gpu-unlock-ada: Linux kernel module that doubles RTX 4090 FP16 compute speed

A Linux kernel module that applies GPU SM speed select overrides before the NVIDIA driver loads. It increases FP16 and BF16 tensor processing speeds by approximately 2x on the RTX 4090 (AD102), improving performance from 165.6 TF/s to 331.6 TF/s for FP32 accumulation operations. Similar 2x improvement results were confirmed on modified 48GB RTX 4090 cards.

It also provides experimental support for other Ada architecture GPUs, such as the RTX 4070 Ti SUPER (AD103) and RTX 4070 Ti (AD104). Each chipset requires a separate opt-in flag and an exact PCI address, and on some models, the FP16/FP32 issue rate increased from 95.0 TF/s to 189.9 TF/s under a 240W power limit. Laptop GPUs require per-model verification due to differences in VBIOS and power rails.

Unlike existing methods, it does not require patching the NVIDIA driver or modifying UEFI applications. It executes signed firmware, verifies a completion marker, and then performs a PCIe Function Level Reset (FLR) before the standard NVIDIA driver binds. This process has been confirmed to remain stable during driver initialization and CUDA benchmarks.

INT8, dp4a, and scalar FP32 operations remain at baseline levels, and no performance benefits were observed in quantized mixture-of-experts workloads, which do not use the FP32 accumulation tensor path. Therefore, this module is effective in environments where FP16/BF16-based high-resolution computations are the primary workload. It is distributed under the GPL-2.0 license, and the firmware is generated locally.