What is the best datacentre GPUs for hashcat?

[starfish]

Enterprise hardware
If you wanted to run hashcat on enterprise hardware, thus could not use consumer GPUs, what datacentre GPUs would you recommended?

I have reviewed the benchmarks for the A100 & H100, they do not come close (assuming I am reading it correctly) to the performance of RTX 4090. Plus the A100/H100 are at a much higher cost. I assume this is due to the types of cores they have and the underlying  architecture.

I have noted in the most recent conversation ( thread 11629 ) regarding the most performant GPUs and @Chick3nman suggested Quadro RTX 6000 Ada GPU,  if you cannot use the RTX 4090. However, this is still considered a consumer GPU.

Looking at https://www.nvidia.com/en-us/data-center GPUs, I was thinking the L40 that is based on the Ada Lovelace architecture (same architecture as the RTX 6000 Ada?) might offer the best performance? If this is the case, any idea how it would compare to the RTX 4090?

Apart from the A100/H100 I have struggled to find benchmarks for other NVIDIA Datacentre GPUs - if anyone can point me in the direction that would be great.


Consumer Hardware
In addition to my question above, if you do use consumer GPUs, I think the recommendation (based on what I have read on this forum plus benchmarks on GitHub) would be either:
 4 or 6 x  RTX 4090
 8 x  RTX 6000 Ada

Does anyone know if that is correct?

Thanks

[Chick3nman]

Consumer Hardware

 8 x  RTX 6000 Ada

Why do you think the RTX 6000 Ada is a consumer GPU? Those are Quadros, which are considered professional GPUs and are used in datacenters all the time.

[starfish]

Consumer Hardware

 8 x  RTX 6000 Ada

Why do you think the RTX 6000 Ada is a consumer GPU? Those are Quadros, which are considered professional GPUs and are used in datacenters all the time.

I agree they are professional GPUs but they are not listed on NVIDIA's datacentre GPU products page. NVIDIA Data Center Products | NVIDIA

I am not saying they cannot be used in datacentres and they do blur the line, but I though they are considered high-end workstation level consumer GPUs? When you look at the server product portfolio of the likes of HPE and Dell, their servers support GPUs listed on NVIDIA's datacentre GPU products page not the RTX 6000 Ada or alike. (at least what I have found)

I do note some other vendors offer them in a server chassis.

What are your thoughts on the L40? Do you have any thoughts above marc1n's reply?

Many Thanks

[starfish]

Thank you marc1n for your very detailed reply and confirming my understanding.

Do you have any thoughts on the new L40S vs the L40?

Am I also correct in thinking cards (noting L40 does not) offering NVLink /NVSwitch offer no benefit to hashcat?

[starfish]

The L40S is 350w ( NVIDIA L40S Product Brief ) where as the L40 is 300w ( product-brief-L40.pdf (nvidia.com) - I presume your reply stands but inverse your references between L40 and L40S? The L4 looks to be the original low power one ( PB-11316-001_v01.pdf (nvidia.com) )

FYI: my requirement is highest possible performance, not power consumption.

Regarding

It can only use one GPU at a time

are you referring to if you had them connected with NVlink it could only use one? But if you have multiple GPUs not connected via NVLink it can use them at the same time or can you just not do parallelized processing?

Thank you for taking the time to explain.

[Chick3nman]

The L40S is 350w ( NVIDIA L40S Product Brief ) where as the L40 is 300w ( product-brief-L40.pdf (nvidia.com) - I presume your reply stands but inverse your references between L40 and L40S? The L4 looks to be the original low power one ( PB-11316-001_v01.pdf (nvidia.com) )

FYI: my requirement is highest possible performance, not power consumption.

Regarding

It can only use one GPU at a time

are you referring to if you had them connected with NVlink it could only use one? But if you have multiple GPUs not connected via NVLink it can use them at the same time or can you just not do parallelized processing?

Thank you for taking the time to explain.

The posts you are referencing have been removed as they were clearly not accurate. They were seemingly generated by an LLM such as ChatGPT. Please don't believe what you've been told so far, there are lots of issues with what was said hence my removal.

For example: "It can only use one GPU at a time" This is simply false and I have no idea why anyone would repeat it willfully. Hashcat can use as many GPUs as are present and accessible, there's no reason why it would be limited to only one.

Could you start over with what information you are looking for and what you believe so far? I will follow up with a real answer as soon as I have time.

[ManuB1G]

From my point of view L40 will the best Datacenter GPU for Hashcat. It has more shaders (18176) than 4090, H100 and A100.
The L40S seems to have a higher clock (2490 MHz vs 2520 MHz) but the +50 W are a high price for not even +1%. 8 of them in a server are +400W, this is why we decided to use L40 because they are a third of H100 price and the fast memory of H100 is not a huge performance increase for hashcat. As far as I know the RTX A6000 Ada is close to the L40 but with a fan for workstation usage.

[Chick3nman]

Correct, the RTX 6000 Ada (not A6000) does have a fan, though it can work in a server chassis too. The L40 (or its variants) will work fine as a replacement if you can get them, though I'm not sure if it's as cost effective as the RTX 6000 Ada. If cost effectiveness isn't as big of a deal for you, then I would stick with the L40 GPUs as they seem to fit the idea of a "datacenter GPU" that you have. You may find that they don't behave quite as well as you would expect out of the box though, due to the limited amount of time I've had to use them and test them. If that's the case, please reach out and I'll ensure any tuning or other issues are corrected.

[ManuB1G]

Thanks, I will let you know, still waiting for Nvidia to deliver….

[starfish]

Thanks Chick3nmann - the information from marc1n was getting somewhat contradictory. 

My understanding is as follows, please correct where I am misinformed:

• Hashcat will use as many GPUs as you give it
  
• Hashcat does not benefit from NVLink/NVSwitch
  
• Hashcat does not benefit from Tensor cores
  
• Hashcat does not benefit from RT cores
  
• The number and clock speed of CUDA/FP32 Cores are important to the performance
  

Regarding GPUs (using NVIDIA's definitions for datacentre and workstation)

• The RTX 4090 is the best performing consumer card you have tested.
  
• The RTX 6000 Ada is the best performing workstation card and should be slightly faster than the RTX 4090 due to its increased number of cores (18,176 vs 16,384) and very similar clock speed. 
  
• The L40 is the best performing 'datacentre' card and likely to perform similar to the RTX 6000 Ada.
  

As you note, other factors include card cost and availability of them in the desired server vendor.

You said you did some testing with the L40, do you have any benchmarks?

Many Thanks