Folding Forum

I am an instructor for a Technology Center in Norman Oklahoma. My class is running FAH on 36 Dell Optiplex 9020's with GTX 645 cards, about 8 older dual core machines and 30 Quad Core Mac work stations.

We are moving up in the stats pretty quick. We are making almost a million ppd. Not bad for beginner right?

Last week I bought an ASUS Strix 970 GTX to put in my home machine to contribute to the project thinking it would improve our stats. Higher stats means that when budget time comes, my board can see we are doing a lot of good which means more funding for more computers or at least GPU's.

My issue is that when I brought the 970 online, it grabbed a core 18 WU and the PPD is aweful. currently my PPD for the 970 machine is about 76,000.

Since these are much more complicated, AND take more time to process (20 hours), shouldn't they be worth a lot more points?

I am running an older ASUS M2N32 SLI Deluxe MB, 4gb Ram, Dual Core AMD.

Any comments?

Thanks

Tommy

The problem with Core_18, and Core_17 as well, is your 970 GTX card. All Maxwell based cards from nVidia are having the same problem. There is a bug in the implementation of a key OpenCL feature used by these cores in the drivers, etc. that nVidia has released for Maxwell based GPU's. To get folding to work at all on Maxwell's, software code to work around the problem was added. Unfortunately that results in a significant loss in processing speed on the Maxwell cards compared to running on Fermi and Kepler based ones.

Hopefully nVidia wil fix this in their drivers fairly soon. No idea how soon that will be. Some folders have posted in the developers forum at nVidia requesting this fix be a priority.

Here is link, not very successful yet though
https://forums.geforce.com/default/topi ... river-bug/

And @OP, you could try to run GTX 970 on Linux, no Core 15 nor core 18. While it safe to assume that Core15 as EOL will not make it to Linux we patiently wait for a core 18 deployment to Linux

ChristianVirtual wrote:Here is link, not very successful yet though
https://forums.geforce.com/default/topi ... river-bug/

Understatement of the century… in communication terms NVidia make PG look positively garrulous

ChristianVirtual wrote:And @OP, you could try to run GTX 970 on Linux, no Core 15 nor core 18.

With the caution that your PPD goes to zero if P9201's should dry up for any reason.

I also own a GTX 970. I have not encountered Core 18 mainly because it's a Windows-only core and I'm running Linux, but the card is very new and the drivers are new and buggy across the board. The 970 is a phenomenal card hardware-wise, but in terms of drivers it hasn't been smooth for F@h users. I hope they can get the issues sorted soon.

In the meantime, I recommend continuing to fold until the PPD issue is resolved by Nvidia. Core 18 uses the latest OpenMM scientific software so it's very valuable to the researchers. The Maxwell thing is holding it up at the moment.

I sure wish PG can come up with a smooth transition when/if NV fixes their driver, as opposed to having the existing core18 version around "forever" causing donor anxiety. The way I see it, an upgraded core18 would have to detect the NV driver version, among other things, and make a runtime decision whether or not to fall back on the workaround. Hardly the kind of cross-platform compatibility FAH is looking forward to, not even within driver releases for Fermi+.

OTOH, the performance penalty of the workaround seems to be very big and Maxwell GPUs still seem very promising, so I keep hoping that PG is able and willing to bridge the gap somehow.

I used to believe that the Red team's drivers were Bad and the Green team's drivers were Good, but looks like they're both consistently Ugly.

Napoleon wrote:OTOH, the performance penalty of the workaround seems to be very big and Maxwell GPUs still seem very promising, so I keep hoping that PG is able and willing to bridge the gap somehow.

This is the kind of thing that needs to be addressed. From a researcher's perspective, running the latest scientific code at a slow pace is better than it not running at all, but from a donor's perspective it makes core18 look really bad, or it makes F@h look inconsistent. Maybe core18 should be moved back to beta so it doesn't annoy so many others until Nvidia fixes their bug.

Napoleon wrote:I used to believe that the Red team's drivers were Bad and the Green team's drivers were Good, but looks like they're both consistently Ugly.

What you did there, I see it. weeewww weeeww weeeeewwwwwwwww

Jesse_V wrote: Maybe core18 should be moved back to beta so it doesn't annoy so many others until Nvidia fixes their bug.

I'm not sure that it helps. Core18 was rolled out because their weren't enough Core17 WUs that would run on Maxwell. Maybe that has changed, but maybe without Core18 for Maxwell we would soon be back to Maxwells sitting idle

And with the current (non) response from nVidia to a threadful of reports of the problem, it doesn't look like we'll get a fix any time soon.

Other than moving to a linux platform, is there a way to refuse the core 18 WU's?

My concern is that I do not want to invest our money into cards that are not optimized for this application.

Is there a different card that we can look at?


Thanks

Tommy

No, we all fold what we get. It's a central premise. All the work needs to get done.
Have patience, it will be optimized eventually.
Yes, but it won't change anything. All newer cards are in the same boat, as are all of us donors.

There is no Linux Core18, yet. It may be coming.

My personal decision is to delay investment in Maxwell GPUs until nVidia release a fix (and it is known to work on Linux). You must make your own decision.

davidcoton wrote:There is no Linux Core18, yet. It may be coming.

My personal decision is to delay investment in Maxwell GPUs until nVidia release a fix (and it is known to work on Linux). You must make your own decision.

+1

My Machine with the 970 finished a core 18 WU and drew a core 17 WU. PPD for this was 223,000. THAT is more like it.

Still not good compared with a 780Ti -- 240K ppd on 9201, currently 270K+ ppd on a 10469. Come on nVidia, get it sorted.

davidcoton wrote:Still not good compared with a 780Ti --

But not bad at all if you compare number of cores and, especially, the power consumption (heat production).

davidcoton wrote:240K ppd on 9201, currently 270K+ ppd on a 10469.

That's about what my 780 Ti gets. I can overclock it to get more, but it's not the easiest card to keep cool and then it tends to fall over on the 104xx's

Currently 390k+ with a P9201 on a GTX 980. GPU temp <60ºC and quite often in this weather the fans aren't even running.

davidcoton wrote:Come on nVidia, get it sorted.

+1

Until they do I'm not looking forward to a Linux Core_18. Unless P9201's dry up first of course… it's a complicated world