building 4x 4090 rig

[Chick3nman]

https://www.aliexpress.com/item/10050037...5617Ic3BuR

These are very bad for hashcat, I would strongly recommend AGAINST using risers and splitters such as these as they will significantly impact both performance and stability.

[slyexe]

These are very bad for hashcat, I would strongly recommend AGAINST using risers and splitters such as these as they will significantly impact both performance and stability.

I think people are not even listening to your input chickenman. Literally you just explained to them the significance of using proper PCIE lanes and they just completely obliterated any suggestions you had provided lol. Sorry man, its a losing battle it seems with this one.

[174region174]

https://www.aliexpress.com/item/10050037...5617Ic3BuR

These are very bad for hashcat, I would strongly recommend AGAINST using risers and splitters such as these as they will significantly impact both performance and stability.

* Hash-Mode 22000 (WPA-PBKDF2-PMKID+EAPOL) [Iterations: 4095]
-------------------------------------------------------------

Speed.#1.........:  1107.0 kH/s (73.50ms) @ Accel:8 Loops:1024 Thr:512 Vec:1
Speed.#2.........:  1103.2 kH/s (73.73ms) @ Accel:8 Loops:1024 Thr:512 Vec:1
Speed.#3.........:  1069.8 kH/s (75.67ms) @ Accel:8 Loops:1024 Thr:512 Vec:1
Speed.#4.........:  1123.1 kH/s (72.42ms) @ Accel:8 Loops:1024 Thr:512 Vec:1
Speed.#*.........:  4403.2 kH/s

* Hash-Mode 1000 (NTLM)
-------------------------------------------------------------

Speed.#1.........: 46052.0 MH/s (31.45ms) @ Accel:128 Loops:1024 Thr:256 Vec:8
Speed.#2.........: 46472.3 MH/s (30.98ms) @ Accel:128 Loops:1024 Thr:256 Vec:8
Speed.#3.........: 45206.4 MH/s (32.15ms) @ Accel:128 Loops:1024 Thr:256 Vec:8
Speed.#4.........: 46785.9 MH/s (30.89ms) @ Accel:128 Loops:1024 Thr:256 Vec:8
Speed.#*.........:  184.5 GH/s

[Chick3nman]

https://www.aliexpress.com/item/10050037...5617Ic3BuR

These are very bad for hashcat, I would strongly recommend AGAINST using risers and splitters such as these as they will significantly impact both performance and stability.

* Hash-Mode 22000 (WPA-PBKDF2-PMKID+EAPOL) [Iterations: 4095]
-------------------------------------------------------------

Speed.#1.........:  1107.0 kH/s (73.50ms) @ Accel:8 Loops:1024 Thr:512 Vec:1
Speed.#2.........:  1103.2 kH/s (73.73ms) @ Accel:8 Loops:1024 Thr:512 Vec:1
Speed.#3.........:  1069.8 kH/s (75.67ms) @ Accel:8 Loops:1024 Thr:512 Vec:1
Speed.#4.........:  1123.1 kH/s (72.42ms) @ Accel:8 Loops:1024 Thr:512 Vec:1
Speed.#*.........:  4403.2 kH/s

* Hash-Mode 1000 (NTLM)
-------------------------------------------------------------

Speed.#1.........: 46052.0 MH/s (31.45ms) @ Accel:128 Loops:1024 Thr:256 Vec:8
Speed.#2.........: 46472.3 MH/s (30.98ms) @ Accel:128 Loops:1024 Thr:256 Vec:8
Speed.#3.........: 45206.4 MH/s (32.15ms) @ Accel:128 Loops:1024 Thr:256 Vec:8
Speed.#4.........: 46785.9 MH/s (30.89ms) @ Accel:128 Loops:1024 Thr:256 Vec:8
Speed.#*.........:  184.5 GH/s

This is in benchmark mode, which uses bruteforce/mask mode where the bus width and speed doesn't have almost any impact. In straight/wordlist, you will see speed/stability issues as the bus will be completely saturated trying to move candidates to the GPU from the host and the GPU will not be fully utilized at all.

[174region174]

https://www.aliexpress.com/item/10050037...5617Ic3BuR

These are very bad for hashcat, I would strongly recommend AGAINST using risers and splitters such as these as they will significantly impact both performance and stability.

* Hash-Mode 22000 (WPA-PBKDF2-PMKID+EAPOL) [Iterations: 4095]
-------------------------------------------------------------

Speed.#1.........:  1107.0 kH/s (73.50ms) @ Accel:8 Loops:1024 Thr:512 Vec:1
Speed.#2.........:  1103.2 kH/s (73.73ms) @ Accel:8 Loops:1024 Thr:512 Vec:1
Speed.#3.........:  1069.8 kH/s (75.67ms) @ Accel:8 Loops:1024 Thr:512 Vec:1
Speed.#4.........:  1123.1 kH/s (72.42ms) @ Accel:8 Loops:1024 Thr:512 Vec:1
Speed.#*.........:  4403.2 kH/s

* Hash-Mode 1000 (NTLM)
-------------------------------------------------------------

Speed.#1.........: 46052.0 MH/s (31.45ms) @ Accel:128 Loops:1024 Thr:256 Vec:8
Speed.#2.........: 46472.3 MH/s (30.98ms) @ Accel:128 Loops:1024 Thr:256 Vec:8
Speed.#3.........: 45206.4 MH/s (32.15ms) @ Accel:128 Loops:1024 Thr:256 Vec:8
Speed.#4.........: 46785.9 MH/s (30.89ms) @ Accel:128 Loops:1024 Thr:256 Vec:8
Speed.#*.........:  184.5 GH/s

This is in benchmark mode, which uses bruteforce/mask mode where the bus width and speed doesn't have almost any impact. In straight/wordlist, you will see speed/stability issues as the bus will be completely saturated trying to move candidates to the GPU from the host and the GPU will not be fully utilized at all.

You are absolutely right. And to avoid such a situation, you need to use rules or use another type of attack. I'm sure the same thing will happen to you if you install your GPU in pci-e slots. And there will be no difference...

[Chick3nman]

You are absolutely right. And to avoid such a situation, you need to use rules or use another type of attack. I'm sure the same thing will happen to you if you install your GPU in pci-e slots. And there will be no difference...

This is simply wrong. The difference between x1 lanes at 3.0 and x16 lanes at 4.0 is almost 32x faster. PCIe x1 @ 3.0 maxes out at ~1GB/s, x16 @ 4.0 can do 31.5GB/s. You can push a hell of a lot more candidates per second to your card with x16 @ 4.0 than you can with x1 @ 3.0.



I implore you to read through the FAQ link: https://hashcat.net/wiki/doku.php?id=fre...king_speed

Even on very old GPUs, the difference can be 15%+ performance loss at x1. Yes rules can latency hide by amplifying, but raw chunk load speed is still important because rules and amps are not always feasible.

[174region174]

You are absolutely right. And to avoid such a situation, you need to use rules or use another type of attack. I'm sure the same thing will happen to you if you install your GPU in pci-e slots. And there will be no difference...

This is simply wrong. The difference between x1 lanes at 3.0 and x16 lanes at 4.0 is almost 32x faster. PCIe x1 @ 3.0 maxes out at ~1GB/s, x16 @ 4.0 can do 31.5GB/s. You can push a hell of a lot more candidates per second to your card with x16 @ 4.0 than you can with x1 @ 3.0.



I implore you to read through the FAQ link: https://hashcat.net/wiki/doku.php?id=fre...king_speed

Even on very old GPUs, the difference can be 15%+ performance loss at x1. Yes rules can latency hide by amplifying, but raw chunk load speed is still important because rules and amps are not always feasible.

Thanks for the link you wrote. the fact is that I never search with pure dictionaries. Therefore, it doesn't matter to me which PCI-E slot the GPU is installed in.

This is exactly what I wrote about. I have known about this for many years. And not from this article, but from my personal search experience.

[Chick3nman]

Thanks for the link you wrote. the fact is that I never search with pure dictionaries. Therefore, it doesn't matter to me which PCI-E slot the GPU is installed in.

This is exactly what I wrote about. I have known about this for many years. And not from this article, but from my personal search experience.

It is also good to note that this will not only impact dictionary attacks without rules, but also ANY attacks against hash modes that use a host side hook for a portion of the computation. Several modes, such as 7z, do a portion of the computing on the GPU such as the KDF and then hand the data off to the CPU to do an AES/Compression step partway through the kernel. These modes will be severely impacted by bus speeds as well, regardless of your ability to amplify work on the GPU with rules or similar. There's more than one reason why sticking GPUs on x1 lane risers for hashcat is a bad idea and they may not be easy to spot if you aren't familiar with the underlying code.

[evilmog]

I used to work with x1 lanes and risers back in 2016, they absolutely cratered systems then and tanked many attack types, I can't imagine it being any better now

[q6h7jsvdf]

@Chick3nman

To my understanding, current RTX 4090 doesn't saturate PCIe 4.0 x16 in gaming, showing only 2% loss on PCIe 3.0 x16. I'm wondering if next-gen GPUs like RTX 5090 or upcoming AMD cards will saturate PCIe 4.0 x16, and how this applies to compute tasks like hashcat versus gaming.
I have a PCIe 4.0 x16 board, - what's the risk of performance loss with future PCIe 5.0 x16 GPUs?
Essentially, I'm trying to gauge if PCIe 4.0 x16 will suffice for upcoming GPUs in compute scenarios (hashcat specifically). Your thoughts?