GPU CPU bound?

[[email protected]]

Hi,

I am runnng a GTX1070 on a low end MSI H110 mobo with a Pentium G4400 and Windows 10, running only a single GPU slot for a couple of months now. I never get above 600k PPD, even if I increase the clockspeed of the GPU. The CPU is running at 53% on FacCore_21, probably meaning that one core is fully occupied. Is it fair to conclude that my system is CPU bound, and that with a faster CPU I would get more PPD out of the GPU?

Any suggestions on how to get to 700k PPD, what seems to be more average for a GTX1070?

[Nathan_P]

The long term average for a 1070 is around 600k, I've just checked mine and its currently showing 512k PPD on a 9839WU, when I came in this afternoon it was showing 665k on a different project, that's on a PCIe 3.0 x16 slot running heavily overclocked on a 4c/8t cpu at a much lower clockspeed than yours.

Whilst you may think you are cpu bound, you aren't. It would be a waste of money to upgrade to a better cpu for running just one card.

[bruce]

FAHCore_21 on windows will use 100% of a single core, so you can expect to see 53% whether or not that's limiting GPU performance. The driver is written to use a spin-wait so there's no way to judge how much capacity is available. While it's possible that you are CPU bound, it's not too likely since the main function of the CPU task is to move data to/from the GPU for processing there and then wait for the next block of data to be ready to be moved.

[[email protected]]

The long term average for a 1070 is around 600k, I've just checked mine and its currently showing 512k PPD on a 9839WU, when I came in this afternoon it was showing 665k on a different project, that's on a PCIe 3.0 x16 slot running heavily overclocked on a 4c/8t cpu at a much lower clockspeed than yours.

Whilst you may think you are cpu bound, you aren't. It would be a waste of money to upgrade to a better cpu for running just one card.

600k? That sounds like mine. My reason for asking was that I saw much higher numbers in https://docs.google.com/spreadsheets/d/ ... tput=html# Any idea how these people get these high numbers? I thought it was ok taking the cheapest socket 1151 CPU around, glad to see that you can confirm that.

[rwh202]

Linux might give you 10% boost over windows. Serious overclocking and selective reporting of PPD also come into play, so take the spreadsheet with a grain of salt!

I also use G4400 CPUs - driving pairs of 1080 and 1080 ti - A single 1070 won't be bottlenecked.

[foldy]

I made an experiment and limited my CPU i5-2500k from 3.3Ghz down to 1.6Ghz. That hurt my gtx 1080ti PPD. But going up to 4.4Ghz did not make a difference over 3.3Ghz.

So maybe a slow and old gen CPU with < 2Ghz may limit a fast GPU? (But the G4400 with 3.3Ghz is fine)

[[email protected]]

Linux might give you 10% boost over windows. Serious overclocking and selective reporting of PPD also come into play, so take the spreadsheet with a grain of salt!

I also use G4400 CPUs - driving pairs of 1080 and 1080 ti - A single 1070 won't be bottlenecked.

Running Linux may be worth a try. Any preference for a particular Linux distribution?

[Nathan_P]

I use Linux mint 18.2 mate edition on all my folding rigs - got everything set up in a couple of hours - That 600k is on Linux. I don't push my clocks any further than what the card wants to do on its boost clocks so my guess is that those 700k PPD are on specific projects that can give a reported 700k PPD in hfm and a balls to the wall overclock.

I've checked a couple of those project numbers and my 1070 is showing slightly worse numbers, but then i'm not pushing an extreme overclock on my card

[[email protected]]

Cranking up the base clock on the GPU with 100MHz already seems to deliver an extra 5%. Maybe I will try some more serious overclocking, as mentioned here earlier, and see what comes out of that. This little extra already proves though that the CPU is indeed not a bottleneck.

[Kuno]

Linux is the way to go as there is literally no CPU bound limits in Linux as there are in Windows. Windows is a hog when it comes to folding, and Linux, well it's apparent you could have your cards in a toaster and they will fold at full speed. I've seen reports of a 1080 in an a6-5200 APU under linux getting well over 900k ppd, where as in windows it was only doing 400k. If you think your processor is limiting you, install Mint on a usb drive and run it that way for a day or so and compare your scores.

[Duce H_K_]

Well actually I was searching the CPU requirement for GPU folding (NV) topic but I guess it's allowed to write here. I run folding on Windows 8.1 use Single GPU GTX1070@2100MHZ chip & 9000MHZ memory clocks. Driver 372.54. So my observation is when running FAHCoreA4/A7 (CPU) make GPU fold a bit slower. Even if I take away 2-3-4 cores from CPU slot for FAHCore21.exe process with WinAFC Utility

[foldy]

Maybe that is a heat issue, GPU downclocking when CPU runs all cores? Does it make a difference if you keep one CPU core idle? Or set FahCore_21 priority to high?

[Nathan_P]

What CPU?

[bruce]

What CPU?

The CPU that's processing one of the FAHCore_a* tasks -- which if one thread per GPU is left idle, is whichever one is NOT processing other things.

[Nathan_P]

What CPU?

The CPU that's processing one of the FAHCore_a* tasks -- which if one thread per GPU is left idle, is whichever one is NOT processing other things.

I meant what make/model of cpu is he using