Are you self-hosting LLMs (AI models) on your headless servers? I’d like to hear about your hardware setup. What server do you have your GPUs in?
When I do a hardware refresh I’d like to ensure my next server can support GPU(s?) for local LLM inferencing. I figured I could put in either a 4090 or x2 3090’s(?) maybe into an R730. But I’ve only barely started to research this. Maybe it isn’t practical.
I don’t know much other hardware lineups besides the Dell R7xx lineup.
I host oobagooba on an R710 as a model server API, and host sillytavern and stable diffusion which use oobagooba as clients. I use an R710 using a CPU, so as you can imagine inferencing is so slow it’s basically unusable. But I wired it up as a proof of concept.
I’m curious what other people who self-host LLMs do. I’m aware of remote options like Mancer or Runpod. I’d like the option for purely local inferencing.
Thanks all
You must log in or register to comment.
At least according to this fairly detailed eBay listing you might be limited in what GPUs you can run in an R730. It states a 300w max per card and double width, which would eliminate both the power and physical requirements of the 3090/4090. You could run say some Tesla P40s but they would be a bit slower.
Another option would to be just buy a rack mount 4U case and say a X99 motherboard (same era CPU as the R730) which would give you a bit more flexibility in running 3090/4090 cards so long as you had a 1200w or so PSU.
Yeah running a 4U case and assembling it with “plain desktop” hardware but rack mounted and headless is definitely an option too. I might be asking too much of server hardware to take R730s (or any racked datacenter hardware) and fit them to a role they weren’t designed for. These are good thoughts and useful links, thank you.
One challenge with the 4090 specifically is I don’t believe there are any dual-slot variants out there, even my 4080 is advertised as a triple-slot card (and actually takes four because Zotac did something really, really annoying with the fan mounting)…you could liquid-cool and swap the brackets, but then you have the unenviable task of mounting sufficient radiators and support equipment (pump, res, etc) into a rackmount server. That assumes you’re looking at something 2-3U, since you mentioned an R730; if you’re willing to do a whitebox 4U build it’s a lot more doable.
Of course if money is no object, ditch plans for the GeForce cards and get the sort of hardware that’s made to live in 2U/3U boxes, i.e. current-gen Tesla (or Quadro, if you want display outputs for whatever reason). If money is an object, get last-gen Teslas. Tossed an old Tesla P100 (Pascal/10-series) into my Proxmox server to replace a 2060S with half the VRAM, for LLMs I didn’t really notice an obvious performance decrease (i.e. still inferences faster than I can read), and in a rack server you won’t even have to mess with custom shrouds for cooling, since the fans in the server are going to provide more than enough directed airflow.