GPU Folding

[K1tty]

I've recently started folding on my main machine. I've hit a few hiccups and will eventually get a dedicated Linux install to run folding on, but noticed a few things.

I've so far tried 3 different clients on this machine. First the standard Folding@home SMP client to run the CPU only. With this of course you're limited on what you can see. It worked alright, but seemed to hang up (hence my low PPD currently. The first 3 days were basically shot from running this client and its hangups.

Next I tried the FAH GPU Tracker V2. This worked great for CPU as the PPD showed around 30k which I think is inline with others on a similar system. However, the GPU would not work at all (7900 series not supported yet).

Finally I tried the newest beta release which identified my GPU and worked on both CPU and GPU (The Beta here)
However, my CPU PPD dipped way down to around 10k.... and the GPU (which takes forever to complete a WU was only at around 5k when I checked it).

So, all in all, is it really worth it to run this on a 7900 series GPU? Power consumption wise, wouldn't it be better to run it on another system of mine (2600k) than to use these GPUS?
Also, why all the variation in the clients? I guess it just depends on the system. I am hoping to get my Ubuntu install up tonight to see if that helps any!

PS: Thankfully I've now passed the 10 WU mark

 

[Axdrenalin]

AMD GPU's have never been able to match the output of the NVidia series GPU's, as they've (Pande Labs) never really tweaked the client to take full advantage of the AMD/ATI cards. At least that used to be the issue...I'm certain that some of the more up to date folks might have more insight on this.

Robert

 

[K1tty]

Thanks for the info! Yeah, I've debated trading out the 7950s for Nvidia once the new cards come out. I guess we'll have to wait and see. I've also been trying to get more info on whether or not I should run bigadv. For the time being I'll leave it on, but not sure if I should enable bigadv for my 2600k system.

 

[402blownstroker]

If you care only about PPD, the GPU clients for the most part are not worth running. If you are most interested in the science, yes by all means run to both types.

With a bit of tweaking you can find a happy medium of running both clients and maximizing PPD. I have a dual Xeon E5440 with a 460 GTX and a 450 GT. The 460 is limited to 90% and the 450 is limited to 75%. This combo produces about 36K PPD +- 3K depending on WUs being processed.

 

[Mr. Pedantic]

Dunno if you still get bigadv on Windows with 8-threaded machines...?

If you don't, then the GPU client is well worth running, even for points. My experience is that you should dedicate a thread to running the GPU client - run the SMP client with -smp 6 or -smp 7, leave at least a core for the GPU client, and that way the GPU won't interfere with the output of the SMP.

 

[K1tty]

Thanks Pedantic,
I think that may have been the problem. Well, I'll get to do a fresh config on my Linux install hopefully tonight when I get home. I'm installing the newest version of 64-bit Ubuntu and going to use the Beta. Think that would work alright?

 

[W.Feather]

1) Your still in windows
2) Have you done 10 work units? The beta client may take that into account (not sure, never used it)
3) If 2 is a no, thats yoru reason
4) If 2 is a yes, then the cores are not being assigned correctly.

 

[K1tty]

meh ... scratch that. Ubuntu has no wireless support for my adapter apparently. Yet another day wasted

 

[rhavern]

Dunno if you still get bigadv on Windows with 8-threaded machines...?

If you don't, then the GPU client is well worth running, even for points. My experience is that you should dedicate a thread to running the GPU client - run the SMP client with -smp 6 or -smp 7, leave at least a core for the GPU client, and that way the GPU won't interfere with the output of the SMP.

You need to be careful about using some prime numbers, some of the WU don't like it.

 

[K1tty]

1) Your still in windows
I've been trying to get up and running on Ubuntu, but I'm currently stuck using wireless and my wireless adapter doesn't play well
2) Have you done 10 work units? The beta client may take that into account (not sure, never used it)
I surpassed 10 the other day. I think one of the clients was taking this into consideration however even before I reached 10 WUs.
3) If 2 is a no, thats yoru reason

4) If 2 is a yes, then the cores are not being assigned correctly.
I am now exclusively using the beta client since I can do both GPU and SMP. I think this client takes quite a while (and probably more accurately) calculates the PPD. I'll look and see what it is when I get home. Last I checked the SMP was climbing into the 20k range and the GPU was around 4k (only using one atm). I also checked the option in the beta client to assign core affinity. Currently not running -bigadv due to all the down time I've been having while messing with this

 

[K1tty]

You need to be careful about using some prime numbers, some of the WU don't like it.

So, I'd be better off designating 5 cores/10 threads to the SMP and 1 core/2 threads to aide the GPU?

 

[Mr. Pedantic]

You need to be careful about using some prime numbers, some of the WU don't like it.

I've heard that, but I've run -smp 7 for a long time before and there were no problems. The reason I use -smp 6 now is because I actually have to use my computer, and having 2 free threads actually makes for faster WUs than 1 free thread.

 

[rhavern]

So, I'd be better off designating 5 cores/10 threads to the SMP and 1 core/2 threads to aide the GPU?

That would be advisable. If you get into a stretch of WU that rapidly fail due to the prime number issue, you can kill your 80% completion rate that will nuke your QRB for some time.

 

[K1tty]

Alright, I've updated it so the next run it should leave 2 threads for the GPU to work better, or so I hope.
I've also ordered a wireless adapter that works with ubuntu linux

However, I've been contemplating trying out Linux Mint, or just running it on my OSX install....