Guru3D Folding@Home Information/Help Thread {Start Here New Users}

[Athlonite]

No config change except for driver everything else is the same infact I installed over the top of 7.1.48 and everything worked fine

but going through my F@H log I counted 68 WU's completed and returned but EOC and Stamford was only showing I had completed 3 (it's now at 5) but thats a long way off of what I'd counted in the log

it shows the only errors where some clock skew error:
13:01:22:WARNING:WU00:FS02etected clock skew, adjusting time estimates

[k1net1cs]

Are you sure those other WUs were finished before its expiration deadline?

In any case, if possible, upload the log file which contains these uncounted finished WUs.
Don't forget to redact the user name within the log file in case you want to.
(Notepad in Win 7 should be able to do just that)
Also, you might want to report this in F@H Forum.

[Athlonite]

here's the log it may not be complete but you'll get the idea

https://skydrive.live.com/redir.aspx...59A9CD9BB7!136

EOC now says I have done 7 WU's


This this for the SMP WU's

02:23:51:Unit 02: 32.27%
02:23:57:Unit 02: 49.43%
02:24:03:Unit 02: 66.60%
02:24:09:Unit 02: 83.76%
02:24:15:Unit 02: Upload complete
02:24:15:Server responded WORK_ACK (400)
02:24:15:Final credit estimate, 793.00 points

and for the GPU WU's I see this

07:54:24:Unit 01: 15.17%
07:54:30:Unit 01: 23.59%
07:54:36:Unit 01: 32.00%
07:54:42:Unit 01: 40.24%
07:54:48:Unit 01: 48.49%
07:54:54:Unit 01: 56.90%
07:55:00:Unit 01: 65.31%
07:55:06:Unit 01: 73.07%
07:55:12:Unit 01: 81.48%
07:55:18:Unit 01: 89.89%
07:55:24:Unit 01: 98.14%
07:55:26:Unit 01: Upload complete
07:55:26:Server responded WORK_ACK (400)
07:55:26:Final credit estimate, 1835.00 points

[Psychlone]

I think I need a little help from PanterX or someone else that can walk me through some optimization.

I currently am just running the SMP client on my i7 3820, and am pulling roughly 18,000 to 20,000 PPD, but I'm wondering if there's a little more that I could do on my own rig here.
I've already tried the GPU client for AMD, but the wattage used (~450W constantly) weighed heavily compared to the extra 10,000 PPD.
I just need to know if there's anything more that I can do on my own rig.



Psychlone

[k1net1cs]

Aside from OC-ing your CPU more or getting an extreme part? Unlikely.
Of course, there's also the option of getting another system which could be identical only on processor alone, and getting non-enthusiast parts for the rest of it...but I don't think the added wattage would be smaller than if you just using your card.

Wattage will always be a problem, especially if you put GPU into the equation.
Even if you end up utilizing your card, you have to keep in mind that you'll lose around two active CPU cores to the card since it's dual AMD GPUs.

[sykozis]

My CPU isn't getting enough PPD to even make it worthwhile to keep it going.... Thinking about putting the 560Ti back in....again....and letting it chew through another couple dozen. I need a better PSU soon....

[Psychlone]

Originally Posted by k1net1cs

Aside from OC-ing your CPU more or getting an extreme part? Unlikely.
Of course, there's also the option of getting another system which could be identical only on processor alone, and getting non-enthusiast parts for the rest of it...but I don't think the added wattage would be smaller than if you just using your card.

Wattage will always be a problem, especially if you put GPU into the equation.
Even if you end up utilizing your card, you have to keep in mind that you'll lose around two active CPU cores to the card since it's dual AMD GPUs.

OK, so what if I wanted to try a different GPU client? I know literally NOTHING about this anymore since there's a GUI now and everything is different than it was a few years back.
I tried just adding GPU0 and GPU1 to my current setup and resulted in the higher wattage (which is clearly expected)... but what about running a separate client? I did a little reading on the GPU Beta 3 client, but it's over 2 years old now - are ALL the clients inter-related and can be used from within the single client control now??

Just looking to do a little more...


Psychlone

[k1net1cs]

Originally Posted by Psychlone

OK, so what if I wanted to try a different GPU client?

Ever since v7, everything is under one GUI.
You just set what kind of client to use for a slot, and the GUI will do the rest, such as downloading the proper client for the slot.
e.g. if you set a GPU slot, it'll automatically download the proper (usually the latest) client for the GPU you're using; in your case, the OpenCL-based client for AMD GPUs.

You can still manually set separate clients using CLI like before, I think...but CMIIW.
Haven't done that for quite awhile, to be honest, ever since trying v7 betas.

Originally Posted by Psychlone

... are ALL the clients inter-related and can be used from within the single client control now??

Not really "inter-related", just being managed under one GUI.
They're still separate clients (slots) running on their own, working on their own WUs.

[Psychlone]

Got it. Thank you for your input!


Psychlone

[Psychlone]

Quick question that I haven't found a good answer for, but I think I already understand (just need clarification)

I started a WU on GPU1 of my 5970 and left GPU0 alone. It increased my watts used to 405 from ~350/360, which isn't much, but my PPD went WAAAY down from ~20,000 to just a hair under 12,000.
The SMP client has completed more than 1 unit, but the GPU picked up a big one and will take some time before it's completed.
I can stop the GPU client since it's only a few percentage into it - they're not losing much if I pull the plug on it, but I don't fully understand the implications of why my PPD dropped so significantly by ADDING a client.

My SMP client is pushing a 9188 credit WU as I write this, and the GPU client is only pushing an 1825 credit WU but is going to take 1 full day longer.


Is this because each GPU client also requires part of the actual CPU cores to function, and where they're already crunching numbers in their own SMP client, the SMP client takes precedence over the GPU and forces it to a lower priority, thus making the GPU client ???



Thanks for all the help. I want to do what I can! I'm all-in.

Psychlone


***EDIT: Just opened up HFM and saw my PPD values were way different than the console is displaying.
The SMP client is pushing 4114 PPD and will be done in an hour.
The GPU client is pushing 7585 PPD and will be done in a day.

So the PPD values are way off... Do you think I should let the GPU client fininsh it's work on this WU or just ditch it in favor of the quicker SMP on my rig and quit worrying about GPU at all?

And why is it that AMD just doesn't crunch numbers as good as nVidia? No one has ever really explained why - they're clearly different approaches to the same problem, they're BOTH clearly parallel-processing, and they're both highly optimized for their jobs... so why the bias towards nVidia?

[k1net1cs]

Originally Posted by Psychlone

... , but I don't fully understand the implications of why my PPD dropped so significantly by ADDING a client.

When you only have one slot (e.g. the SMP client) working, then the approximate PPD has only one variable to average from, which is the SMP client's PPD.
If you have two slots (e.g. SMP & GPU) and one being stopped halfway, the GUI will still average from both PPDs, which one of them will have decreasing PPD overtime and thus dragging the average PPD for both slots smaller and smaller.

So on and so forth.

Keep in mind that SMP WUs employ bonus points that are derived from how fast you can finish the WU, while points for GPU WUs are static.
This is probably why including GPU WUs to the mix impact the average PPD being shown by the GUI more than SMP WUs, because there's no bonus points from GPU WUs.

CMIIW, though.
These are purely from my observations so far.
You'd probably get more concrete answers in F@H forum.

Originally Posted by Psychlone

My SMP client is pushing a 9188 credit WU as I write this, and the GPU client is only pushing an 1825 credit WU but is going to take 1 full day longer.


Is this because each GPU client also requires part of the actual CPU cores to function, and where they're already crunching numbers in their own SMP client, the SMP client takes precedence over the GPU and forces it to a lower priority, thus making the GPU client ???

IIRC the GPU client has a higher priority than the SMP client (in terms of CPU usage), so the SMP client shouldn't be taking over the GPU client's "needs", so to speak...but maybe PantherX can clarify this.

It's also possible at the time there is something else using the GPU, such as watching a video using DXVA for example.

Originally Posted by Psychlone

Just opened up HFM and saw my PPD values were way different than the console is displaying.
The SMP client is pushing 4114 PPD and will be done in an hour.
The GPU client is pushing 7585 PPD and will be done in a day.

AFAIK the GUI approximates each slot's (client's) PPD by how fast you're getting to each 1% checkpoints for that single WU being worked on, but HFM averages from the WUs you've finished before...though I'm probably mistaken.

Originally Posted by Psychlone

Do you think I should let the GPU client fininsh it's work on this WU or just ditch it in favor of the quicker SMP on my rig and quit worrying about GPU at all?

Personally I'd wait at least a day, but it's entirely up to you.
If you're that curious then wait.
If you're more concerned about electricity bill, ditch the GPU slot, but it's recommended to at least finish its current WU first.

Originally Posted by Psychlone

And why is it that AMD just doesn't crunch numbers as good as nVidia? No one has ever really explained why - they're clearly different approaches to the same problem, they're BOTH clearly parallel-processing, and they're both highly optimized for their jobs... so why the bias towards nVidia?

nVidia GPU client uses CUDA, and AMD GPU client uses OpenCL.
It just so happens that the OpenCL-based client has to use some CPU resources.

Personally, I don't think of it as a bias, but probably either a limited understanding of OpenCL (from the client's coder) or the limitation of OpenCL itself...or both.
nVidia is also known to be more pro-active in helping developers than AMD in order to implement their standard, irrespective of whether it's proprietary (CUDA) or not.
The last thing I heard from Stanford (in F@H forum) is that there are simply not much resources to develop and/or support AMD GPU client further.
Also the reason why there is no way to fold with GPU in GNU/Linux if you have an AMD card, at least back then when I've finally lost interest in folding in Linux Mint.

[Psychlone]

Thank you for breaking all that down for me. I had a bit of a grasp, but really didn't understand a chunk of it until now.

I'll go ahead and let the GPU client finish just so it's not gone to waste... it's several more percent into it now anyway... it's only a day.
I'm not super energy-bill conscious, even though it's me that pays the bills, but I certainly don't want to go overboard. I've yet to see what the power bill is going to look like for a whole month since I just started a week ago... I can make more concrete decisions on this after this month is over.

Thanks yet again,
Psychlone

[k1net1cs]

A bit of a correction regarding the GUI's method for counting the PPD of all slots.
It should be the sum of the PPD from each slot, not the average.
My thinking was a bit carried away by EOC stats.

[PantherX]

Originally Posted by k1net1cs

...IIRC the GPU client has a higher priority than the SMP client (in terms of CPU usage), so the SMP client shouldn't be taking over the GPU client's "needs", so to speak...but maybe PantherX can clarify this...

The GPU Priority is higher than SMP. In order to explain their interaction for F@H, I will use some easy-to-understand numbers which aren't real but represents the real situation.

If you are folding on your Quad Core CPU with all the Cores (100% Usage), and the GPU wants to use 25% of the CPU, the SMP will use 75% of the CPU and the GPU will use 25% to give the overall CPU usage of 100% (assuming that nothing else is running). That is normal understanding which is correct. However, with F@H, if you start the SMP Client on a Quad core, it will create 4 threads, one for each CPU to make the folding faster. If the GPU uses one CPU, you would assume that the slow down is by 25% but that isn't the case. The slow down is significantly more because the 4 threads created needs to be synchronized after X seconds but since the GPU is using 1 CPU, the 4 threads are left to "fight" over the remaining 3 CPUs which makes the PPD drop significantly. Thus, we recommend that if the GPU uses a significant amount of CPU Usage, reduce the number of CPUs allocated to the SMP Client to the next lower even number, i.e. from 8 to 6.

I hope that you have understood this better now.

[Psychlone]

Originally Posted by PantherX

The GPU Priority is higher than SMP. In order to explain their interaction for F@H, I will use some easy-to-understand numbers which aren't real but represents the real situation.

If you are folding on your Quad Core CPU with all the Cores (100% Usage), and the GPU wants to use 25% of the CPU, the SMP will use 75% of the CPU and the GPU will use 25% to give the overall CPU usage of 100% (assuming that nothing else is running). That is normal understanding which is correct. However, with F@H, if you start the SMP Client on a Quad core, it will create 4 threads, one for each CPU to make the folding faster. If the GPU uses one CPU, you would assume that the slow down is by 25% but that isn't the case. The slow down is significantly more because the 4 threads created needs to be synchronized after X seconds but since the GPU is using 1 CPU, the 4 threads are left to "fight" over the remaining 3 CPUs which makes the PPD drop significantly. Thus, we recommend that if the GPU uses a significant amount of CPU Usage, reduce the number of CPUs allocated to the SMP Client to the next lower even number, i.e. from 8 to 6.

I hope that you have understood this better now.

Understood. I'm going to try to set affinity of the SMP to 6 cores and try GPU0 and GPU1 clients again just as a test for the next few days.
I'm pumping out a pretty decent PPD from my rig right now but always think I can do a little better.
I'll also monitor my wattage usage since that's going to be a little bit of a problem... I'm the one paying the bills!


Psychlone

[Psychlone]

Didn't work. For some reason, whenever I enable a GPU core (or cores), then my total PPD drops significantly.
I've set the SMP to 6 instead of -1 and that didn't help, so I'm going to try one last thing before writing off the GPU client... I'm going to set the SMP to 4 and try enabling a single GPU core (GPU0) just for a day to see.

Before I started playing with GPU cores, I had my most productive day ever @ 33,711 ... and that was JUST running the SMP on my rig with no other clients. Since then, I've been hovering in the 20,000 +- range. The only, and I mean ONLY thing that I can attribute it to is the GPU client. It's apparently just not primed for AMD's dual GPU - at least not the 5970 Engineering Sample that I'm running.

Anyway, off to set SMP to 4 and run 1 GPU client and see if I can crank out more total PPD daily...


Psychlone

[Athlonite]

Hey Psychlone what sort of gpu load are you getting on your GPU slot mine for some reason are sitting on 41~43% and wont go higher last time I folded on these (5770's) cards the GPU clients all maxed out at 99% workload....

and makes no difference if I pause or unpause the SMP slot it still oscillates between 41~43%

temps are fine nothing out of the ordinary top cards temp is 51c and 56c for the bottom card so well within the limit for these GPU's

[Athlonite]

seems after setting the client-type = beta in the client options I am now getting 99% GPU load on WU 7662 (24, 22, 1) with an est PPD of 16804 sweet news for ATI folders




and even better news the CPU usage for folding on an AMD/ATI GPU has gone down significantly, it used to average (for me at any rate) around 35~42% CPU usage for a single GPU slot now it's down to 7~10%

[iancook221188]

the funny thing Athlonite is the problems you were having with the AMD client are now happening to the nvidia client

[Athlonite]

well all I can say is turn around is fair play nVidia have had it there way for so long now it's not funny

personally though I tend to think evenly matched NV and AMD cards should always pull roughly the same PPD when using F@H as its mostly OpenCL doing the same work just a slightly different GPU architecture

[PantherX]

It is possible that with better/improved drivers, the Nvidia GPUs Usage may drop. However, it is anyone's guess as to when that will happen.

[PantherX]

Originally Posted by Athlonite

...personally though I tend to think evenly matched NV and AMD cards should always pull roughly the same PPD when using F@H as its mostly OpenCL doing the same work just a slightly different GPU architecture

The issue is that both comparable GPUs may not have the same OpenCL version which may make all the difference. AMD is currently ahead of Nvidia in terms of the latest OpenCL version. Let's hope that both vendors will get their OpenCL optimized so everybody benefits.

BTW, in theory, FahCore_17 can run on any supported OpenCL platform including CPUs, iGPUs, Xenon Phi, etc. Hopefully, that will be the future which would be awesome!

[iancook221188]

yep let hope Nvidia update there driver (we can dream) with better OpenCL support when this is out of beta or more stable we need to let people know there must be lots of amd card out there