AMD FirePro S9150 for PrimeGrid Genefer

Discussion in 'Distributed Computing' started by TennesseeTony, Mar 31, 2017.

  1. TennesseeTony

    TennesseeTony Elite Member

    Joined:
    Aug 2, 2003
    Messages:
    1,770
    Likes Received:
    292
    I'm trying to find something this card will excel at doing. It was mentioned a few weekly stats ago, that GFN tasks might be heavily dependent on double precision. So I figured what the heck, I'll see how it does and ask you guys for some comparison numbers. Then I thought, well, there's no reason I can't contribute some comparison numbers.......and that's when this got ugly. lol

    I'm going to go out on a limb and say that, no, GFN has little to no use for double precision, at least not for these little ones. Maybe GFN21/22 does, but I'm not going crunch a single task for days on end just as a test. ;)

    PrimeGrid Genefer
    FirePro S9150
    GFN15 completed in 2m12s
    GFN16 completed in 4m38s
    GFN17-Low finished 9m13s
    GFN17-Mega was at 14m14s
    GFN18-28m
    GFN19-Still running, but looks like 2h8m


    GTX1070 --This contest is over. About half the time.
    GFN15-0m50s
    GFN16-2m44s
    Skip to the good stuff
    GFN18-16m
    GFN19-62m

    The card was purchased in hopes of utilizing it's massive 2.53TFLOPS of double precision power (compare that to 0.368TFLOPS of the GTX1080Ti). Looks like you have to specifically code for this card, to get all those horses running at once. Sigh. Sad day.
     
  2. Loading...

    Similar Threads - FirePro S9150 PrimeGrid Forum Date
    PrimeGrid: CPU benchmarks Distributed Computing Jan 5, 2017
    GTX 1080 vs. 1070 in Folding@Home and PrimeGrid Distributed Computing Jan 2, 2017
    PrimeGrid Races 2017 Distributed Computing Dec 29, 2016
    PrimeGrid Olympics Challenge August 2-5 Distributed Computing Aug 1, 2016
    1GB ATI Firepro 5800 - will it crunch? Distributed Computing Nov 16, 2010

  3. TennesseeTony

    TennesseeTony Elite Member

    Joined:
    Aug 2, 2003
    Messages:
    1,770
    Likes Received:
    292
    From Michael Goetz-One thing to remember is that ONLY GFN 21 and GFN 22, amongst all the apps at PrimeGrid, still use double precision. All the other GFN apps, as well as AP27 and PPS-Sieve, do not use double precision at all.
     
  4. Orange Kid

    Orange Kid Elite Member

    Joined:
    Oct 9, 1999
    Messages:
    3,089
    Likes Received:
    139
    That's a no good.
    Looks like the DC community is programming for masses. I guess I can understand that as the more work that can be completed the better for science. I think that cards like the FirePro are being used mostly for a specific purpose and software is written specifically for it, ie; 3d rendering or modeling. Not something the average Joe uses. I would guess it comes down to drivers or programming.
    Bummer, would have liked to see it be king. :)
     
  5. TennesseeTony

    TennesseeTony Elite Member

    Joined:
    Aug 2, 2003
    Messages:
    1,770
    Likes Received:
    292
    Yep. I can barely get it to match an ancient R9-280X. Trying my hand (again) at Linux to see if that can fully utilize the card, but so far, Linux can't even recognize properly use an even more ancient Quadro FX1800 just for a basic adapter.

    It seems double precision is going away, and getting replaced with "run it twice" since the 'new' standard is high single precision vs double precision. That's my take on how it's done anyway, but I'm sure it's more complex than just running it twice, or even three times.
     
  6. StefanR5R

    StefanR5R Senior member

    Joined:
    Dec 10, 2016
    Messages:
    324
    Likes Received:
    136
    In DC and mainstream GPU computing perhaps. But GP100 is evidence that there is still a niche for FP64 GPU computing in one or another well-funded HPC endeavor. I take it from rumors that Vega2 is supposed to have maybe at least one professional SKU with proper FP64 support too.

    From what I understand, FP64 GPU computing in engineering and science has always been a small niche, but having competition more in CPU cluster computing rather than FP32 GPU computing. Challenges in GPU computing include the latency of moving jobs onto and off from the GPU, memory limits of GPUs, and chronically cumbersome and buggy libraries with ever evolving APIs. None of these points would be addressed by an FP32 GPU algorithm which can emulate the precision of an FP64 GPU algorithm.

    I concede though that I do not know whether the rarity of FP64 GPU computing is the chicken, and availability of good FP64 support almost only in Quadro/Tesla and FirePro lines (and there only in higher-end SKUs and only every second or third architectural generation) is the egg, or vice versa.
     
  7. StefanR5R

    StefanR5R Senior member

    Joined:
    Dec 10, 2016
    Messages:
    324
    Likes Received:
    136
    So, are you considering to test how S9150 compares to R9 280X in GFN 21-22? And if it does well, will you be readily shifting your focus from cosmology to number theory...?
     
  8. TennesseeTony

    TennesseeTony Elite Member

    Joined:
    Aug 2, 2003
    Messages:
    1,770
    Likes Received:
    292
    I've just been desperately trying to find a way to see if this S9150 is defective or just unsuitable. I've determined it is just unsuitable for DC. Linux didn't help. I removed the card and boxed it back up, reinstalled some R9-280Xs, inserted my Win10 USB key back into the test computer, and was back to BOINCing along as usual in record time.
     
  9. crashtech

    crashtech Diamond Member

    Joined:
    Jan 4, 2013
    Messages:
    6,515
    Likes Received:
    205
    Well, that's a disappointment, but at least now we know... :(
     
  10. TennesseeTony

    TennesseeTony Elite Member

    Joined:
    Aug 2, 2003
    Messages:
    1,770
    Likes Received:
    292
    Anybody wanna buy a S9150 for $600? lol Hopefully the seller will take it back, maybe a small restocking fee.