Folding Forum

My GTX460@800 MHz gets ~14k PPD with projects 8005-8009 and 8018-8020. In 2012 <10k PPD is below average for a GTX460. PPD is increasing from year to year so it would be of interest when exactly the 9787 PPD was measured.
And please remember that DoctorsSon complains a PPD drop of 4-5K PPD that would be a third to a half of average PPD!

Heiko

k1wi wrote:You seem to be overlooking the benchmark information that was quoted by 7im - 9787ppd is how much a stock GTx460 should get. The extra 10% that yours is over clocked by would put the 11k bang in line with the stock standard benchmark card... Anything higher than that would be 'above average'. I don't know what WUs earn the 14k PPD but perhaps they're smaller WUs that are on the older 480 benchmarked WUs.

Here's the rundown I'm pulling from HFM's WU History of my machine. Dates are when the project(s) were first released to beta. All of the WUs should have been completed at the same 820 MHz core frequency.

7612/3: 11.2k PPD Feb 2012
7622/3: 11.1k PPD 7622: Dec 2011 7623: June 2012
7624-6: 10.9k PPD June 2012
7640-4: 9.5k PPD Feb 2012
8005: 13.5k PPD July 2012
8006-10: 13.9k PPD 8006-9: April 2012, 8010: May 2012
8018: 14.8k PPD July 2012
8019-20: 15.4k PPD April 2012
8031: 14.6k PPD Feb 2012
8033: 14.1k PPD Feb 2012
8034: 14.6k PPD March 2012
8035: 14.1k PPD April 2012
8036: 14.6k PPD April 2012

I believe I have the exact same model as the benchmark card, based on the part number listed here, apart from the fact that I've overclocked mine from 720 to 820 MHz core clock.

Assuming a linear scaling with clock frequency, I should be earning 9787*820/720=11.1k PPD. Due to when they were all released, I don't believe any of these were benchmarked on the 480.

In that case perhaps the 14k ppd WUs need to be benckmarked lower... the 7623-25 of artoar_11 looks bang on the benchmark...

Two things are standing out (to me at least). The first is that everyone other than artoar_11 who is reporting an issue seems to be reporting a pretty high OC. Maybe there are differences between the WUs that is affecting how well/badly they scale with frequency. artoar_11's point delta seems to negate this however.

It is possible that different WUs react differently to a fully loaded CPU and GPU folding, but that stock GPUs appear to be hitting the benchmark to me this would indicate that other WUs are actually over valued.

This is an interesting topic since it gets to the basic issues of benchmarking. I wouldn't even read it if we were talking about some other model of GPU. I have a couple of remarks, however. First, there are now at least three different versions of GTX460. Can anyone guarantee that all three get the same PPD? Second, although Zagen30 is apparently using the same model, on what basis can we assume linear scaling with clock frequency? Can the time spent transferring data always be zero? Overclocking can speed up the calculations, but if data has to wait for the pcie data transfer, overclocking will not speed up that transfer so the total time will be longer than linear scaling will predict.

Back to my first point: since the total time to finish a WU is the sum of data compute and data transfer and possibly other things, when nvidia changes the number of compute engines and their clock rate but doesn't change the model number, the same statement about non-linearity applies to comparing those models. Benchmarking cannot guarantee identical results unless the hardware is identical, and even when it's identical, there is an expected variation between WUs of up to 10%. Put that all together and we can explain maybe half of the variation but not all of it. The project owner should rebenchmark this one.

Zagen30 wrote:I believe I have the exact same model as the benchmark card, based on the part number listed here, apart from the fact that I've overclocked mine from 720 to 820 MHz core clock.

Assuming a linear scaling with clock frequency, I should be earning 9787*820/720=11.1k PPD. Due to when they were all released, I don't believe any of these were benchmarked on the 480.

gwildperson wrote: The project owner should rebenchmark this one.

This is certainly all I'd ever suggest or hope for. I believe it's warranted by the degree of divergence involved as well as its frequency.

gwildperson wrote:Second, although Zagen30 is apparently using the same model, on what basis can we assume linear scaling with clock frequency? Can the time spent transferring data always be zero? Overclocking can speed up the calculations, but if data has to wait for the pcie data transfer, overclocking will not speed up that transfer so the total time will be longer than linear scaling will predict.

I don't know how valid of an assumption that is myself. I've always heard that FAH isn't bottlenecked by PCIe unless it's got very few lanes available, but I do acknowledge that that transfer time is non-zero and obviously doesn't decrease just due to a GPU OC. I'm not sure what percentage of the processing time of a packet of data is PCIe transfer and how much is actual GPU crunching; I've always thought PCIe time was small in comparison, but that's not grounded in any real evidence.

I apparently had nothing better to do tonight and calculated a bunch of stats for all the GPU data in HFM's history (my previous figures were from eyeballing HFM). Here's what I got:

Project Avg PPD Std. Dev. Min Max Median Mode Num WUs
7612 11109 118 10917 11201 11143 11201 5
7613 11044 696 09084 11316 11289 11316 9
7622 11004 352 08927 11121 11093 11121 56
7623 11121 000 11121 11121 11121 11121 4
7624 10941 014 10931 10957 10931 10931 5
7625 10957 000 10957 10957 10957 10957 3
7626 10931 000 10931 10931 10931 10931 7
7640 09369 219 08939 09595 09459 09499 9
7641 09124 104 08987 09274 09116 09210 8
7642 09473 110 09353 09650 09472 09472 5
7643 09395 105 09262 09486 09446 09262 6
7644 09361 208 08867 09526 09406 NONE 8
8005 13500 165 13223 13813 13470 13554 15
8006 14023 113 13470 14082 14082 14082 32
8007 14028 054 13902 14082 13991 13991 20
8008 13953 196 12757 14082 13991 13991 177
8009 13938 064 13813 14082 13902 13902 68
8010 13869 084 13305 13902 13902 13902 203
8018 14963 190 14673 15305 14982 14982 11
8019 15396 016 15305 15400 15400 15400 54
8020 15375 076 14804 15400 15400 15400 76
8031 14597 064 14501 14629 14629 14629 4
8033 14017 336 12626 14131 14131 14131 20
8034 14375 525 12299 14629 14564 14564 35
8035 13999 257 13177 14131 14131 14131 41
8036 14629 000 14629 14629 14629 14629 9

My GTX460 is a a normal GTX460 (neither a SE nor a SE2) like in the benchmark machine, but OC by Gainward.
There´s no bonus factor for GPU and in my experience GPU WUs scale linear, so you don´t get unexpected more PPD with higher shader frequencies.

But the nVidia driver of the GTX460 benchmark machine is version 258.96, it was published in Juli 2010. Is it possible that the 9787 PPD were measured 2 years ago?
We´ve seen a PPD inflation in the last years. Maybe because of optimized WUs or just by higher points per WU.
If you search for old PPD values for GTX460, 10k PPD were normal for a GTX460@stock clock in 2010, but not in 2012!

Heiko

It is possible that there have been driver improvements that have led to better computational improvement, however I would have expected that someone would have noticed it and strongly advocated the use of the newer drivers based on a notable increase.

From Zagen30's data, it appears that on *overclocked* GTX460s, the 80xx WUs appear to return notably higher PPD than 76xx WUs, which are much more in line with the stated PPD of the benchmark GPU. Clearly 10k PPD is still normal for 2012, just less common (particularly given the frequency of folded units).

Just to reiterate my perspective: If there is any rebenching, given the evidence in this thread I would expect that there will be a revision of over inflated PPD *down*, rather increasing the PPD of the 76xx WUs...

For what it's worth, I have done 80 units 7624-7626 on GTX570s and have experienced an average reduction in ppd of 20% - so its not just on lower-end cards

Anyone getting more than 9800 PPD on a 460 GPU should be thankful, and should be careful about the tone of how a drop in points is reported for 762x WUs. So far, no one has gotten below the benchmark PPD, so this is a non-issue. If it becomes an issue, the PPD on the other projects are more likely to come down than 762x going up. Be careful what you wish for.

Ok, so I´m the lucky one with the GTX460@800Mhz getting the ~14k PPD (which the exception of 762xP of course).

@7im
Are you trying to scare us with points reductions? It´s very unlikely that this will happen.

Heiko

If my 2 full time 460's dropped to 10k each I'd quit folding with them, my 260 does 8k and I only use that and a few others during contests.

heikosch wrote:@7im
Are you trying to scare us with points reductions? It´s very unlikely that this will happen. It´s very unlikely that this will happen.

If people push and push for PG to address why the PPD is so much lower for the WUs are properly valued then I would say there is a possibility that PG will attempt to address those WUs that are over valued and thus lead to the perception that the properly valued WUs are undervalued. As to the actual odds of a devaluation, only PG would know, but as to the odds of increasing the value of 7624/5 WUs I'd say they are pretty much zero.

Nobody is paying attention to the published polciy. If projects are rebenchmarked, PG will adjust them to conform to the policy. Overvalued WUs will be reduced;undervalued ones will be increased TO CONFORM TO POLICY. It has nothing whatsoever to do with what you've gotten accustomed to.

http://folding.stanford.edu/English/FAQ-Points
•GPU: VGA GeForce GTX460 768MB DDR5 PCI-Express 2.0 Graphics Card 768-P3-1360-TR
. . .
•PPD: 9787

Zagen30 wrote:I don't know how valid of an assumption that is myself. I've always heard that FAH isn't bottlenecked by PCIe unless it's got very few lanes available, but I do acknowledge that that transfer time is non-zero and obviously doesn't decrease just due to a GPU OC. I'm not sure what percentage of the processing time of a packet of data is PCIe transfer and how much is actual GPU crunching; I've always thought PCIe time was small in comparison, but that's not grounded in any real evidence.

The statement that FAH isn't bottlenecked was based on the small WUs that we had at the time (several years ago). I doubt it's necessarily true with some of the current WUs which have gotten much bigger in the last few months. (I don't remember seeing data reassessing the original statement.) [Projects 7624-25 are not huge, however, like some of the upcoming projects that are in beta.]

Does anybody know of a utility that can report how busy the PCIe bus is?