building 4x 4090 rig

[Siegfried1986]

I am currently using 2x 4090 rig which is simply attached via PCIex 4.0 and 5.0 slots. I suppose both x16. 

 planning to reassamble those GPUs into 4x 4090 rig on minig motherboard. 


I have question, especially to people who have 3090 rigs.  Is it so important , that GPUs will be connected via PCIex x1, will it result in great speed loss ? 
Or I could connect them via PCIex X1 and use for hashcat with no problem ?


P.s using 2x 4090 for hashcat mode 400 for long time, and totally satisfied with it

[Siegfried1986]

I am already reading some big posts regarding this topic from "Chick3nman", but maybe some other people could give their advice!
I have some doubts about how the GPU bus is important

[Chick3nman]

Unfortunately, there's not a lot of other people active on here that work as closely with the hardware as I do, so a lot of the relevant posts will be from me. I'm not sure why you would think that bus speed isn't important to cracking, the transfer of candidate data from the host to the GPU happens across the bus and it usually ends up being one of the largest sources of bottlenecking, aside from work distribution issues. It's detailed a bit more thoroughly in the hashcat FAQ here: https://hashcat.net/wiki/doku.php?id=fre...king_speed

[Siegfried1986]

Unfortunately, there's not a lot of other people active on here that work as closely with the hardware as I do, so a lot of the relevant posts will be from me. I'm not sure why you would think that bus speed isn't important to cracking, the transfer of candidate data from the host to the GPU happens across the bus and it usually ends up being one of the largest sources of bottlenecking, aside from work distribution issues. It's detailed a bit more thoroughly in the hashcat FAQ here: https://hashcat.net/wiki/doku.php?id=fre...king_speed

Hello Dear Chickenmann. I strongly appreciate your determination to the things you are involved to. Your posts have answers on many hash-cracking-related questions. I would like to ask you a straight question. I know there are many many things which impact on speed , but it will try to explain : 

I am going to expand my rig which consists of : 
2x4090 on x16 pci's
i7 13700k
64gb ddr5 ram

I want to make new one on asus server motherboard with
4 x 4090 connected via 8-16x PCIs
i7 14700k
64 or 128 GB of ram ? 1)I would be so gratefull if you could advice me here

I am using hashcat for 2 types of attack
phpass (mode 400) with large wordlists (like 10 billion strings) 
md5 (mode 0)  mask attack , like ?a?a?a?a?a?a etc

On my 4090 , using -m400 I have 12kH\s speed\per one . (md5 -- like about 68 GH\s) 
2 cards = about 24000 So my aim is to get at least 1.8-1.9x on speed by adding 2 cards.
2)Would not be that a problem for i7 14700k to work with 4 GPUs on those attack modes ? 

That is all. 
If you just say "14700 will bottlenceck 4 cards " -- I skip this idea. (or just make 4k more bucks for Threadripper)
If you say like "yes 14700k is strong enough and you just need 128 RAM (or do not need)" --  I will make that rig in 2 weeks and name it into your honor. I have laser engraving machine so I will engrave your nickname on the box with 4 gpus 


P.S what about bus speed? I decided making that rig on 1x PCI's is wrong idea, so I am not even looking at it as a variant. I am currently inquiring about possibilty to make that on normall motherboard.


Execuse me for being noobish in some terms, but I tried to elaborate that the most clearly.


P.S.S If I disconnect 1 GPU now, and try running md5 or phpass hash mode -- speed is not incrasing, so I could make a conclusion, 13700 is not bottlenecking 2 GPUs

[Siegfried1986]

I have been inspired by rig made by Pufet systems

[Chick3nman]

To be clear, it's not the CPU power/performance that will cause issues. It's the lack of available PCIe lanes. The 14700k only has 20 available lanes. The 13700k is actually the same, give or take on chipset switched lanes(usually for storage devices). You are already not running at x16 on both cards, regardless of what the slots look like. The reason people use threadripper and xeon/epyc CPUs is not because we need the CPU performance necessarily, but because we need the available PCIe lanes. Just because the motherboard has a slot that appears to be x16 physically, does NOT mean that it's connected to x16 lanes electrically. If you want to populate a board with more GPUs, you need to look at how the available lanes will be split up from the CPU/Chipset. A 13700k/14700k can run in x16+x4 with 2 GPUs, which is enough to not notice any major performance losses. Adding more GPUs, however, is either not possible or will require reducing the bus width across all of the GPUs. In THEORY you have enough to give every GPU x4 at least, but I don't know of a motherboard that is setup in a way that allows that right now. Most likely the only configs possible are x8+x4+x1+X1 or siimilar.

The reason Puget is using a threadripper on that board is because it has 128 PCIe lanes available, which is more than enough to provide each GPU with x16. You will not be able to replicate their results on just any motherboard/CPU combo. You would need something that has and exposes enough PCIe lanes for all of your GPUs to get at least x4 lanes to avoid performance issues.

[CmdFlaz]

Have you thought about that the new RTX series is out late this year with Blackwell Architecture, thought to be twice as fast as the current RTX 4090, Might be most cost efficient to wait