Return to Level1Techs.com

Home server for deep learning and plex

So I have an old server I bought and subsequently abused to make into my home plex server, it happens to be alright at deep learning too but slow. I have two Xeon X5670 6 core cpus and 24gb of ram.
I something more powerful and silent as my roommates don’t allow me to keep a server in the living room and sleeping with a turbine in the room doesn’t fly with the girlfriend. You should really see my solution to a silent server, it is some of the highest tier jank.
Powerful, silent, cheap, gpu for deep learning. higher single core performnace or more cores witha gpu, which one trains faster? What has good software support for pytorch and tensorflow? could I use a mining card for cheap gpu compute with deep learning? I need a lot of fast storage for my plex library and for large data sets.
What do you (all) suggest for this build?
Are there any places I can go to find answers to my questions?
I feel like level1 would be the place to have these deep learning configuration benchmarks, unless I missed it and these benchmarks do exist.
thank you

Deep learning is still most accessible in the Nvidia ecosystem. Ubuntu 18.04 with nvidia-docker is an easy platform to get a large variety of software running via containers. AVX instructions are required for some pre-compiled software so you’ll need a Sandy Bridge or newer CPU. You’ll also want a Pascal or newer GPU; an RTX 2060 Super would probably be a reasonable starting point if you can budget for it.

2 Likes

Without Budget, it’s hard to recommend somehting.
Lot’s of fast storage can be anything from Raid 5 2TB Disks, to 24 Disk NVMe Array.

For Plex alone, it would probably be a 3rd gen Ryzen of some sort (CPU or APU), 8-16G of RAM on a B450 Board with a Noctua Cooler. Add storage as needed. Not “cheap”, but affordable.

When it comes to Deep Learning and AI stuff, my latest knowledge is, that Nvidia GPU rains supreme. No clue if CPU or RAM play a major factor. I’d assume you want more ram to potentially hold your entire Dataset in RAM.

You COULD go first gen Threadripper with something like this:
https://pcpartpicker.com/list/rpCP3t
Just add storage as needed. I put in the 2060 Super as that seems to be the best bang for the buck from Nvidia atm. You can save some money by reusing components, and i’d look into ECC RAM for the build. YOu might also want to go with more but slower RAM.

If GPU is all we care about, a Ryzen APU can save you some serious money though.

1 Like

Usually when talking GPU’s these are kind of exclusive properties, either you go powerful, or silent/cheap e.g. nvidia 710/730 or what ever they’re called now a days.

If i were you, id skip worrying about plex, just make a VM that runs on your server, and dedicate and thread or two to the VM, your current CPU(s) are quite power full, or heck even just install the service, and run bare metal.
plex is really easy on your system. Basically if your system can transcode “insert codex here”, then it like Hodd Towards would have said, “it just works”, and is usually done on your GPU, taking up next to no resources.
As for a “powerfull”, silent, and cheap GPU, what is your metric?.
whats your budget?.

1 Like