Why every running CUDA programs force the GPU to P2 state. It’s really annoying:
- There is no way to disable that behavior. (no NVreg_* options in nvidia kernel module).
- It prevents power saving when decoding with NVDEC (GPU clock doesn’t decrease bellow base clock).
- It reduces GPU performance of CUDA programs and games during encoding sessions.
There are topics from 2017/2018 complaining about that:
A least add an option such as kernel parameters to disable “P2 forced” state when running CUDA.
It is 2024 - and this issue make Nvidia GPU video acceleration pointless in Linux.