It's actually not that straightforward to plug in these consumer cards as 4x setup. We spent weeks researching how to achieve up to 7x RTX 3090 setup in a single rig. Could write up our method if anyone is interested.
It's not even just about the slots, it's about the PCIe lanes (which is something I never had to worry until now, though I built countless PCs in the past).
We tried bunch of setups with Threadrippers and EPYC, at the end settled for the ROMED8-2T which is a monster motherboard.
We run 4x 2080s on threadripper systems. What sort of trouble did you run into? I thought threadripper has plenty of PCIe lanes. We didn't have any trouble but it could be I missed something, we had to get it working quick and I didn't do very much benchmarking.
Threadrippers are great and I had 4x Threadripper setup for the longest time, but they are a bit more expensive.
The advantage of EPYC is that because it's so common, we can find used cheaper ones on ebay. They are a bit slower I believe, but we can deal with that by using Nvidia's DALI and decoding images on the GPU rather than CPU.
Ohh I hadn't thought of there being cheaper ones on ebay. That's a good tip, I'll check it out for our next upgrade.
We're doing photogrammetry not machine learning, running some blackbox software that scaled best with clock speed so threadrippers were the most efficient option.
I put them in 4U supermicro boxes with a noctua cooler with 2 9000RPM delta fans attached to it with rip ties.
I just built a rig with a Romed8-2T as well. I got pcie 3 16x risers and zip-tied them above the tower into a rack shelf above it. It's super ghetto, and I can't believe it works, but it totally does. I'm hosting on vast.ai hoping someone will train with my 4-6 3090s, but everyone wants their large language models and image generation models that require more than the 24GB of RAM. shrug
Maybe some day I'll use it myself to train my plastic surgery outcome estimation visualization GAN or diffusion model if I can figure out how to fine-tune one.
Never thought I'd get a response from Jeremy Howard, now I have to post a well polished article! Thanks for all your teachings btw, I really enjoyed and learned a lot from the fast.ai ML course!
would love to see the riser setup that you're using for such a monster!
we mostly gave up and just got barebones machines since the cabling situation becomes pretty tricky, and the barebones total cost is low relative to the GPUs anyways.
I posted a link just below to twitter with an image of the riser setup. That setup worked well for 4x, but for the 7x we're moving the cards upside down and setting them up like tree branches if you will. So the trunk/floor is the motherboard and you get close to edges, the cards are angled and use longer riser cables.
The issue we had with barebones was cost and cooling, we use 30$ racks from Target and hang the GPUs with metal zip ties and a box fan from below, so they get lots of air and we don't break the bank and can easily roll them around.
Sure, will do, though it might take some time to finish writing the blog post, you can get a preview of our previous setup with 4x GPUs here: https://twitter.com/ftufek/status/1569367127878139905. For those that are curious, that's running a Threadripper 3970x.
It's not exactly a "clean" one, like a proper 2u/4u chassis and server grade GPUs but it does the job for 70-90% cheaper.
