It seems that a lot of GPU problems revolve around specific versions of drivers. Though NVidia has their own support structure, you can often learn from information reported by others who fold.
I've put 2 GTX960's in an old S939 motherboard with an Athlon X2 3800+. The PSU is a new one, more than capable of powering 3 high-end GPU's. The motherboard only has PCIe 1.1 slots and with 2 GPU's, can only provide 8 lanes to each.
The GPU's are both at stock setting, and have been getting approx 100kPPD each on P9411. I'm using the latest 347.52 drivers.
The GPU load hits a maximum of 85% and generally runs at 77-80%. There are no CPU slots configured. The CPU load is 100% for both cores. Both GPU's have completed all P9411 WU's without any issues
Its a clean install of Win7 pro. I think the PPD is a little low compared to the figures I've seen other users get. How likely is it that the CPU or PCIe 1.1 interface is holding the GPU's back?
I don't think the CPU has much effect. It seems to sit in a polling loop, so 100% usage for not much real work. The PCIE could well slow things down marginally, since all the data transferred on and off the GPU has to go that way. I have no figures as to what the impact is, and isolating one factor from the other could be tricky. (Is it possible, with a faster PCIE motherboard, to configure it to work in a legacy mode to do a direct comparison with full bandwidth?)
1. As long as you have idle CPU cycles when the GPUs are folding you won't be held back. In your case it seems you're CPU limited if the CPU is 100% utilized because of the GPU slots. How much I don't know - you should check what is the normal for 960 / this project when not CPU bound.
2. Now provided 1) is resolved, to what extent a faster CPU would influence GPU folding - I don't know, I'm guessing any difference would be marginal.
3. My 970 also doesn't exceed 90% GPU utilisation, but I am told this is because it's held back by shaders.
Windows 11 x64 / 5800X@5Ghz / 32GB DDR4 3800 CL14 / 4090 FE / Creative Titanium HD / Sennheiser 650 / PSU Corsair AX1200i
Breach wrote:1. As long as you have idle CPU cycles when the GPUs are folding you won't be held back.
2. Now, to to what extent a faster CPU would influence GPU folding - I don't know, I'm guessing any difference would be marginal.
3. My 970 also doesn't exceed 90% GPU utilisation, but I am told this is because it's held back by shaders.
Both CPU cores are showing as 100% load, and that is constant. I'm also using an R9 270X on a old Core2Quad. The CPU core that feeds the GPU is showing a usage of between 5-10%, which is what I expected to see from the Athlon cores.
I would be happy to see 90% load on the GPU's, but it very rarely goes over 80%.
Oh well, I didn't really think the CPU and PCIe interface would make much difference, it's just the figures (ppd, cpu load, gpu load) don't seem to be quite right
Yes, I edited my post, as you had already stated that CPU usage is 100%. You can try stopping one slot and see what's the CPU utilisation in that case. It would give you some relative idea of how much performance you are losing when running both cards.
You can try to overclock the CPU to compensate.
Windows 11 x64 / 5800X@5Ghz / 32GB DDR4 3800 CL14 / 4090 FE / Creative Titanium HD / Sennheiser 650 / PSU Corsair AX1200i
Yes, I've seen that thread - it's one of the reasons I questioned my PPD. I don't think the older PCIe interface has any impact, and an Athlon X2 3800+ should be more than capable of feeding the GPU's - but something isnt quite right.
I've paused 1 of the GPU's now and the CPU load has changed from 100%/100% to ~25%/~85%, with the 85% being the core feeding the active GPU. I'll keep an eye of the TPF and see if it makes any difference.
For reference my 3770k @ 4.6Ghz (quad-core w/HT on) utilizes 13% CPU for maintaining a single 970 slot (core 17). Looking at this: https://cpubenchmark.net/cpu.php?cpu=AM ... 2B&id=1511
I think there's nothing wrong, that CPU is just old (2005).
Windows 11 x64 / 5800X@5Ghz / 32GB DDR4 3800 CL14 / 4090 FE / Creative Titanium HD / Sennheiser 650 / PSU Corsair AX1200i
Breach wrote:For reference my 3770k @ 4.6Ghz (quad-core w/HT on) utilizes 13% CPU for maintaining a single 970 slot (core 17)....
That is, probably, one thread at 100%. That is normal for Core 17 (and, I think, current Core 18) on nVidia GPUs. A separate thread is recommended (and reserved by the automatic install) for each nVidia GPU. This thread runs at 100%, but as I said above, actually does very little between data transfers.
Breach wrote:For reference my 3770k @ 4.6Ghz (quad-core w/HT on) utilizes 13% CPU for maintaining a single 970 slot (core 17)....
That is, probably, one thread at 100%. That is normal for Core 17 (and, I think, current Core 18) on nVidia GPUs. A separate thread is recommended (and reserved by the automatic install) for each nVidia GPU. This thread runs at 100%, but as I said above, actually does very little between data transfers.
I can confirm that. The CPU is at 100% load but it isnt generating much heat. Is the 100% thread usage particular to Nvidia? The ATI GPU running on a C2Q only results in a CPU load of ~10%
Breach wrote:For reference my 3770k @ 4.6Ghz (quad-core w/HT on) utilizes 13% CPU for maintaining a single 970 slot (core 17)....
That is, probably, one thread at 100%. That is normal for Core 17 (and, I think, current Core 18) on nVidia GPUs. A separate thread is recommended (and reserved by the automatic install) for each nVidia GPU. This thread runs at 100%, but as I said above, actually does very little between data transfers.
So you're essentially saying that regardless of clock speeds one core (or logical processor) will always be running at 100% doing pretty much nothing but polling? Cool, I didn't know that, so I stand corrected (though obviously why this is like this escapes me, isn't this a waste of CPU cycles...?).
In this case PCI Express limits indeed seem likely - this board is PCI Express 1.0 only, right? This would already bottleneck a single card I believe. Edit: OK, reading some more, maybe not a single one at x16, but definitely 2 cards at PCI Express 1.x at x8 would be affected. Easy to test too - just leave one card in the 16x slot and see whether PPD improves.
Windows 11 x64 / 5800X@5Ghz / 32GB DDR4 3800 CL14 / 4090 FE / Creative Titanium HD / Sennheiser 650 / PSU Corsair AX1200i
Tried same thing with S939 and a Dual core opteron and a MSI 970. Had similar results. Move 970 to an I5 system and got 97% usage in Ubuntu and 90 to 95% in Win7. Upgraded to MSI Ax88m motherboard with APU. Now running at 97% and 62c. Win7 shows 93 to 95% usage. The old motherboard was definitely slowing me down. This is a folding machine only. Recommend Ubuntu and latest Nvidea drivers. Folding 3 cores on APU and 1 core for gpu. Hope this helps
wilding2004 wrote:I've paused 1 of the GPU's now and the CPU load has changed from 100%/100% to ~25%/~85%, with the 85% being the core feeding the active GPU. I'll keep an eye of the TPF and see if it makes any difference.
What I would try next is setting affinity for the thread that's running the fahcore, leaving everything else to use the other CPU. Does that change ~25%/~85% to ~25%/~100% or to ~15%/~100% or does it stay at ~25%/~85%? If all testing is done on the same WU, did the frame time change or is it still the same as it was with both slots running?
My final guess would be that 8 lanes isn't enough.
Very relevant question.
I have been paying attention to this topic during my period of time folding GPU mostly. Both the CPU and the PCIe have some impact on the PPD.
Nvidia is the main concern regarding CPU performance. And old CPU will decrease your PPD on GTX 970 or 980 by far.
I just proved this myself using 2 x 980 on a X79 motherboard. When disable PCIe 3.0 in bios, I got 690k PPD with two 980. When enabling PCIe 3.0 I got 740k PPD with same CPU clock and same Core Clock on GPU.
The numbers will not be the same if you use a GPU weaker than 980, or if you use AMD GPU.
When it comes to CPU clock, it seems 3,1Ghz is to low when we talk about the potential of GTX 970. If you get past the 3,1Ghz you will earn about 10k PPD on each GPU in a combo using 2 x 970.
I have tested this on 1150 platform and PCIe 3.0 x8/8
When it comes to i7 4 core 4 treads and 13% load, my theory is 100/8 = 12,5%
In other words 13% is one core of the i7 100% use.
Each GPU needs 1 core, and treads are not supported. If folding 2 GPUs, you can fold 5 cores CPU (i7), but not 6 (V7.4.4 sees 8 cores (4 core and 4 treads) ).
And if you fold 3 GPUs you can fold 3 cores and not 4 and so on (if X79 and 4280k).
Why Nvidia and AMD is so different I do not know. But this is what I have learned trough my time of folding.
