Seeking some basic hardware advice.

[scoochykid]

I have 2 existing RTX 4090 GPUs, I am soon going to buy another 2.

I just would appreciate some thoughts on my upgrade path.

I know preferably, GPUs would be running at 4.0/5.0 x16, with 64GB RAM per GPU {256gb total}

But I am not sure I want to go to those lengths of buying a server motherboard, threadripper etc.

If I just have 1 card on x16 and the other 3 at x1. With 64 GB (total system memory). Am I missing out on "A LOT"

I'm not a professional, I'm a hobbyist - but I also don't want to lose out on crazy performance gains.

Also slightly worried the new big hashcat update may somehow utilise x16 cards more and have major performance improvements etc.

Any comments are appreciated!

[Chick3nman]

Not sure about the 64gb of RAM per GPU thing, that really shouldn't matter so much and as long as you aren't seeing out of memory issues, you're fine on that front. Typically the "suggested" RAM is "system RAM >= combined VRAM" so for 4x 4090, that'd be 96gb or more, but in reality this is only to stop out of memory/out of resources issues from the runtime and if you haven't been hitting those then nothing to worry about. If you DO run into those on some attack, however, the fix will unfortunately be to either change the attack in some way or to add RAM.

As for the bus width, this is a much bigger issue. The minimum bus width I would suggest is ~3.0 @ x4 before you start to see serious performance degradation with cards that big. In reality though, it doesn't have to matter a ton as in most attacks you should be trying to create an ideal workload with high amplifiers to help with the PCIe bottleneck _anyways_, but it can make a pretty significant difference in certain attacks or if you're not structuring them well. If we do some rough guessing for a 4090, (assume MD5 @ N=256, 128 SMs, ideal threads 32, w/ 3.0 @ x1=8Gb/s, saturation is 2.281Gb/s) we see an upper bounds on the loss as being ~27.9%, a little more than 1/4 of your theoretical performance, while using 3.0 @ x16 shows only ~1.7% theoretical loss. This sounds really bad, and it is if you're attack is heavy on the bus such as running just a wordlist, but you can certainly restructure attacks to try and make some of this up so if it's all you have it may be good enough for you.

Please see here for some, albeit outdated and a little wonky, example math:
https://hashcat.net/wiki/doku.php?id=fre...king_speed