CPU For Folding Rig

[cysquatch]

I understand PPD of cpu's is negligible.  Should I just grab whatever cpu/motherboard for a dedicated folding rig?  I plan on running at least 2 gpu's on the build.  Thanks in advance 

[BWG]

Which GPU's?

[Avacado]

27 minutes ago, cysquatch said:

I understand PPD of cpu's is negligible.  Should I just grab whatever cpu/motherboard for a dedicated folding rig?  I plan on running at least 2 gpu's on the build.  Thanks in advance 

In a 2 GPU scenario, yes. Most any CPU/MOBO combo will do. 

[cysquatch]

17 minutes ago, Avacado said:

In a 2 GPU scenario, yes. Most any CPU/MOBO combo will do. 

 What about 3 if I end up delving that deep into this?  Sorry,  I have zero experience with multiple gpu's.  Thanks

[cysquatch]

20 minutes ago, BWG said:

Which GPU's?

 I plan on getting a 3080 minimally when they become more readily available.  Eventually I would Like 2 of them

[Bastiaan_NL]

One of my dedicated systems is using an i3 9100f with a cheap motherboard and 8gb ram. It has been online 24/7 for almost a year now, without any problems and always with 2 GPU's. Folding on that CPU is not a good idea with 2 GPU's though, even with only 2 cores for the CPU client it'll use 100% CPU with every checkpoint and that will have a big impact on the GPU PPD.

 

11 minutes ago, cysquatch said:

 What about 3 if I end up delving that deep into this?  Sorry,  I have zero experience with multiple gpu's.  Thanks

The problem is having enough PCI-e lanes on the CPU/Motherboard. With 3080's you'll need 8x lanes per GPU if you don't want to bottleneck them.

Edited March 19, 2021 by Bastiaan_NL

[cysquatch]

6 minutes ago, Bastiaan_NL said:

One of my dedicated systems is using an i3 9100f with a cheap motherboard and 8gb ram. It has been online 24/7 for close to a year now, without any problems and always with 2 GPU's. Folding on that CPU is not a good idea with 2 GPU's though, even with using only 2 cores for the CPU client it'll use 100% CPU with every checkpoint and that will have a big impact on the GPU PPD.

 

The problem is having enough PCI-e lanes on the CPU/Motherboard. With 3080's you'll need 8x lanes per GPU if you don't want to bottleneck them.

What gpu's are you using with that cpu?

[Bastiaan_NL]

A pair of 5700xt's for around 6 months, and for the other 5 months a 2080ti and a 2080 super. I've been switching cards in and out of that system a few times, for testing etc.

That system is at 9 million PPD at this moment with the current Beta units!

 

[edit]

In the previous post I mentioned that with 3080's you'll need at least 8x lanes per GPU, but that's the case for all the fast GPU's out there.
I'm not sure where the tipping point is, maybe someone else over here knows that. 

Edited March 19, 2021 by Bastiaan_NL

[Avacado]

51 minutes ago, Bastiaan_NL said:

A pair of 5700xt's for around 6 months, and for the other 5 months a 2080ti and a 2080 super. I've been switching cards in and out of that system a few times, for testing etc.

That system is at 9 million PPD at this moment with the current Beta units!

 

[edit]

In the previous post I mentioned that with 3080's you'll need at least 8x lanes per GPU, but that's the case for all the fast GPU's out there.
I'm not sure where the tipping point is, maybe someone else over here knows that. 

You'll still get it's full potential with at least 4x. Even with higher end cards. What matters more is proper power regulation. Ensure the cards are getting enough juice. 

Edited March 19, 2021 by Avacado

[cysquatch]

9 minutes ago, Avacado said:

You'll still get it's full potential with at least 4x. Even with higher end cards. What matters more is proper power regulation. Ensure the cards are getting enough juice. 

 

I'm hoping thats the case, there isnt much out there in the HEDT market right now... 

[Avacado]

2 minutes ago, cysquatch said:

 

I'm hoping thats the case, there isnt much out there in the HEDT market right now... 

Most modern chips have 16 lanes anyway. You'll be fine as long as its within the past 8 years or so.

[Bastiaan_NL]

45 minutes ago, Avacado said:

You'll still get it's full potential with at least 4x. Even with higher end cards. What matters more is proper power regulation. Ensure the cards are getting enough juice. 

I'm sorry if I misinformed anyone, for some reason I was sure that someone tested it but I can't find that. 

Is the bus interface load from GPU-z a good measuring tool for PCI-e bandwidth? If so I'll have some numbers which would indicate that you are right besides the 3090 spiking a few times above 4GB/s.

 

 

 

[Avacado]

14 minutes ago, Bastiaan_NL said:

I'm sorry if I misinformed anyone, for some reason I was sure that someone tested it but I can't find that. 

Is the bus interface load from GPU-z a good measuring tool for PCI-e bandwidth? If so I'll have some numbers which would indicate that you are right besides the 3090 spiking a few times above 4GB/s.

 

 

 

Absolutely. GPU-z will list PCI-e speed, multiplier and bandwidth. I don't have a 3090, so maybe you could experiment for us. At least for the 2080's and Teslas, I haven't noticed a performance impact between 16/8/4. Maybe 10% at most with 4, but usually not. 1x for sure there is a huge impact. 

Edited March 19, 2021 by Avacado

[cysquatch]

I may have found a decent deal on an I7 6700k combo, with a z170 board, that should be good right?

[Bastiaan_NL]

That should be no problem

I've got a z170/7700k combo running without an issue, has been folding with 2 GPU's too. 

 

As for the 3090 bandwidth experiment, I've done some basic testing and have not seen more than 14% (4.48GB/s) usage of a 4.0 x16 slot with an average of 9% (2.88GB/s). 
So a 3090 will hit the max of a 3.0 x4 (4GB/s) slot every now and then, but I've not tested what that means in TPF/PPD. I could not get FAHBench working with a copied beta workunit, will try that again later.

 

[BWG]

On 3/20/2021 at 4:08 PM, Bastiaan_NL said:

That should be no problem

I've got a z170/7700k combo running without an issue, has been folding with 2 GPU's too. 

 

As for the 3090 bandwidth experiment, I've done some basic testing and have not seen more than 14% (4.48GB/s) usage of a 4.0 x16 slot with an average of 9% (2.88GB/s). 
So a 3090 will hit the max of a 3.0 x4 (4GB/s) slot every now and then, but I've not tested what that means in TPF/PPD. I could not get FAHBench working with a copied beta workunit, will try that again later.

 

 

Geesh, that's a lot of headroom still on 4.0!