Multi-GPU support?

[gts250gamer]

Recently, I decided to switch to an Intel Arc A310 as my Jellyfin transcoding card, in an attempt to ditch all of my aging Nvidia hardware (Tesla P4). I have two low profile A310 Eco cards from Sparkle, and I have currently been using one of these cards solo as my transcoding card.

It has largely worked great, and the power consumption has decreased dramatically from the old server, but I have also noticed that the A310 struggles a bit with multiple HDR streams simultaneously, functionally impairing SyncPlay for a group larger than two people.

I have done some initial searching online to try to see if JF supports multiple GPUs simultaneously for transcoding, but it appears as though it was not supported, at least as of the latest posts I was coming across from around a year ago. I am simply curious as to whether this has changed at all recent times, or if it is still under development.

Thanks!

[TheDreadPirate]

It has not changed. ffmpeg cannot dynamically decide which GPU to use, meaning Jellyfin would have to coordinate that and it currently doesn't and there are currently no plans to implement something like that.

VRAM is extremely important for tone mapping, so going with the A310 over the A380 or higher limits your tone mapping potential. I think the math is ~1GB of VRAM per tone mapped stream.

Make sure you enabled low power encoding. This setting is important when using OpenCL (which is used during tone mapping).

https://jellyfin.or/docs/general/adminis...r-encoding

If your motherboard supports resizeable BAR, enable it. It sounds like you have an older board so that may not be an option.

If your server is on Linux you can also experiment with updating your Arc GPU's firmware. Another user reported issues with newer firmware on their pretty old motherboard. They were able to use another newer system to flash back to the factory firmware to get it working again. So attempt at your own risk. Flash one, reboot and verify it still works before flashing the other.

https://forum.jellyfin.org/t-flash-intel...e-in-linux

[gts250gamer]

It has not changed.  ffmpeg cannot dynamically decide which GPU to use, meaning Jellyfin would have to coordinate that and it currently doesn't and there are currently no plans to implement something like that.

VRAM is extremely important for tone mapping, so going with the A310 over the A380 or higher limits your tone mapping potential.  I think the math is ~1GB of VRAM per tone mapped stream.

Make sure you enabled low power encoding.  This setting is important when using OpenCL (which is used during tone mapping).

https://jellyfin.or/docs/general/adminis...r-encoding

If your motherboard supports resizeable BAR, enable it.  It sounds like you have an older board so that may not be an option.

If your server is on Linux you can also experiment with updating your Arc GPU's firmware.  Another user reported issues with newer firmware on their pretty old motherboard.  They were able to use another newer system to flash back to the factory firmware to get it working again.  So attempt at your own risk.  Flash one, reboot and verify it still works before flashing the other.

https://forum.jellyfin.org/t-flash-intel...e-in-linux

This is a perfect explanation. Thank you, DreadPirate!