F@H double vs single precision

[imreloadin]

So I've been away from folding for a while now and want to get back into it and make a machine that only does that and run it 24/7.

I've been using the FAHBench results to try and find the most cost effective GPU to run but I'm a little confused in regards to Single Precision vs. Double Precision.

 

I know Double Precision is more work for a GPU than Single Precision but how does that relate to PPD? Are you able to choose whether you want to only do Double Precision over Single Precision or does it choose that based on the WU?

Any help on clarification from the gurus here would be much appreciated

[Electronics Wizardy]

double precision uses 64bit float values instead of 32bit. If you on a gtx gpu, use single precision, as the double precision is very heavily cut down. Id only use double precision if you have a orginal titan, a quadro(except maxwell and pascal) or a tesla.

[imreloadin]

2 minutes ago, Electronics Wizardy said:

double precision uses 64bit float values instead of 32bit. If you on a gtx gpu, use single precision, as the double precision is very heavily cut down. Id only use double precision if you have a orginal titan, a quadro(except maxwell and pascal) or a tesla.

Yeah that's one thing I noticed in my comparisons, so I take it that the points you receive for those work units isn't weighted differently than single precision work units?

[alextheaverage]

I know this is a late reply but as far as I know F@H leans much more heavily on single precision hardware as in many cases the double precision just isn't there on a lot of cards, along with this the card needs to be optimized for to be able to be effectively used by processing programs such as F@H.  Take for example the R9 290 compared to the GTX 970.  The GTX 970 having 3494 GFLOPS of processing power under a perfect load and producing around 340,000 PPD, while the R9 290 having 4848 GFLOPS under a similar load yet only producing around 250,000 PPD.  From that we can gleam the raw theoretical specifications are not always indicative of real world performance even in raw computation such as F@H.

 

I would recommend looking at this list on overclock.net for looking at around what each card will produce if you haven't found it yourself already.

http://www.overclock.net/t/475163/gpu-projects-ppd-database

 

Now double precision tasks are a different matter,  the difference in power between the GTX 970 and the R9 290 is huge, one of the many reasons so many were bought for coin mining a few years ago and are still being purchased for Etherium mining today.  The GTX 970's theoretical double precision output is just 109 GFLOPS while the R9 290 holds of to a fair bit more with it coming in producing 606 GFLOPS.  Now this difference of power doesn't show in F@H due to it's preference toward single precision tasks.  Where it does show is in some task on the BOINC platform, for this I'm going to use Primegrid as an example as it is my primary project on all of my rigs, plus it has a GPU performance list built into their site.

 

On GFN 22 tasks which are the most intense tasks for GPUs at the moment on Primegrid the R9 290 pulls far ahead in work unit completion time,yet for PPS Sieve it falls far behind as that's a single precision load.  Now for the PPS sieve that's not all that big of a deal as my R9 290s complete one of those tasks in around 20 minutes, yet for the GFN 22 tasks my R290s can take up to 3 and a half days, so the 970's lack of power for double precision really impacts it here where I've seen it reported as taking up to 5 days.  Just for a comparison beyond those card the HD 7950 produces 717 GFLOPS of double precision work and complies it's GFN-22 tasks in just over 3 days, and ive seen it complete one in under 3 before, but only once.

 

Sorry for the wall of text, was more so for my self than anything else and there's no reason to delete it now that I've already types it out.

 

TL;DR  Look at preference lists to see real PPD numbers as it's not always straight forward.  But generally for F@H Nvidia>AMD.

F@H   http://www.overclock.net/t/475163/gpu-projects-ppd-database

Primegrid  https://www.primegrid.com/gpu_list.php