Comment on Finally happish with my all-in-one homelab (20+ services and way too much tinkering)
tomkatt@lemmy.world 14 hours ago
This is pretty cool, and I’m guessing the hardware doesn’t even break a sweat.
My entire setup runs on two AMD mini-PCs, one running a Ryzen 5700u with Proxmox for VM hosting, the other is a 5500u running bare metal Ubuntu.
Proxmox server hosts:
- Asset UPnP
- Jellyfin
- (formerly) Plex (I ditched it)
- Lyrion
- LRCLIB offline lyrics server (via LRCLIB DB dump)
- Home Assistant VM
- Navidrome instance
Probably a few other things I forgot about…
The bare metal Ubuntu server currently just runs Audiomuse-AI for music analysis, but when not scanning it’s generally idle. Only set this up because early versions of Audiomuse-AI didn’t play nice for me in a VM even when r running with CPU in host mode.
When the hardware gets cheaper I wouldn’t mind a full size server again with a dedicated GPU (I decommissioned my old one due to defective RAM), but for now the mini-PCs have been great. They’re quiet and absolutely sip power, so efficient and quiet.
Yea! It really doesn’t break a sweat. Most of the time it sits around 5-20W (legit jumps between 5W and 20W) while hosting everything, with around 5-15% CPU utilization. Though I do share it with other people, so it may not be as “idle” as I think it is. With all services stopped and nothing running, it obviously uses way less power.
I also have a Windows VM for someone else, which basically acts as a terminal server for development. They run Linux and need Windows for developing something, but they said “I don’t want Windows anywhere on my physical system”, so I gave them a VM to use instead.
I also like using my server as a playground to test things. It saves me having to install VM software on my PCs and keeps experiments separate.
Originally, I used GPU passthrough to a VM, but I stopped doing that because I wanted multiple services to be able to use the GPU while still keeping everything isolated. Unfortunately, my GPU isn’t nice enough for NVIDIA vGPU support.
So passing it directly into containers was the next best option - although definitely the more tedious route to set up. The efficiency gains are pretty nice though.
It also only pulls around 300W at max load (CPU + GPU), but the load is very variable and heavy usage isn’t that dramatic. Normal “heavy” use is closer to around 120W (CPU + GPU).
In the future, I’d love to move to an ARM-based server for even better efficiency, once the hardware becomes more readily available and practical for server use. The efficiency gains are really interesting, but the ecosystem and availability still need some time to catch up.