FAHBench, Turing

[porina]

I restarted folding, and am digging further to optimise my resources and how best to apply them. My optimisation will be leaning towards PPD/watt as I have to limit my heat output, and by implication, power consumption.

 

I just noticed there is FAHbench. Can't run it right now, but general question to the folding community, how useful a measure is FAHbench compared to "real" folding work? I'd like to compare not only how my specific GPUs compare in performance, but also what if I over/underclock them, and its impact on power consumption.

 

I also saw hints that Turing might be highly efficient at folding. Been holding out for 2080Ti availability, but I do note that 2070 is comparable in pricing to 1080, so could get one to play with while I wait...

 

Yeah, when I do things... I really do things. Rough calculation suggests it would take me under 2 weeks running hard to get into team top 100, if I put all current GPUs on it  November might get interesting...

[GOTSpectrum]

FAHBench is great because it actually runs real work units through your cards it just doesn't submit them I believe. 

[porina]

1 hour ago, Ben Quigley said:

FAHBench is great because it actually runs real work units through your cards it just doesn't submit them I believe. 

Ok, sounds like I have some testing to do this weekend!

[BuckGup]

Why would turing be any better? It performs the same if not worse than Pascal considering the gains from a dye shrink

[porina]

15 minutes ago, BuckGup said:

Why would turing be any better? It performs the same if not worse than Pascal considering the gains from a dye shrink

I saw hints elsewhere they were more energy efficient for a given performance. Ordered a 2070 to play with and will test myself this weekend. At the least you'd hope to gain something from the shrink, only performing worse if you run it closer to the limit.

[BuckGup]

2 hours ago, porina said:

I saw hints elsewhere they were more energy efficient for a given performance. Ordered a 2070 to play with and will test myself this weekend. At the least you'd hope to gain something from the shrink, only performing worse if you run it closer to the limit.

Maybe if you can turn off RT and tensor cores as the tdp is higher than pascal 

[porina]

18 minutes ago, BuckGup said:

Maybe if you can turn off RT and tensor cores as the tdp is higher than pascal 

2070 175W (FE 185W), 1080 180W

2080 215W (FE 225W), 1080Ti 250W

2080Ti 250W (FE 260W)

 

So comparing ball park gaming performance it's similar or lower power. 

 

Besides, folding is not necessarily the same as gaming, which is what I will be testing.

[Gorgon]

FAHBench works great for testing stability and power consumption but does not output in Points per Day but in a measure relevant to biologists. Make sure to use a current 11720 or such MM_21 WU.  Essentially while folding a current WU copy the relevant folder to your FAHBench folder then create a JSON file and tell it to use that.

 

Anand Tech actually provides numbers for the GTX and RTX cards running FAHBench in their reviews and by this you can see that the RTX cards are, as you say, not particularly better at folding using the Single Precision that F@H uses but may be more power efficient. From a purely PPD/$ perspective the RTX cards are way worse values though but this will depend on local market pricing. Currently in Canada the best value for folding is the 1070ti.

 

My own testing using FAHBench on EVGA 1060 6GB and 1070 ti Blacks show that power limiting the cards to 90W and 110W respectively results in a 15% decrease in PPD for a 40% decrease in power consumption. If your concerned about power consumption and heat load then also moving to a Gold or higher efficiency power supply can also show results.

 

While I find most power sizing calculators are way too conservative, running in the 30 to 70% utilization level for a power supply puts you in the highest efficiency sweet spot. High efficiency folding with the 1070ti and 1060 I’m using 210W and full tilt 370W measured at the wall using a Kill-a-Watt but the CPU in my main folding rig is a Pentium G5400 with a 65W TDP. I currently have a 750W Gold power supply but a 650W would definitely work. A 500W would likely work but might not handle transient load requirements.

 

The hardest part of setting up a dual card dedicated folding rig was getting consistently good airflow to the GPUs to keep them cool and running at higher clock speeds. I’m using fractal Define R4 and S chassis with 6 or seven 140mm Noctua iPPC2000 and 3000 PWM fans and have the fan curves for most of the fans controlled by the temperature sensor on the upper PCIe3x16 slot running about 30% PWM at 43C to 70% at 46C but these need to be determined empirically and some headroom left for fluctuations in ambient temperature.

[Gorgon]

FAHbench will also crash just like the client if your GPU or memory offset is unstable

[porina]

4 minutes ago, Gorgon said:

FAHBench works great for testing stability and power consumption but does not output in Points per Day but in a measure relevant to biologists. Make sure to use a current 11720 or such MM_21 WU.  Essentially while folding a current WU copy the relevant folder to your FAHBench folder then create a JSON file and tell it to use that.

I don't actually need the ppd, as long as I have a good relative measure of a card's performance. I've been doing some initial testing with the default settings and things are pretty much unexciting so far. I'll have to see if I can figure out the custom WU later.

4 minutes ago, Gorgon said:

From a purely PPD/$ perspective the RTX cards are way worse values though but this will depend on local market pricing. Currently in Canada the best value for folding is the 1070ti.

I don't care about up front costs as I'm not buying for FAH, but multiple uses. What I care about if I'm going to run it is PPD/W. It isn't cool enough yet for me to hit this hard so I'm limiting my power output.

 

4 minutes ago, Gorgon said:

My own testing using FAHBench on EVGA 1060 6GB and 1070 ti Blacks show that power limiting the cards to 90W and 110W respectively results in a 15% decrease in PPD for a 40% decrease in power consumption. If your concerned about power consumption and heat load then also moving to a Gold or higher efficiency power supply can also show results.

One of my hopes were, by setting power limit, could I get power usage down more than the output goes down. If so, I think overall I'd be better off running more GPUs at reduced power, than a few at full power. That's before taking into consideration rest of system power. I've no plans to change PSUs.

4 minutes ago, Gorgon said:

While I find most power sizing calculators are way too conservative, running in the 30 to 70% utilization level for a power supply puts you in the highest efficiency sweet spot. High efficiency folding with the 1070ti and 1060 I’m using 210W and full tilt 370W measured at the wall using a Kill-a-Watt but the CPU in my main folding rig is a Pentium G5400 with a 65W TDP. I currently have a 750W Gold power supply but a 650W would definitely work. A 500W would likely work but might not handle transient load requirements.

I've ran a 980Ti + 6700k off a 450W PSU before! I might be close to the rated limit... similar again, with a 550W PSU, two RX 580's and another 6700k. Fortunately I'm in a higher voltage part of the world and gain a small increase in efficiency from that. 

 

4 minutes ago, Gorgon said:

The hardest part of setting up a dual card dedicated folding rig was getting consistently good airflow to the GPUs to keep them cool and running at higher clock speeds.

Had similar problem when testing SLI. I wouldn't consider my setup to be used long term in this way. At least I intend to participate in the November event, and get some data leading up before that. Two GPUs per system, especially ones with larger mobos, might help a bit.

[GOTSpectrum]

If you are going for max PPD/w then ignore CPU folding all togeter, it burns 90+watts for very limited PPD with most CPUs. 

[Gorgon]

2 minutes ago, porina said:

One of my hopes were, by setting power limit, could I get power usage down more than the output goes down. If so, I think overall I'd be better off running more GPUs at reduced power, than a few at full power. That's before taking into consideration rest of system power.

Yes you can. Run FAH bench at the maximum Power limit then the minimum and compare the results. I see 800-900 thousand PPD on the 1070ti running full tilt and 700 to 800 thousand at 10W above minimum

[porina]

1 hour ago, Ben Quigley said:

If you are going for max PPD/w then ignore CPU folding all togeter, it burns 90+watts for very limited PPD with most CPUs. 

Already turned off. I can't recall the exact numbers, but a 6600k on 3 cores was like 1% the PPD of the 1070 in the same system.

1 hour ago, Gorgon said:

Yes you can. Run FAH bench at the maximum Power limit then the minimum and compare the results. I see 800-900 thousand PPD on the 1070ti running full tilt and 700 to 800 thousand at 10W above minimum

I already tried 80% on nvidia cards, and -20% on AMD ones. There was a small drop to PPD, but I haven't measured power at all yet. Also, on the 1070, I did wonder if I should try lower, since even when I was at 100% PL, the average was somewhat below 100% anyway. So capping at 80% isn't 80% of running at 100%, if that makes sense. In GPU-Z the power limit was only flicking on now and then, not constantly.

[porina]

Done some testing on 2070, with the default FAHBench settings... it is interesting. Following are at stock setting (including factory OC)

 

135 1080Ti Asus OC

111 2070 Gigabyte Windforce

74 1070 Zotac FE

73 980Ti Gigabyte Windforce

57 RX 580 Gigabyte Windforce

 

Yeah, I have a lot of windforce cards, they seem to be amongst the cheapest without being MSI.

 

System power was around 175W with the 2070, compared to 45W idle. Now here is where it gets weird. I fire up MSI Afterburner, set it to min allowed power limit of 71%. Rerun the bench. Score and power went up! Saw up to 120 and over 180W power draw. Went back to 100% PL, got original result. Used Gigabyte's too, same thing happened with MSI afterburner. Monitored clocks were the same regardless of power setting.

 

Also tried the OC scanner on Gigabyte's tool. Crashed the system. Currently trying a manual OC.

[Gorgon]

9 hours ago, porina said:

I did wonder if I should try lower, since even when I was at 100% PL, the average was somewhat below 100% anyway. So capping at 80% isn't 80% of running at 100%, if that makes sense. In GPU-Z the power limit was only flicking on now and then, not constantly.

Sorry, I use Linux on the command line so I meant 80w on a 1060 or 100W on a 1070ti not 80% on a slider in afterburner. I found my sweet spot to be about 10w over the minimum power.

 

Nice to have the 2070 to play with.  I was hoping the 1070s would drop in price with its launch but here in Canada they went up!  The EVGA Black was $534 last week and is now $649. If I were suspicious I would think that they’re trying to make the 2070 look more attractive.

 

Oh well, I’ll just have to wait for Black Friday and see if I can another 1070ti for $499. After all, I do have to give @Ben Quigleya run for his money 

[porina]

5 hours ago, Gorgon said:

Sorry, I use Linux on the command line so I meant 80w on a 1060 or 100W on a 1070ti not 80% on a slider in afterburner. I found my sweet spot to be about 10w over the minimum power.

I noticed you used W not %, but it is easier for me to think in %.

 

5 hours ago, Gorgon said:

Nice to have the 2070 to play with.  I was hoping the 1070s would drop in price with its launch but here in Canada they went up!  The EVGA Black was $534 last week and is now $649. If I were suspicious I would think that they’re trying to make the 2070 look more attractive.

Sellers don't want stock to sit around forever so keep an eye on what pricing does. Usually if there is a rise then they are the last few units they try to sell to people who have to have a specific model.

 

5 hours ago, Gorgon said:

Oh well, I’ll just have to wait for Black Friday and see if I can another 1070ti for $499. After all, I do have to give @Ben Quigleya run for his money 

I'm looking to move up the team rankings too... if I went flat out I think top 10 in a month would be not out of the question, but I think my house would melt if I tried that...