Wow, that's a huge feature. Thank you, guys. By the way, does anyone have a preferred case where they can put 4 AMD 7900XTX? There's a lot of motherboards and CPUs that support 128 lanes. It's the physical arrangement that I have trouble with.
You don't need 128 lanes. 8x PCIe3 is more than enough, so for 4 cards that's 32. Most CPUs have about 40lanes. If you are not doing much that would be more than sufficient. Buy a PCIe riser. Go to amazon and search for it, a 16x to 16x PCIe riser. They go for about $25-$30 often about 20-30cm. If you want really long one, you can get one from China a 60cm for about the same price, you just have to wait for 3 weeks. That's what I did. Stuffing all those in a case is often difficult, so you have to go open rig. Either have the cables running out your computer and figuring out a way to support the cards while keeping them cool or just buy a small $20-$30 open rig frame.
Most CPUs have about 20 lanes (plus a link to the chipset).
On the one hand, they will be gen 4 or 5, so they're the equivalent of 40-80 gen 3 lanes.
On the other hand, you can only split them up if you have a motherboard that supports bifurcation. If you buy the wrong model, you're stuck dedicating the equivalent of 64 gen 3 lanes to a single card.
Edit: Actually, looking into it further, current Intel desktop processors will only run their lanes as 16(+4) or 8+8(+4). You can kind of make 4 cards work by using chipset-fed slots, but that sucks. You could also get a PCIe switch but those are very expensive. AMD will do 4+4+4+4(+4) on the right boards.
Ah ha, that's the part I was curious about. I was wondering if I could keep everything cool open rig. I'm waiting for the stuff to arrive: risers, board, CPU, GPUs. And I've been putting it off because I wasn't sure how about the case. All right then, open rig frame. Thank you!
Crypto mining didn't require significant bandwidth to the card. Mining-oriented motherboards typically only provisioned a single lane of PCIe to each card, and often used anemic host CPUs (like Celeron embedded parts).
Does LLM inference require significant bandwidth to the card? You have to get the model into VRAM, but that's a fixed startup cost, not a per-output-token cost.
