Folding Forum

I'm curious if anyone else has noticed that the P92xx project series works a treat on Maxwell GPU's, but sucks for Pascal GPU's?

If I take a snap shot of two GTX 1080 cards working on similar projects; PPD for one card on P9207 is 500k, and PPD for the other on P9209 is 305k. This compares with the same snap shot of GTX 970 cards; PPD for P9205 is 350k and P9207 is 300k.
In contrast, on other, non-P92xx projects, the GTX 1080's can typically be in the 700-900k PPD range.
Is this project series, then, optimised for Maxwell?

I'm also seeing PPD all over the place on Pascal. Unfortunately I've just migrated over to a new HFM install, so I don't have many results to share, but top and bottom 3 WUs from a pair of 1080 FE:

11430 - 927k PPD
9675 - 855k PPD
11429 - 844k PPD
...
9211 - 486k PPD
9209 - 485k PPD
9213 - 480k PPD

I'm currently working a 9206 and that's showing a middle of the road 744k PPD. It's also worth mentioning that there is easily a 10% usual variation in PPD between WUs of the same project - it suddenly seems to be more significant when it's worth 80k PPD!

Of course, PPD is going to vary on any machine that differs from the benchmark one, but as hardware evolves and diverges further from that standard, PPD becomes more inconsistent and gets further amplified by the QRB. I think it would be beneficial if WUs were routinely rebenchmarked every year or so on 'representative' hardware to keep things a bit more consistent.

DocJonz wrote:Is this project series, then, optimised for Maxwell?

Using the same FAHCore, it would seem that any such differences are in the way NV optimizes OpenCL in their drivers (Pascal vs. Maxwell).

And Pascal is so new that Nvidia may not have done much OpenCL optimizations for it yet, in the future it may well be even stronger.

Just to add to this, I'm seeing some higher run 9208s taking 50% longer on a mix of pascal, maxwell and Kepler. Looks more like a WU variation than an architecture/driver one.

I've noticed a very consistent pattern with these projects, all run by Jade SHi where the TPF is always 2 to 4 times what I normally see and the estimated PPD is 40 to 50 percent less. I get a lot of these and they are really hurting my PPD. It's especially annoying when I get a couple of these at the same time. I don't see these issues with any of the other project WUs I get. I'm just wondering if anyone else has seen this? I'm folding on 3 nvidia gtx 1080 cards on a linux Mint system. I don't over clock, my temperatures are good, and I'm not getting any errors.

Also reported here:
viewtopic.php?f=74&t=29035

I suspect it's just part of the normal variation in WUs, but getting amplified by super fast cards and the QRB, but hopefully the project owner can confirm.

I guess what has me most upset about these is just how many of them I'm getting. Most of the time I'm working on at least one, sometimes two of these. I was looking at the project list and since there are so many projects, if assigned randomly, I shouldn't get but a couple a month. However they had among the highest number of atoms being studied. So I'm guessing that since I have pretty fast cards that the algorithm is intentionally matching them with me.

SteveWillis wrote:I guess what has me most upset about these is just how many of them I'm getting. Most of the time I'm working on at least one, sometimes two of these. I was looking at the project list and since there are so many projects, if assigned randomly, I shouldn't get but a couple a month. However they had among the highest number of atoms being studied. So I'm guessing that since I have pretty fast cards that the algorithm is intentionally matching them with me.

The project list might be long, but there's probably significantly fewer projects with sufficient remaining WUs and high enough assignment priority available for your mix of hardware and OS.

Right now, my ppd across 7 GTX 1080:
530k PPD (p9209)
730k PPD (p9206)
850k PPD (p13500)
850k PPD (p11400)
680k PPD (p10496)
500k PPD (p9212)
900k PPD (p11707)

Quite a spread! On 'average' I trust the PPD to even out, but I agree, it's frustrating to have multiple 'slow' WUs and seeing a million less PPD. It's compounded by not only are the 'slow' workunits worth fewer points, but they run for longer (a 50% mix of fast WUs earning 1M PPD and 50% slow earning 500k PPD doesn't average at 750k, but rather 670k). Short of Stanford rerunning a more rigorous benchmarking process or modifying the QRB, I don't see much changing unfortunately.

p9210 R42-C24-G2 best TPF 00:05:57 resulting PPD 358,110. This is the first time I see such little PPD on 92xx projects
I use GTX 1070 GPU 2012 / MEM 2003 WHQL 373.06 driver

Gonna Leave it crunching as is. No harm to science my consciousness will allow Hope next projects assigned to my client will be quite fine
*EDIT* Finished with 353076.8PPD At least it' like a GTX970, not lower
p.s. jumped from this topic

I've attempted to contact the owner of these project and have him re-benchmark his projects. I'm also merging another discussion of the same issue.

That is very cool. Thanks!