Getting the most PPD out of your hardware for F@H

[TennesseeTony]

Any word on the street regarding how the 1080Ti performs at F@H tasks?

 

[Markfw]

I will let you know sometime after May 8th. I have a guaranteed preorder of the FTW3 at $750. Its suposed to ship then. I expect based on 3500 cuda cores to have well over 1.2 million

 

[StefanR5R]

With 100 % power target and when cooling is sufficient to keep up good clocks, 1080Ti outputs about 1.1 M PPD average. Perf/Watt is comparable to 1080 then.

With somewhat reduced power target, you can dial to a more effective point. I think I had it at about 0.9 M PPD at 70 %, or maybe even 65 %.

(F@H is currently off here because I need the CPU. I could possibly dig up logs and try to make some sense of them, but no promises. No spare time ahead.)

Edit:
Theoretically, 1080Ti and Titan X(P) should perform identically. OCN's PPD database currently lists the latter with 1.4 M PPD. But this is most definitely only from a few, rare, cherry-picked WUs, and far off from what can be achieved in the daily churn.

Edit 2:
Also, considering that I am still quite a n00b at F@H, it is possible that a more versed DCer can get more out of 1080Ti than I did so far.

 

[StefanR5R]

A pro pos 3500 cuda cores: Is it possible to run two F@H jobs simultaneously on the same card? Depending on the WU type, a single job utilizes the GPU more or less below full tilt. The bigger the card, the lower the average utilization it seems.

 

[Markfw]

A pro pos 3500 cuda cores: Is it possible to run two F@H jobs simultaneously on the same card? Depending on the WU type, a single job utilizes the GPU more or less below full tilt. The bigger the card, the lower the average utilization it seems.

Not that I know of. Only one task per card. BUT I can say that some units are better utilizes than others, as with the card running at the exact same GHZ and voltage, the temps can vary really high on some units

 

[Aurum]

It seems to me that disabling hyperthreading may actually slow the calcs down. The way I do it I open the Windows Task Manager Performance tab and watch the CPU Usage graphs. I usually find that I need to dedicate a thread per folding GPU plus two more. E.g., one folding rig (I believe you call them boxes here) with four 1070s on a Gigabyte GA-X99-SLI (x8x8x16x8 @3.0) with an E5-2603 v4 that has 6 cores and 6 threads, an M.2 SSD and G.SKILL Ripjaws5 Quad DDR4 3000. I require that total CPU Usage remain mostly at 85% or lower. With four 1060s it takes all 6 cores to do that with an E5-2603 v4 so no CPU folding.

On another rig with four 1070s on a Gigabyte GA-X99-SLI (x8x8x16x8 @3.0) with an E5-2673 v3 that has 12 cores and 24 threads, an M.2 SSD and G.SKILL Ripjaws5 Quad DDR4 3000. This rig can easily support four 1070s with 6 threads and use 18 threads for SMP client CPU folding.

On another rig with three 1070s on an ASRock Z97 Extreme4 with an i7-4771 with 4 cores and 8 threads and an 8 GB (666 MHz) DDR2 dual-kit. It only takes one thread to support three 1070s and 7 threads can do CPU folding.

My i3-4130T rigs are borderline capable of handling multi-GPU folding so no CPU folding.

For AMD rigs my Athlon II X4 640 can handle multi-GPU with no CPU but I only use it for legacy GPUs like 7970 and 7800. FX-6300 and FX-8350 can easily handle four 1070s plus CPU folding. I prefer Intel X99 motherboards since they do not use the Northbridge chip to switch the PCIe bus. I won't build any more AMD rigs since all PCIe busses are switched via the Northbridge chip.

My next rig is going to be a Gigabyte GA-X99-UD3P with an E5-4660 v3 (14/28), four 1080s, a 16 GB Quad-kit Corsair Vengenance DDR (2133) and an M.2 SSD.

 

[TennesseeTony]

Hello @Aurum ! Heck of a multi-set up you have there! I don't suppose you fold for TeAm Anandtech? I appreciate the insights. I've tried HT off and it didn't seem to help, so it's good to see some confirmation there, on the X99 chipset, at least.

 

[StefanR5R]

my four GPUs shall crunch some extra rounds in Folding@Home. The GTX 1080/ 1080Ti's may underperform though because their 4.0 GHz feeder CPU had to be replaced by a 2.6 GHz one.

Actually this 2620v3 seems to be doing quite fine as F@H triple GPU feeder. From what I remember, GPU core utilization and PPD were quite similar with the i7 before.

I am still debating whether I should get a replacement i7, but maybe I leave the Xeon in there and relegate this box to GPU-only projects. Even BOINC GPU projects with higher CPU demand than F@H might be OK if I run enough tasks per GPU in a staggered fashion.

 

[Markfw]

My E5-2683v3 has 2 video cards, and each one only takes 3.8% of the CPU, or basicly one cpu core. I have a G4400 system with one card and it takes more than one of the 2 cores, almost the whole box.

So your Xeon should be more than enough for those 3 cards.

 

[LightningZ71]

Hello all. Wonderful discussion you all have going here. I once folded a long time ago for a while, but got away from it (and SETI, and wcg...). I'm finally getting back to where I have some time to devote to it again. Currently, I've got some old phones and a few clunkers rotting into obsolescence while cranking out SETI packets and a phenom X6 doing WCG work. As the result of a recent project, I have received a handful of Intel I5-2300 boards with RAM, a bag full of Nvidia NVS 510 video cards (underclocked GT 630 2GB DDR3 128 bit), and a few spare drives and a case. I have taken that case and installed one of the motherboards, 16GB ram, and two of the nvs 510 video cards. I fully realize that the video cards are old and slow, but this isn't really meant to be a prime crunching rig.

I would like to use this as a "proof of concept" box for making a farm of curecoin miners. The boards are low end, with one pcie2x16 slot and one x4 slot and no overclocking, but they're free. The power is free to me for the next 6 months or so (solar array at work, currently out of production due to age and production related issues but still power certain outlets while it waits to be replaced. I have permission to use it at my own risk) and the internet is public wifi from next store. So, cost and environmental impact are not a concern.

So, here's the question: how best do I set this up? I plan to use linux (mint is my weapon of choice). I have no experience with coin mining or the wallets associated with it. What are the gotchas? Does anyone know any setup guides? Given how low end those cards are, can I get away with running a cpu process or two as well?

Any insight would be greatly appreciated.

 

[Markfw]

Hello all. Wonderful discussion you all have going here. I once folded a long time ago for a while, but got away from it (and SETI, and wcg...). I'm finally getting back to where I have some time to devote to it again. Currently, I've got some old phones and a few clunkers rotting into obsolescence while cranking out SETI packets and a phenom X6 doing WCG work. As the result of a recent project, I have received a handful of Intel I5-2300 boards with RAM, a bag full of Nvidia NVS 510 video cards (underclocked GT 630 2GB DDR3 128 bit), and a few spare drives and a case. I have taken that case and installed one of the motherboards, 16GB ram, and two of the nvs 510 video cards. I fully realize that the video cards are old and slow, but this isn't really meant to be a prime crunching rig.

I would like to use this as a "proof of concept" box for making a farm of curecoin miners. The boards are low end, with one pcie2x16 slot and one x4 slot and no overclocking, but they're free. The power is free to me for the next 6 months or so (solar array at work, currently out of production due to age and production related issues but still power certain outlets while it waits to be replaced. I have permission to use it at my own risk) and the internet is public wifi from next store. So, cost and environmental impact are not a concern.

So, here's the question: how best do I set this up? I plan to use linux (mint is my weapon of choice). I have no experience with coin mining or the wallets associated with it. What are the gotchas? Does anyone know any setup guides? Given how low end those cards are, can I get away with running a cpu process or two as well?

Any insight would be greatly appreciated.

This thread has nothing to do with curecoin, only FAH. Please start your own thread.

 

[LightningZ71]

My apologies if I have offended. From the curecoin.net website, it appeared that curecoin was entirely based on Folding@Home and the coins were just a secondary reward for the Folding work. It appears that I must have misread it.

 

[Markfw]

I don't know anything about curecoin.

That being said, purely from a folding perspective, one good video card requires a minimum of one fast, or 2 slower CPU's. For example, I had a G4400 and it was about all it could do, to keep up with a 980TI. If you are buying new cards, the GTX 1070 appears to be best bang/buck including power usage. But the 1080TI is a beast ! The same points as 2 1070's, over one million ppd.

 

[StefanR5R]

I have received a handful of Intel I5-2300 boards with RAM, a bag full of Nvidia NVS 510 video cards (underclocked GT 630 2GB DDR3 128 bit),

Folding@Home's credits are nonlinear: Amount of work is given credit too (like in almost all other distributed computing projects), but there is additional credit for how fast the work is done. Simplified, a fast card can be better than two slow ones.

Overclock.net has a list of points-per-day at F@H with various cards:
http://www.overclock.net/t/475163/gpu-projects-ppd-database
There is a GT 630 entry, but only with two samples (i.e. only from two work units, which is far too few for proper statistics because of high variability between F@H WUs). Still, the number given there is very small; a fast CPU might do a better job than this card.

 

[LightningZ71]

On overclocker's list, those aren't even representative of the NVS 510. The NVS510 uses a kepler chip, the GK106. There were three or four different versions of the GT 630 using two different chips and three different memory configurations. The closest thing on the list to the GK106 NVS510 on the list is the GTX 650ti, which is just about 4 times the performance of the 510. It still leaves it at about 5000 PPD (based on scaling down the 650ti's numbers) which is still very slow.

Again, the only purpose of this first system is to just prove that I can get the system up and running and figure out what I need to get to get more of them running (when it's all done, I should have about 6 or so systems all running at once.)

 

[StefanR5R]

@LightningZ71 - another solar powered distributed computing thread, but about off-grid solar: http://www.portvapes.co.uk/?id=Latest-exam-1Z0-876-Dumps&exid=threads/solar-anyone.2497083/

 

[StefanR5R]

On the question how much GPU performance in Folding@Home depends on single-threaded CPU performance:

GTX 1080 Ti, driver version 384:

• Linux, i7-7700K, 4.4 GHz:
  GPU utilization 94...95 %, 1.1...1.15 M PPD

• Linux, E5-2696v4, no load besides F@H, 3.4 GHz:
  GPU utilization 92...94 %, 1.1 M PPD

• Linux, E5-2696v4, WCG besides F@H (3C/6T set aside for F@H and OS), 2.8 GHz:
  GPU utilization 88...93 %, 1.05 M PPD

• Windows 7, E5-2696v4, 2.8...3.x GHz:
  GPU utilization ~75 % (rarely ~80 %), 0.85...0.90 M PPD

Conclusion: CPU performance does matter for GPU performance in F@H, but not as much as I initially thought. The influence of the operating system is much higher than the influence of CPU performance.

 

[Markfw]

On the question how much GPU performance in Folding@Home depends on single-threaded CPU performance:

GTX 1080 Ti, driver version 384:

• Linux, i7-7700K, 4.4 GHz:
  GPU utilization 94...95 %, 1.1...1.15 M PPD

• Linux, E5-2696v4, no load besides F@H, 3.4 GHz:
  GPU utilization 92...94 %, 1.1 M PPD

• Linux, E5-2696v4, WCG besides F@H (3C/6T set aside for F@H and OS), 2.8 GHz:
  GPU utilization 88...93 %, 1.05 M PPD

• Windows 7, E5-2696v4, 2.8...3.x GHz:
  GPU utilization ~75 % (rarely ~80 %), 0.85...0.90 M PPD

Conclusion: CPU performance does matter for GPU performance in F@H, but not as much as I initially thought. The influence of the operating system is much higher than the influence of CPU performance.

First, what linux tool can you use to see GPU temps, and utilization% ?

 

[TennesseeTony]

Psensor

 

[StefanR5R]

What Tony said. But besides psensor, I also use something like

nvidia-smi dmon -d 120​

in a console window. This shows power usage, temperature, core utilization, memory utilization, and clocks, every 120 seconds. See "nvidia-smi -h" and "nvidia-smi dmon -h" for more options.

(BTW, one of the first extra packages which I installed on Mint was Openssh-server. That way I can use an ssh client on my main PC to log into a command line on a DC worker PC, for example to run tools like nvidia-smi remotely.)

 

[Aurum]

(...extra packages which I installed on Mint was Openssh-server. That way I can use an ssh client on my main PC to log into a command line on a DC worker PC, for example to run tools like nvidia-smi remotely.)

I've heard Linux folds faster than Win7, guess they weren't kidding. I'm an old geezer who'd like to maximize my folding but can an old dog learn the new tricks of Linux? My folding rigs are headless & Win7 TightVNC makes that easy to do. Is there a good book or resource I can use to break the Linux ice??? TIA

 

[Howdy]

I've heard Linux folds faster than Win7, guess they weren't kidding. I'm an old geezer who'd like to maximize my folding but can an old dog learn the new tricks of Linux? My folding rigs are headless & Win7 TightVNC makes that easy to do. Is there a good book or resource I can use to break the Linux ice??? TIA

I too am an old geezer............... If I can manage to do it pretty sure you can too. Look HERE for install guide, follow to the letter and you should be good to go.

 

[Markfw]

I too am an old geezer............... If I can manage to do it pretty sure you can too. Look HERE for install guide, follow to the letter and you should be good to go.

Well that was close, but I had to tweak it to work for me. Check out my procedure, based ont that thread in my "help with linux" thread.

 

[biodoc]

What Tony said. But besides psensor, I also use something like

nvidia-smi dmon -d 120​

in a console window. This shows power usage, temperature, core utilization, memory utilization, and clocks, every 120 seconds. See "nvidia-smi -h" and "nvidia-smi dmon -h" for more options.

(BTW, one of the first extra packages which I installed on Mint was Openssh-server. That way I can use an ssh client on my main PC to log into a command line on a DC worker PC, for example to run tools like nvidia-smi remotely.)

nvidia-settings is a gui app that gives you similar information along with the ability to to control fan speed and overclocking functions as well.

 

[TennesseeTony]

What team are with, @Aurum ? We are glad to help out and share knowledge with anyone, to advance science, but just curious. And if you aren't already affiliated with a team for your crunching/folding, we'd love to have you! Folding team is 198, BOINC is TeAm AnandTech.