads. Dense GEMM throughput increased by 25% on cards capped at 225 W, while other configurations showed gains up to 36%. Specific benchmarks included a 3090 Ti dense 8192³ GEMM test at 450 W with FP32 accumulation. LLM prefill, diffusion, and FlashAttention tasks also saw speedups between 14% and 36%. However, LLM decode and INT8 workloads remained approximately the same, as these are limited by memory bandwidth rather than instruction issue rate.

Reports confirm that computational outputs are bit-identical before and after the unlock. The module supports up to four cards in a single machine, including display cards and those passed through to virtual machines using vfio-pci. Loading the module manually after boot does not affect cards already bound by the nvidia driver. The kernel module and local packaging are licensed under GPL-2.0-only.