GPU General Concept questions

[Pedulla]

I have some general concept type questions about GPU passthrough on Proxmox.
I'm currently running pve 6.4, my VM is Mint 20.
The passthrough card in question is a NVidia Quadro RTX6000, running on a Dell 7525 with Epyc processors.

When I do GPU passthrough is it presumed that I will use the passed through GPU for the primary desktop Display for the VM as opposed to the qxl Display?

Does the above impace the accelerated processing tasks like video transcoding in that VM?


If the app running in the VM is display intensive (like say OBS Studio or a video editing app) can using the Spice VGA Display kill the overall VM performance or is this the scenario where use the passed through GPU for desktop display?

If using the passed through GPU for the desktop Display, I assume the Spice/noVNC console no longer work. Is this correct?
If so, what is the best remote desktop to tool to replace spice/novnc from Proxmox?

 

[Dunuin]

If using the passed through GPU for the desktop Display, I assume the Spice/noVNC console no longer work. Is this correct?
If so, what is the best remote desktop to tool to replace spice/novnc from Proxmox?

As soon as you use the passed through GPU your virtual GPU wont work anymore so no VNC/spice build into the PVE webui. But you can always install a remote protocol server of your choice (VNC, RDP, ...) inside your VMs to access it remotely with clients. So VNC would still work, just not over the webUI.

 

[Pedulla]

So VNC would still work, just not over the webUI.

Got it, that makes sense.

As soon as you use the passed through GPU your virtual GPU wont work anymore

Wouldn't I need to set the passed-through gpu as the primary GPU and have no display option set in this case?

 

[Pedulla]

Okay, having tremendous success here.
Am able to pass through the RTX6000 to multiple LXC's each adding it's encode/decode load as needed. Pretty cool.
The pixie dust comes in ensuring you:
- Install the same version of the Nvidea driver on the host as in the container
- Install the Nvidea driver in the container w/o kernel modules
- Get the LXC permissions right in the .conf file for the container

I'm also able to install a desktop GUI in the lxc and then NoMachine for remote access. Order of install is important here.
But the RTX6000 is a headless GPU, so the UI/Console in the LXC is still coming from the integrated graphics chip on the MB (correct me if I'm wrong please).

So now I'm down to my last question:
When I do have to run a GUI desktop environment in one or more of these LXC's would having a second actual GPU/Video card in the host as the primary graphic card, improve the desktop performance in my LXC's?

PS: Just in case anyone is wondering why I'm doing this, one of my applications is to run multiple remote OBS-Studio sessions to live stream in multiple languages simultaneously.