As announced in another thread the stats pages now list the relative ranking of GPUs in AP27, where fastest = 1 and everything else is lower than that.

1.000 1.000 GeForce GTX 1070
0.996 0.963 GeForce GTX 980 Ti
0.629 0.662 GeForce GTX 970
0.439 _.___ GeForce GTX 1060 6GB
0.206 0.197 GeForce GTX 750 Ti
0.153 0.168 AMD Radeon HD 7870/7950/7970/R9 280/R9 280X series (Tahiti)

Selected copy/paste above, first value is from primegrid stats, 2nd is what I worked out of my GPUs, normalising for my 1070=1. Obviously the relative scores will vary as I don't know where my 1070 sits relative to the primegrid average.

More interesting is 1060 6GB is quite a lot lower than the 970 which I wouldn't expect. Based on cores and typical clocks I'd expect the 1060 to be faster than a 970. Now, I don't know how many 1060 results there are, so it could just be one person with a slow configuration for all I know.

By multiplying the number of cuda cores, boost clocks and 2, you get the theoretical FLOPs. From that you can work out the relative placing expected of those cards:

1.000 GeForce GTX 1070
0.937 GeForce GTX 980 Ti
0.677 GeForce GTX 1060 6GB
0.607 GeForce GTX 970
0.606 GeForce GTX 1060 3GB
0.215 GeForce GTX 750 Ti

I do think if buying a new card optimised just for this project, the 1060 3GB would be an interesting choice, as even though it has fewer cores than the 6GB model, the cost per core works out lower. For me at least, cheapest 3GB is about 20% less than a 6GB, for 10% fewer cores.

In other testing, Pascal does show a performance per watt improvement over Maxwell, at roughly 25-33% less energy per unit done.

In team red there's only limited info to go on, but they're not fast at this application and gobble up a lot of power getting there. I don't see any Polaris results yet and I'm still curious if they are any different than previous generations. Still got a temptation to pick up a 460 or 470 to bench.

For reference, my 970 is consistently doing tasks at around 2200s, which you can check out here. It's OCed to 1430mhz, but I figure I'd mention it, given that you posted on the other topic that yours is taking ~2500. No info on power, though, and I can't really be bothered to do it right now.

I ran a few units through on a 480, quickly drawing the conclusion it was "not good" at them.

6 completed, low 6233, high 6410.

[Edit] Best time I got on a 290x, 5096. But I suspect I was consuming way more power.

Rafael, that's a nice clock boost and likely explains your faster times.

Van Zimmerman, I think that pretty much confirms AMD GPUs are not the best choice for AP27 project.

Just to add the performance per watt results... to recap, my method was to simply look at the system idle power, then the system power with AP27 running. The difference is then the extra power required to run AP27. This is measured at the wall, so will include any PSU inefficiency. Most if not all my GPU systems are running 80+ rated units.

By factoring in the amount of time it took to run the unit, I can work out the Wh consumed per unit, which is how I previously presented the results. This is a little less intuitive to digest as lower is better, so I just flipped it over instead. Following results are units per kWh, so kinda like a miles per gallon measure where higher is better.

These are for my systems, and results may vary depending on how the clocks and voltages are set per card, and other factors.

14.8 - GTX 1070 (Zotac FE)
11.2 - GTX 980 Ti (Asus reference, after market water cooling)
9.9 - GTX 970 (EVGA SC ACX 2)
10.1 - GTX 750 Ti (PNY)
2.9 - R9 280X (XFX DD Black edition - un-overclocked, lowered voltage)

I should comment that the 980 Ti was measured for a new unit this morning. The time that took was a fair bit shorter than the one I did last night. The GPU is in my main machine and I was using it at the time, so that might have made the 1st unit worse. Looking at the previous 1070 results where I did 34 units in total, the variation between units was 1.4% so not significant.

Also the 980 Ti GPU is watercooled. On the R9 280X I observed what looked like thermal runaway previously, in that as it got hotter, it took more power, which made it hotter and so on... thankfully the built in throttling stopped it before it got worse. So back to the 980 Ti on water, it would run much cooler than with the standard cooler on it, so there may be an efficiency benefit from that.

For reference, my 970 is consistently doing tasks at around 2200s

For reference, my 970 is consistently doing tasks at around 2200s

That makes your GTX 970 faster at this than my GTX 1060 6GB which gets around 2450s per task.

Please remember that the numbers on that page are NOT Gospel. There's lots of reasons that the numbers can be distorted. Certainly don't base purchasing decisions upon the information on that page without verifying that the information is correct.

For example, you'll see a few places where a 1070 is the fastest GPU -- and a 1080 is slower. I doubt anyone believes that's accurate.

That said, adding one more datapoint, my GTX 580 does AP27 tasks in about 4500 seconds, and draws about 145 watts while running, as measured at the UPS.

(Wattage consumption between computers isn't really accurate because you're not measuring how much power the GPU is using. You're measuring how much power the power supply is using, so what you're actually seeing is GPU consumption plus power supply inefficiency.)
____________
My lucky number is 75898⁵²⁴²⁸⁸+1

It was implied that the primegrid stats are probably an average of all the valid results for a given reported card, and thus if there are few results then outliers might swing the results more significantly. The 1060 certainly is not where I think it should be in the stats as they were.

Taking Keith's 1060 6GB time relative to my 1070's time, it would be 0.679. Pretty close to the 0.677 where I'd expect it to be based on reference boost clocks.

It would be easier to compare if, instead of the normalised results, the actual average time was available. Even if just for the top fastest card, the rest could be calculated from there.

On the topic of including PSU efficiency in the results, that could be a reason why my 980Ti result might be above other Maxwell cards. It sits in my main system where I bought a nicer PSU than for the other systems. It is 80+ platinum rated, as opposed to the lower 80+ tiers of the others.

Taking Michael's reported values for the GTX 580, I make that 5.5 units/kWh. Comparing the time against my 1070, that puts its relative performance at 0.370, compared to 0.340 showing in current stats so that's pretty close. Again, I have no idea where my 1070 sits relative to primegrid average 1070.

It was implied that the primegrid stats are probably an average of all the valid results for a given reported card, and thus if there are few results then outliers might swing the results more significantly. The 1060 certainly is not where I think it should be in the stats as they were.

BOINC limitation? So you don't really know what the returned result was done on? That might explain some of the unexpected results, if people are running multiple different GPUs on uncommon cards.

I also found it a little confusing as there's two lines with 280X in it, and I picked the one described the same as mine for the earlier comparisons.

Will be interesting to see how this evolves assuming activity picks up later.

BOINC limitation? So you don't really know what the returned result was done on? That might explain some of the unexpected results, if people are running multiple different GPUs on uncommon cards.

I also found it a little confusing as there's two lines with 280X in it, and I picked the one described the same as mine for the earlier comparisons.

Will be interesting to see how this evolves assuming activity picks up later.

gpu AMD Radeon R9 280X

linux - 5216.82s

windows - 9349.71s

NVIDIA GeForce GTX 970 on Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz HT on with ubuntu lts 64 bit: 2,568.64
____________
Sysadm@Nbg
my current lucky number: 321303946 ^ 65536 + 1
PSA-PRPNet-Stats-URL: http://u-g-f.de/PRPNet/

So, RTX 2080 made it to the top of Fastest GPU PPS Sieve and I was wondering how it would do on AP27.

Going a month back to my previous posts in HARDWARE ⋮ RUMOR [Confirmed] NVIDIA Launching RTX 2080 Ti
AP27: 750 sec (12m30s), twice as fast comparing to GTX 1070.

My recent test are showing 760-830 sec, depending on GPU clock, CPU used and perhaps driver version.

Might do a bit more testing and wondering when it will appear on Fastest GPU AP27 list


EDIT: OK, some more run times to compare.

GTX 980 Ti ~ 1550-1690 secs
GTX 1080 ~ 1090 secs
GTX 1080 Ti ~ 940 secs
Titan V - 410 secs
And just for fun - i7-4960X with 62 900 secs.
____________
My stats

So... updated AP27 with RTX 2080
"faster" of AMD is at position 20 :-(

(1.000) TITAN Xp
(0.942) GeForce RTX 2080
(0.729) GeForce GTX 1080 Ti
(0.625) GeForce GTX 1080
(0.538) GeForce GTX 980 Ti
(0.470) GeForce GTX 1070
(0.401) GeForce GTX 980
(0.350) GeForce GTX 1060 3GB
(0.338) GeForce GTX 1060 6GB
(0.276) GeForce GTX 1070 Ti
(0.248) GeForce GTX 1050 Ti
____________
My stats