>>936325855

I've been thinking about building an "everything" server that would fit into one compact-ish package (something like a Jonsbo N3 or N4)
By everything I mean I want it to have two "parts".
One is something like Truenas Scale, running Jellyfin, the various *narrs, nextcloud, MAYBE the occasional game server, mail server, typical homelab shit.
The other is essentially a machine I can SSH to to run machine learning using whatever Nvidia GPU I could get my hands on.

What's the most "humane" way to use two virtual machines like that with one GPU?
(not necessarily, literally two VMs, any other solution is welcome)
By that I mean that while I like having a server, I don't like maintaining it and tinkering with it beyond the initial setup, so I'm looking for a solution with minimal headache.

My initial thought was to use something like an Arc pro A40 for Jellyfin encoding and the other nvidia gpu for ML tasks, but considering size, power and money limitations it would make more sense to use just the one GPU for everything, especially since encoding videos isn't very demanding for even 3-4 concurrent viewers.
So that's option 2, but from what I've read about TrueNAS Scale if I want to pass a GPU to a VM it would require another GPU to be used by the bare-metal system.
So I guess Proxmox with two virtual machines?

I'm certain I'm missing something simple, which is why I'm asking.