I’m trying to build system with two T1000 video cards, total 8 outputs in order to play 8 4K video files - one at each output. There were problems with configuring 8 separate screens and I tried to discuss it in another topic, but meantime I was able to build configuration with two screens each with four monitors.
Now I’m trying to play video file with mpv player.
If I start player for screen :0.0 it shows C+G process type for that player in nvidia-smi, but when I start second player for screen :0.1 it have spread processes across two GPU and it looks like NVDEC of only one (first) GPU is used here:
Choose the GPU device used for decoding when using the cuda or nvdec hwdecs with the OpenGL GPU backend, and with the cuda-copy or nvdec-copy hwdecs in all cases.
For the OpenGL GPU backend, the default device used for decoding is the one being used to provide gpu output (and in the vast majority of cases, only one GPU will be present).
For the copy hwdecs, the default device will be the first device enumerated by the CUDA libraries - however that is done.
For the Vulkan GPU backend, decoding must always happen on the display device, and this option has no effect.
So maybe you can choose the gpu per command-line for each mpv instance.
So far tried vlc player and found that it use NVDEC of corresponding video card depending on which screen playback is started. So GPU memory usage and NVDEC engine usage is distributed evenly for both channels!