B85M-E/CSM for 2 GPUs?

[TheHeroofCanton]

The Asus B85M-E/CSM has 2 physical 16x pci-e slots, 1 of which is pci-e 3.0/2.0 @ 16x electrically, and the other being pci-e 2.0 @ 4x electrically. I don't think the board supports SLI, but I imagine that doesn't matter for folding. So could I use two pascal GPUs in it for folding to their full effectiveness?

[bruce]

If "full effectiveness" is defined as 100% utilization 100% of the time it's probable that two powerful GPUs may not to meet your requirments, no matter what other components are provided to complete all of the steps required to fold without some limitations. Along with the PCIe bus, the CPU and main RAM are also involved in the process and you can't depend them to create zero lag in filling the GPU with work to be done.

Nevertheless, those two slots will make the best approximation that's possible after subtracting whatever time is needed to actually process your computer's graphics.

[foldy]

There were reports a pcie 2.0 x4 slot could limit the bus usage on Windows but not on Linux for a fast GPU and so you could loose e.g. 20% performance but it is still worth it.

[7im]

There were reports a pcie 2.0 x4 slot could limit the bus usage on Windows but not on Linux for a fast GPU and so you could loose e.g. 20% performance but it is still worth it.

I doubt there is much difference between the OS types. And I ran a GTX 760 in an x16 2.0 slot (x1 electrically) and only lost 15% in Windows, compared to what others have reported for their top PPD numbers. Obviously, the faster the card, the more potential for PPD loss because of the (ludicrously) exponential nature of the Quick Return Bonus.

[foldy]

Maybe I calculated wrong, there may be some percent lower speed on pcie 2.0 x4 but not 20%, more like 2%.

But for fast GPUs and when going down to pcie 2.0 x1 the performance loss is 50%.
https://forums.anandtech.com/threads/pc ... t-38470000

[Aurum]

The Asus B85M-E/CSM has 2 physical 16x pci-e slots, 1 of which is pci-e 3.0/2.0 @ 16x electrically, and the other being pci-e 2.0 @ 4x electrically. I don't think the board supports SLI, but I imagine that doesn't matter for folding. So could I use two pascal GPUs in it for folding to their full effectiveness?

Your MB can run 3 cards for folding as you have:
1 x PCIe 3.0/2.0 x16
1 x PCIe 2.0 x16 (x4 mode)
1 x PCIe 2.0 x1
1 x PCI
https://www.asus.com/us/Motherboards/B8 ... fications/

You can use 1x risers on any PCIe slot:
https://www.amazon.com/gp/product/B008B ... UTF8&psc=1

This style will work if you file a slot on the inboard side (without damaging the contacts).
https://www.amazon.com/gp/product/B01IU ... KHESM7DAAM

I have a PCI to 16x PCIe adapter card from the old days I'm going to try to see if I can fold on the PCI slot. I just have one lying around, your $39 is better spent elsewhere.
https://www.amazon.com/StarTech-com-Exp ... ie+adapter

[foldy]

PCI has 500MB/s bandwidth and equals pcie 2.0 x1. This will slowdown a fast GPU by 50%.

[Aurum]

There were reports a pcie 2.0 x4 slot could limit the bus usage on Windows but not on Linux for a fast GPU and so you could loose e.g. 20% performance but it is still worth it.

I doubt there is much difference between the OS types. And I ran a GTX 760 in an x16 2.0 slot (x1 electrically) and only lost 15% in Windows, compared to what others have reported for their top PPD numbers. Obviously, the faster the card, the more potential for PPD loss because of the (ludicrously) exponential nature of the Quick Return Bonus.

When you compare PPD estimates are you using FAHbench

[Aurum]

PCI has 500MB/s bandwidth and equals pcie 2.0 x1. This will slowdown a fast GPU by 50%.

How about one of my antique GPUs

I'd like to learn how you calculate that.

[bruce]

An "antique" GPU will probably have a reduced number of shaders which will consequently need less bandwidth than one with more shaders. I'm sure there are some GPUs with poor enough performance that a more limited pcie speed won't add any additional limitations to what you already have. Put the fastest GPU in the fastest slot.

[foldy]

The bandwidth numbers are from wikipedia and the pcie 2.0 x1 vs x16 test was done in anandtech forum with a pcie x16 to x1 raiser using FahBench.
slow R9-280X: x16 42.3, x1 38.5 => 9% slower
fast gtx 980 ti: x16 81.4, x1 41.7 => 51% slower
https://forums.anandtech.com/threads/pc ... t-38470000

pcie 2.0 x4 should be fine also for fast GPUs.

[MARSTG]

Yes you could, I am using a H57M USB3 that has a similar configuration with a GTX 1060 3GB and a 280X and they work just fine.

[Aurum]

I'm more interested in the other end of the spectrum as I'm building a beast for my next folding rig with an i7-4970K and this MB:
http://www.gigabyte.com/products/produc ... id=4950#ov
Specs say:
1 x PCI Express x16 slot, running at x16 (PCIEX16)
1 x PCI Express x16 slot, running at x8 (PCIEX8)
* The PCIEX8 slot shares bandwidth with the PCIEX16 slot. When the PCIEX8 slot is populated, the PCIEX16 slot will operate at up to x8 mode.
1 x PCI Express x16 slot, running at x4 (PCIEX4)
* The PCIEX4 slot shares bandwidth with the PCIEX8 and PCIEX16 slots. When the PCIEX4 slot is populated, the PCIEX16 slot will operate at up to x8 mode and the PCIEX8 will operate at up to x4 mode.

My plan was to install 3 GTX 1070s so this means they'll be running at x8, x4, x4. If x4 degrades performance then that's just a waste implying I should install a 1070 in PCIEX16 and 2 weaker cards in PCIEX8 & PCIEX4.

Your data says that building multi-GPU folding rigs with high end cards is a big mistake, especially if one uses 1x risers.

[foldy]

I was a bit misleading, pcie 2.0 x1 is too slow for fast GPUs but pcie 2.0 x4 is fine.
In your case it is pcie 3.0 x4 which equals pcie 2.0 x8 so there is no performance loss

[Aurum]

I was looking for a utility that monitors the traffic on the PCIe bus. It may be like the Heisenberg Uncertainty Principle, to measure it is to change it. It seems that Nvidia includes nvidia-smi.exe that might do it. FAHbench is not useful on a multi-gpu rig because it only runs one gpu at a time. All GPUs need to be under load to learn how to optimize a multi-gpu rig. If all GPUs could be assigned the same WU simultaneously then that would do it. They don't distribute any documentaion with FAHbench so I'm going to have go find the instructions for using FAHBench-cmd.exe again. I think there's a command line option that will select which GPU to benchmark but I don't recall seeing one that allows multiple GPUs to run at the same time. Even if only one GPU reports they all need to be under full load. I can try running F@H 7.4.15 and rightclicking a single GPU, Pausing it, and then running FAHbench on the Paused GPU.