Performance differences on the same video card in F@H

[Markfw]

So we did a test a while back. Tenneseetony was getting over 600k ppd on his 1070s using a 5820k CPU. I thought that the CPU would have little effect. I got one and put it in a E8400 dual-core system. Getting on average 100k less on units. The CPU was almost maxed with no CPU client running. So I got a quad-core Xeon with a 12 meg cache and thought that would fix it. NOT !!

So then after more thought, I realized that the system was 1) DDR2, 2) PCI-e 1.0 and possibly other platform problems. So I got an 860k AMD quad-core, a motherboard with DDR3, and PCI-e 3.0, figuring that must fix it.

So far I have not seen over 500k,c unit 9211 that Tony had 2 units and averaged over 570k ppd.

So what have I missed ? what could be causing this major PPD difference on the same unit, with similar technologies ? And memory is not even close to maxed, or CPU usage. And NOW the platform is the same !

Edit: a few more stats. CPU utilization at 75% pretty constant. Memory at 2.5 gig of 4 used. Video card is a 1070 SC using EVA precision, stock 1594 mhz, but according to EVAPrecision, running consistantly over 1900 mhz. And NO other processes running, and this is a fresh win7 pro 64 bit install (today)

current 488k on a 9211, and Tony was at 570k

 

[crashtech]

I'm confused, are the platforms the same, or is one a 5820K and the other an 860K? Is he GPU only, or letting some CPU cores in on the fun?

 

[Markfw]

OK, His is a 5820k platform, and he allowed at least 4 threads to the GPU (I saw a screen shot and no CPU client, but Tony can confirm). Mine is an AMD 860 platform and NO CPU is running. The platforms are different, but the technologies are basically the same.

Edit: I DO have a CPU client running 2 cores, but the total util is only 75% max, so should not be affecting. I will pause it just in case

 

[Markfw]

So I paused the CPU client, and noticed in task manager that fahcore_21 was pegged at 25% (one core win 7) but on my 3920k box, its only at 8.5% (win 10) Could the windows going from 7 to 10 be the difference ?

 

[crashtech]

The scheduler is said to be improved starting in Win8, but I am not conversant in the details.

 

[Markfw]

OK< upgrade to win 10, and new nvidia win 10 drivers done (373.06, per the problems we recently had with the most current).

Will post again after I get a unit with history to compare. Tony never did a 9179. 1 minute 5 seconds, estimated almost 600k, but I lost 3-4 hours for the upgrade, so next unit will be a valid test if we have history. If this fixes it, all my boxes will get the upgrade.

edit: and its using 28-30% cpu, so more used than before. before on either the Xeon or this cpu in win 7, was maxed at 25%

 

[TennesseeTony]

I too have noticed the 100% usage of a core/thread, and have concluded that FAH GPU tasks are CPU bound. EDIT: and they are certainly PCIe bandwidth bound too, they run VERY poorly at 1x PCIe v2.0, and are still somewhat bandwidth bound at 4x if I remember correctly. 8x PCIe is ok though.

As for the differences in CPU utilization percentages, your 3920k has 8 threads, so 100% divided by 8 threads equals 12.5% for each thread, your AMD 860k has 4 cores, so 100% /4 = 25%.

 

[Markfw]

3930k has 12 threads. but notice the update,,,, Win 10 uses MORE than one core ! 28-30% on my 860.

my 3930 is using 8.3% for core21 which is one whole core, but they are stronger cores than an 860 AMD

 

[crashtech]

I have Celeron G1620 with a single GTX 1060 (2.7GHz Sandy) that seems to put out similar ppd to the same card but an i3-4170 (3.8GHz Haswell). Neither do any compute on the CPU, it's just there to support the GPU. Shouldn't I be seeing some difference between the two if F@H is CPU bound? I'm trying to understand this better so I can produce more. Neither max the CPU, but they do hang out over 50% almost all the time.

 

[TennesseeTony]

I've got an experiment for you Mark. Do this on one of the i7 systems with 980Ti card(s). Make sure this is one of the systems you already have a lot of data on.

• Reboot and enter the BIOS, and disable hyper threading. You'll have to finish/pause/remove any CPU task first.
  
• Keep track of the results of that system for a few days to see if it is CPU bound.

One core is nearly twice as powerful as one thread in many cases. Should be pretty evident (higher scores) if you currently don't have enough CPU feeding the cards.

 

[TennesseeTony]

Peeps always posting relevant stuff while I'm composing my replies, lol. If it were CPU bound crashtech, your CPU's are different enough to prove it.

The 1060 is an impressive beast in it's own right, but I do wonder if it's possible that with the slightly diminished output of a 1060, the CPU can keep up?

*******************************

Mark, might as well try the no hyper-threading on the Skylake too, as the difference should be more substantial on the 1080's if it's CPU bound.

 

[crashtech]

The 1060's seem to be a good value. Maybe the 1070 is better, but the 1060 is an easier pill to swallow and I had a couple spare mobos laying around. Perhaps it's for the best that I didn't spring for the 1070 after all, though I am considering one for the 5820K rig.

 

[Markfw]

I want to finish the 1070 experiment first. Then I can play with the 1080;s

 

[Markfw]

OK, so I installed win 10 on the dual 1080 box, and turned off hyperthreading and paused the CPU client. So, in the morning I should have 2 boxes to compare.

Edit: first real comparison on the 860. unit 13500, Tony got 670k consistently, I am getting 650k. Now we are talking. Can't wait to see what the 1080 box changes do !!!

 

[Markfw]

Well, it looks like disabling hyper-threading helps, and win 10 helps. I just need to turn off all CPU client, instead of pause. I got 100k ppd out of each box !

 

[TennesseeTony]

Is that 100k EXTRA? Woot!!!

 

[Markfw]

Yes, the 1070 is within 5% of your ppd now, and the 1080's have gotten better on a few units. working on updating the 980ti boxes. I am not touching the 3930k, as its working good.. I have one 4790k, I didn't update, but I did remove the cpu client and turned off HT.

My average was 6.3m, and after all this, I might hit 7m average !

 

[Markfw]

OK, only 4 of 9 done(not doing one of them), but I just saw 7,320,158 ppd on HFM. The plan seems to be working

 

[Markfw]

Well, the farm is upgrade, modified and ready. I created another thread so people could use what I learned to maximize their hardware. Averge ppd went from 6.3m to about 6.6m. Not sure until a week goes by. I have one card (a 1080) that is doing almost 1 million ppd all by itself !