I’d like to buy an ex company PC to serve as home server, as an upgrade for my raspberry pi. It should run some selfhosted services like Frigate and Immich, and perhaps Jellyfin. I was thinking to buy an 9th gen Intel or later ex company PC. It should be energy efficient as well.
I tried looking at Trademe but you can’t filter on CPU generation. I could search on the exact Intel cpu number but that seems too restrictive. Pbtech is too expensive what I saw. I haven’t tried Marketplace but guess that’s the same. Some reddit post suggested Dell Optiplex, but no idea how their type numbering works.
Any tips?
I’ve got a few different boxes at home. From an old desktop that I can fit a fanless nVidia GPU in (for Frigate) which i’m running unRaid on as its got lots of slots for HDD. But I kinda hate unRaid, so have since gone the TinyMiniMicro route (ala ServeTheHome - https://www.youtube.com/watch?v=bx4_QCX_khU&list=PLC53fzn9608B-MT5KvuuHct5MiUDO8IF4) and have 3 EliteDesk Minis.
They’re running XCP-NG with a Xen Orchestra VM built from source to manage the 3 boxes. On that i’ve got a handful of VMs, a couple of generic linux builds one for docker, and then 3 VMs running a k3s cluster and i’m slowly migrating services across to. Some stuff like Plex / Frigate etc does better with a GPU though so i’ll keep the unRaid box around running those sorts of things there.
Anywho, with the small form factor PCs - Trade Me’s auction history is terrible so even though I only bought them part way through last year I can’t find the vendor I bought mine from. They were really good which is why I wanted to shout them out - sent a replacement fan & stick of RAM within a day of me emailing a problem with what they’d originally sent. They show up on TradeMe & Facebook Marketplace all the time, they’re all ex-lease from corps I’m guessing.
I’d recommend these boxes over a NUC because they’re cheaper, and by far over a Pi because they are so much more powerful for a similar price if you can get a good deal on a used one. They’re not much louder than a Pi (and because they don’t use those tiny fans the noise they do make is less annoying) and they’re still reasonably lower power use from the wall too.
Thanks for your detailed response! Think just an ex lease pc is sufficient for now, but will keep the ServeTheHome in mind.
That’s all these are, they’re just small - not sure which corps would use them, maybe contact centres as they can probably be mounted behind monitors. My 3 are the G2 so have an i5-6500T @ 2.5Ghz. They’ve got 2 RAM slots so I have 2x8GB sticks in mine (they almost always just come with 1x8GB, so is a nice easy upgrade). The only trick is to check what size HDD they’re going to come with, but even though they’re small they’re quite serviceable - fans, hdd & ram all easy to access.
There’s loads of them on TradeMe at the moment; and these are just the HP EliteDesk Mini’s that ServeTheHome video goes through a few different types, I think Dell & Lenovo both make very similar sized & spec machines:
https://www.trademe.co.nz/a/marketplace/computers/desktops/no-monitor/search?search_string=elitedesk mini&sort_order=priceasc
Hmm, now im searching these boxes im wondering if I want to expand my cluster from 3-6 nodes, though I don’t know if I have enough room in my rack :)
Here’s the Dell & Lenovo varieties:
https://www.trademe.co.nz/a/marketplace/computers/desktops/no-monitor/search?search_string=optiplex micro&sort_order=priceasc
https://www.trademe.co.nz/a/marketplace/computers/desktops/no-monitor/search?search_string=thinkcentre tiny&sort_order=priceasc
Thanks! So I’m new to multi nodes self hosting stuff, so far have used a Rpi for everything. With that configuration, can you run everything? Most of the things I want to run are dockerized, but eg pi hole had to be installed directly. Most important is better performance for Frigate (web cam monitoring) and immich (photo and video management).
There’d be a learning curve if you wanted to make a proper cluster; but you don’t really have to. The simplest way to augment what you’re already doing with the Rpi is to just add another computer as a server and put some of the docker containers on there instead of on the pi. You’d just have to know what IP is for which service, or just setup local DNS via pi-hole and point it to the right IP.
Yeah that’s probably the easiest, just set it up as second server. Thanks again.
Hi sorry to bother you, but I’ve been researching getting a HP elitedesk 800, probably g3. For Frigate they recommend a Coral. Mini PCIE is cheaper than usb, and with a mini pcie to pcie adapter I should be able to fit it in one of the pcie slots. I’m also planning to add a GPU. But perhaps I don’t need the coral if I have a GPU, how’s your experience with using a GPU? Would make it easier as it’s very hard to find coral devices.
Yeah the difficulty of sourcing corals is exactly why i’ve not yet used one. I’ve used both GPU & CPU for detection and both have been fine. I switched to CPU because the GPU I had in this box originally was a Quadro and its power use and heat output was just rediculous. I’ve since bought a passive cooled GT1030 but I might not have even switched back to GPU over CPU since I installed - I only have 4 cameras in Frigate and the CPU is keeping up fine.
I figured i’d have a crack at enabling the GPU detector today and noticed that i’m currently not using CPU detectors anymore, i’ve at some point switched to using the OpenVINO one so its using the Intel iGPU I think. That must have been working well given I forgot that’s what I was using :)
I’ll let you know how I get on switching to the nVidia detector.
Well the nVidia detector is a bit more fiddly to get going than the OpenVINO one; but not too hard, just a matter of going through the steps in the documentation.
You need to run a different image of the frigate container so it includes all the TensorRT stuff and separately you have to generate the models. The only hiccup I had there was setting the right GPU to generate on bc mine is only a 1030 disabling FP16 operations.
That’s all running fine so now i’ll just monitor it for a bit to see if it makes any significant difference to CPU load & case temperatures running the detections on the GPU instead of CPU.