2w ago

need help understanding if this setup is even feasible.

I have an unused dell optiplex 7010 i wanted to use as a base for an interference rig.

My idea was to get a 3060, a pci riser and 500w power supply just for the gpu. Mechanically speaking i had the idea of making a backpack of sorts on the side panel, to fit both the gpu and the extra power supply since unfortunately it's an sff machine.

What's making me weary of going through is the specs of the 7010 itself: it's a ddr3 system with a 3rd gen i7-3770. I have the feeling that as soon as it ends up offloading some of the model into system ram is going to slow down to a crawl. (Using koboldcpp, if that matters.)

Do you think it's even worth going through?

Edit: i may have found a thinkcenter that uses ddr4 and that i can buy if i manage to sell the 7010. Though i still don't know if it will be good enough.

20 comments

I just got second PSU just for powering multiple cards on a single bifurcated pcie for a homelab type thing. A snag I hit that you might be able to learn from: PSUs need to b turned on by the motherboard befor e being able to power a GPU. You need a 15$ electrical relay board that sends power from the motherboard to the second PSU or it won't work.
Its gonna be slow as molasses partially offloaded onto regular ram no matter what its not like ddr4 vs ddr3 is that much different speed wise. It might maybe be a 10-15% increase if that. If your partially offloading and not doing some weird optimized MOE type of offloading expect 1-5token per second (really more like 2-3).
If youre doing real inferencing work and need speed then vram is king. you want to fit it all within the GPU. How much vram is the 3060 youre looking at?
- You need a 15$ electrical relay board that sends power from the motherboard to the second PSU or it won't work.
  If you are talking about something like the add2psu boards that jump the PS_ON line of the secondary power supply on when the 12v line of the primary one is ready. Then i'm already on it the diy way. Thanks for the heads up though :-).
  expect 1-5token per second (really more like 2-3).
  5 tokens per seconds would be wonderful compared to what i'm using right now, since it averages at ~ 1,5 tok/s with 13B models. (Koboldcpp through vulkan on a steam deck) My main concerns for upgrading are bigger context/models + trying to speed up prompt processing. But i feel like the last one will also be handicapped by offloading to RAM.
  How much vram is the 3060 youre looking at?
  I'm looking for the 12GB version. i'm also giving myself space to add another one (most likely through a 1x mining riser) if i manage to save up enough another card in the future to bump it up to 24 gb with parallel processing, though i doubt i'll manage.
  Sorry for the wall of text, and thanks for the help.
  
  No worries :) a model fully loaded onto 12gb vram on a 3060 will give you a huge boost around 15-30tps depending on the bandwidth throughput and tensor cores of the 3060. Its really such a big difference once you get a properly fit quantized model your happy with you probably won't be thinking of offloading ram again if you just want llm inferencing. Check to make sure your motherboard supports pcie burfurcation before you make any multigpu plans I got super lucky with my motherboard allowing 4x4x4x4 bifuraction for 4 GPUs potentially but I could have been screwed easy if it didnt.
I think a dual-channel system with DDR3-1600 isn't what we call fast any more. So you should try to avoid offloading with that. But I'm not an expert on the numbers, and it depends a bit on the specific use-case whether it makes sense to invest in old hardware, or buy a new machine along with a graphics card, since that's quite some money.
Would you be willing to talk about what you're intending to do with this at all? No hard feelings if you'd rather not for any reason.
For context on my request - I've been following this comm for a bit and there seems like a real committed, knowledgeable base of folks here - the dialog just in this post almost brings a tear to my eye, lol.
I work fairly adjacent to this stuff, and have a slowly growing home lab. Time is limited of course and gotta prioritize what to learn and play with - LLMs are obviously both and useful, but I haven't yet encountered a compelling use case for myself (or maybe just enough curiosity about one) to actually dive in.
Selfishly I just wish every post here would give some info about what they're up to so I can start to fill in whatever is apparently missing in my sort of "drum up fun ideas" brain subroutine, regarding this topic. Lol.
- at the moment i'm essentially lab ratting the models, i just love to see how far i can push them, both in parameters and in compexity of request. before they break down. plus it was a good excuse to expand my little "homelab" (read: workbench that's also stuffed with old computers) form just a raspberry pi to something more beefy. as for more "practical" (still mostly to mess around) purposes. i was thinking about making a pseudo-realistic digital radio w/ announcer, using a small model and a TTS model: that is, writing a small summary for the songs in my playlists (or maybe letting the model itself do it, if i manage to give it search capabilites), and letting them shuffle, using the LLM+TTS combo to fake an announcer introducing the songs. i'm quite sure there was already a similar project floating around on github. another option would be implementing it in home assistant via something like willow as a frontend. to have something closer to commercial assistants like alexa, but fully controlled by the user.
  I’ve been following this comm for a bit and there seems like a real committed, knowledgeable base of folks here - the dialog just in this post almost brings a tear to my eye, lol.
  to be honest, this post might have been the most positive interaction i've had on the web since the bbs days. i guess the fact the communities are smaller makes it easier to cobble up people that are genuinely interested in sharing and learing about this stuff, same with the homelab community. like comparing a local coffee shop to a starbucks, it just by nature filters for different people :-)
I have the feeling that as soon as it ends up offloading some of the model into system ram is going to slow down to a crawl.
Then don't offload! Since its 3000 series, you can run an exl3 with a really tight quant.
For instance, Mistral 24B will fit in 12GB with no offloading at 3bpw, somewhere in the quality ballpark of an Q4 GGUF: https://cdn-uploads.huggingface.co/production/uploads/6383dc174c48969dcf1b4fce/tfIK6GfNdH1830vwfX6o7.png
It's especially good for long context, since exllama's KV cache quantization is so good.
You can still use kobold.cpp, but you'll have to host it via an external endpoint like TabbyAPI. Or you can use croco.cpp (a fork of kobold.cpp) with your own ik_llama.cpp trellis-quantized GGUF (though you'll have to make that yourself since they aren't common... it's complicated, heh).
Point being that simply having an ampere (3000 series RTX) card can increase efficiency massively over a baseline GGUF.
- I'll have to check exllama once i build the system, if it can fit a 24B model in 12 gb. It should give me some leeway for 13B ones. Though i feel like i'll need to quantize to exl3 myself for the models i use. Worth a try on a Container though.
  Thanks for the tip.
  
  You can definitely quantize exl3s yourself; the process is vram light (albeit time intense).
  What 13B are you using? FYI the old Llama2 13B models don’t use GQA, so even their relatively short 4096 context takes up a lot of vram. Newer 12Bs and 14Bs are much more efficient (and much smarter TBH).

20 comments