Folding Forum

One of my folding boxes has an H370 chipset, which means the top x16 slot has 16 lanes through the CPU and the next x16 slot has 4 lanes through the PCH. I experimented with putting a GTX980 in the top slot and an RTX2070 in the bottom slot, and I noticed with the RTX2070 in the bottom slot it gets as much as 200k-300k less ppd than when it's in the top slot. It seems most noticable on the 14*** projects.

When the RTX is in the bottom slot, GPU-Z shows the TDP % as significantly less. At first I thought maybe there was a power issue, but upon running a test mining process, the GPU is able to hit 100% TDP in that bottom slot. So that left me wondering if there is a data bottleneck. Even with 4 PCI-E 3.0 lanes through the PCH, I'd figure throughput should be plenty fast. But I don't know a lot about what's going back and forth there.

So all that said, I have two questions:

1) What exactly causes the reduction in ppd in the above case? Is it the 4 lanes going through the PCH, or could it be something else I'm not thinking of? The system CPU is a Pentium Gold G5400 and it runs about 50-60% with two cards folding, so I didn't think there was any cpu bottleneck. The system also has 2x8GB system RAM.

2) If I replaced the H370 board with a Z370/390, that is, something that can split the PCI-E CPU lanes between the cards, would that solve the ppd drop issue?

Yes it is the pcie x4 slowing down the rtx 2070. Guess it doesn't matter if it comes through PCH but maybe that makes things worse.
What mainboard to you have ...H370...? Another mainboard with pcie x4 will not make the rtx 2070 faster.
The gtx 980 is slower and so uses less pcie bandwidth. rtx 2070 to the x16 slot and gtx 980 to x4 slot would be best.
Do you use Windows or Linux? Linux has better performance than Windows especially on slow pcie connections.

The PCIe bandwidth topic is already discussed here : viewtopic.php?f=38&t=28847

Think of it this way: When you watch a streaming video, data comes through your internet connection and must be processed by your computer or TV. Under "normal" circumstances, your internet connection is fast enough to stay ahead of what you're seeing on the screen, but if your internet is too slow (or your satellite connection is interrupted by heavy rain) the picture will break up until more data arrives. The same sort of thing happens when your GPU is slow enough that it never runs out of data. Put a fast GPU on a slow PCIe connection and the GPU's processing will have to pause until more data arrives.

The PCIe 1x to 16x (see picture in the reference topic) will slow down procesing of mid- to high-range GPUs. A 4x slot will do the same, but only with high-range GPUs.

Which saturates first, the GPU or the PCIe connection ... and for what percentage of the time?

doctorlexus wrote:1) What exactly causes the reduction in ppd in the above case? Is it the 4 lanes going through the PCH, or could it be something else I'm not thinking of? The system CPU is a Pentium Gold G5400 and it runs about 50-60% with two cards folding, so I didn't think there was any cpu bottleneck. The system also has 2x8GB system RAM.

My latest build is quite similar - coffee lake pentium and a B360 MB. Running a pair of 2080 Ti, the bottom slot actually gets higher PPD since it boosts higher. However, this is under linux.Folding under windows is known to be more dependent on PCI bandwidth.

The other thing is cooling - is the bottom slot as well ventilated? Sometimes it's pushed right up to the bottom of case / PSU. However, normally it's still better off than the top slot that gets blocked by the hot 2nd GPU...

I wouldn't swap to a z370/390 board - you might just end up having 2 slowed cards! The PCIe switches used in these motherboards can add their own issues too.

I personally put the fastest cards in the full 16x slots.
I think a 1x slot can handle about 100k PPD in Windows, but supposedly more in Linux.
It wouldn't surprise me if a 2070 is too much for a 4x slot in WIndows.
If it's only 200-300k PPD, I don't think Linux will help.
I wished there were some risers that would be able to split 16x slots in 2x 8x slots.

ProDigit wrote:I wished there were some risers that would be able to split 16x slots in 2x 8x slots.

Yes, it would be nice. Is that even possible?

bruce wrote:

ProDigit wrote:I wished there were some risers that would be able to split 16x slots in 2x 8x slots.

Yes, it would be nice. Is that even possible?

Possible, yes. I think I even came across something like that in the past. But it involves a higher level of complexity and would be expensive. Simpler and cheaper to just start with a logic board that is already designed to support two or three video cards.

Interesting to hear Windows handling of pci-e may be the culprit. Also interesting to hear about the two 2080Ti's on a B360. I think, given the information here, I'd probably waste my time replacing my mainboard (although I do think I'll spring for the higher end chipsets in the future).

Yes, I put the GTX980 in the top slot because of heat. The bottom slot is much better for cooling, and my RTX2070 is the more valuable card. I've since removed the GTX980 and am just running the RTX2070 solo in the box now. Less energy, more efficient all around.

I'm currently running the GTX 1050 OC in a PCIE 1x 2.0 slot, and I see less bandwidth usage than with the GT 1030 (possibly due to the board auto switching between 1. v1.1 and 1x v2.0).

The GT 1030 had 29% PCIE traffic, the 1050 has a mere 8%.
Though slightly overclocked, the 1050 hovers around the same PPDs as before.
Still have to verify when the first few runs will be in.

Not sure if this is the cause of running different WUs?

Also nice would be a PCIE 16x slot to 4x full slots at 4x speed.

bruce wrote:

ProDigit wrote:I wished there were some risers that would be able to split 16x slots in 2x 8x slots.

Yes, it would be nice. Is that even possible?

Not as easy as a passive solution. Splitting PCIe buses is what the PLX chips do ... but it often causes troubles, like discussed here : viewtopic.php?f=80&t=30682