MilkyWay@H - Benchmark thread Winter 2016 on (different WU) - GPU & CPU times wanted

Discussion in 'Distributed Computing' started by Assimilator1, Dec 29, 2016.

  1. iwajabitw

    iwajabitw Senior member

    Joined:
    Aug 19, 2014
    Messages:
    632
    Likes Received:
    92
    Put this system together last week.

    AMD FX-6300
    MSI 970A-G43
    8GB Gskill Ripjaws
    2x MSI R9 280x
    Running 3 tasks per card, getting about 400,000 ppd from each card daily. GPU time varies from 70-110 sec per task with very little cpu time 24sec or so.
    https://milkyway.cs.rpi.edu/milkyway/results.php?hostid=723824&offset=0&show_names=0&state=4&appid=

    But goodness does the top card run hot! 80-85c with a side 140mm blowing on the cards, case is well ventilated. Makes a nice bedroom heater through this small cold snap. But I am betting I'll have to shut down for summer or smoke the cards. May try 1 task per card when the time gets here, just to see if I can keep it up. I'd like to hit 100 mil first.
     
    crashtech likes this.
  2. crashtech

    crashtech Diamond Member

    Joined:
    Jan 4, 2013
    Messages:
    6,918
    Likes Received:
    320
    Twin 280X GPUs make for a beast of a M@H cruncher!
     
  3. iwajabitw

    iwajabitw Senior member

    Joined:
    Aug 19, 2014
    Messages:
    632
    Likes Received:
    92
    I have gotten good deals off Ebay on the cards. Waching a few more, I blame Newegg for having the cpu combo cheap. LOL
     
  4. Assimilator1

    Assimilator1 Elite Member

    Joined:
    Nov 4, 1999
    Messages:
    22,846
    Likes Received:
    21
    Nice MW rig! :D

    Lmk when you've done some single tasks & then I can post your times.
     
  5. crashtech

    crashtech Diamond Member

    Joined:
    Jan 4, 2013
    Messages:
    6,918
    Likes Received:
    320
    I have lots of AMD/ATI* cards that I could potentially post times for, if you feel like updating the OP with them. I posted some times here, but since they didn't make it to the OP, I thought this thread was pretty much over.

    *5870, 5970, 280x, 285, 290, and Fury.
     
  6. Assimilator1

    Assimilator1 Elite Member

    Joined:
    Nov 4, 1999
    Messages:
    22,846
    Likes Received:
    21
    Huh? Who said it was over? :confused:
    No it's still going :), it's just that I haven't been on here for a while (partly because I've recently moved into a flat on the 2nd Feb), but also I just don't come on here as regularly as I once did. That said, I should've got some email notifications, but it appears I didn't (I did get 1 this eve) or the email got buried! :eek:.
    So apologies for not replying sooner, but this thread isn't dead :)
    I will update the op with your times, but I can't do it tonight or tomorrow, hopefully later in the week I can though.
     
    crashtech likes this.
  7. iwajabitw

    iwajabitw Senior member

    Joined:
    Aug 19, 2014
    Messages:
    632
    Likes Received:
    92
    Before I shut down for summer, hope I don't have to. I am going to try a week of single tasks to see if the top cards will stay at a decent temp 70C's instead of the mid 80C's, so that will be a good time for me to get those run times. Weather temps have been crazy, last week was cold, today it's warm and time for shorts, after the storms tomorrow it will be cool again. Crazy.
     
  8. TennesseeTony

    TennesseeTony Elite Member

    Joined:
    Aug 2, 2003
    Messages:
    2,010
    Likes Received:
    399
    280X's max at 100C. Don't fret. Although at 98C they tend to lock the computer. ;)
     
  9. iwajabitw

    iwajabitw Senior member

    Joined:
    Aug 19, 2014
    Messages:
    632
    Likes Received:
    92
    Haha, good to know. It's just no one is ever home during the day and I do set backs on my house thermostat, so it can get pretty warm in the house before the kick down around 5PM.
     
  10. crashtech

    crashtech Diamond Member

    Joined:
    Jan 4, 2013
    Messages:
    6,918
    Likes Received:
    320
    I check on mine with remote desktop software.
     
  11. iwajabitw

    iwajabitw Senior member

    Joined:
    Aug 19, 2014
    Messages:
    632
    Likes Received:
    92
    Oh yes, I'm a teamvier'er lol.
     
  12. TennesseeTony

    TennesseeTony Elite Member

    Joined:
    Aug 2, 2003
    Messages:
    2,010
    Likes Received:
    399
    Mark, I have a new, never before reported GP-GPU for you. Single task time on the 'new' 5x-Milkyway tasks......a ridiculously incredible 38.5 seconds. (Well, that's embarrassing, I thought that was a really good time, but I just checked the first page, and it's only a touch better than a 280X)

    I'll only tell what this single card is once you get off your butt and rejoin the TeAm. ;) (miss you bro.)
     
    #87 TennesseeTony, Mar 30, 2017
    Last edited: Mar 31, 2017
  13. StefanR5R

    StefanR5R Senior member

    Joined:
    Dec 10, 2016
    Messages:
    589
    Likes Received:
    231
    How many WUs should be processed simultaneously on this GPGPU for peak throughput? And how does this peak throughput compare to R9 280X's?

    (Edit) Also: How many CFM of air do you pump through this card, and how? :sunglasses:
     
    #88 StefanR5R, Mar 31, 2017
    Last edited: Mar 31, 2017
  14. TennesseeTony

    TennesseeTony Elite Member

    Joined:
    Aug 2, 2003
    Messages:
    2,010
    Likes Received:
    399
    The AMD FirePro S9150 finally arrived, with 2.5 TFLOPS of Double precision, which is 2.5x the output of a single R9-280X, but for only 10 more watts TDP.

    I just looked at the first page, and turns out 38.5 seconds isn't very impressive after all. :oops:

    I'm still testing it on various projects. So far only MW performance is ok. Einstein strangely was not very impressive, not bad, but slower than the 280X. Folding at Home...terrible, 90k ppd. When F@H finishes up I'll put it back to Milkyway and start cranking up the concurrent tasks to see what happens.

    As for cooling, the card requires 20 CFM, and I am feeding it with a blower rated at about 40 CFM, which has been taped to it with painter's tape. :D There are no monitor connections, this is a co-processor only, so it was easiest to tape the fan to what is typically the monitor output end of the card.
     
    StefanR5R likes this.
  15. TennesseeTony

    TennesseeTony Elite Member

    Joined:
    Aug 2, 2003
    Messages:
    2,010
    Likes Received:
    399
    Ok, so not being satisfied with this card so far, I double checked my driver, updated the driver, re-ran MW single tasks and tasks were slower, 42.2 seconds avg, with one thread for MW and one thread free for the system. I stopped all CPU crunching, and boom, 36.1 seconds per task. So these new tasks are most certainly CPU dependant/bottlenecked. I disabled HT and still about 36.1 seconds.

    Time for some multi-tasks.
     
    #90 TennesseeTony, Mar 31, 2017
    Last edited: Mar 31, 2017
  16. StefanR5R

    StefanR5R Senior member

    Joined:
    Dec 10, 2016
    Messages:
    589
    Likes Received:
    231
    Something isn't right with this F@H score. overclock.net lists R9 290 and 290X at 260 kPPD. (R9 390X is listed with 370 k ppd, but from only two samples, so this means nothing.) 90k ppd is quite exactly what my 1280 CUs wide FirePro W7000 is doing. Therefore those 260 k from overclock.net sound plausible for a GCN card with 2816 CUs.

    Regarding possible driver differences between FirePro and Radeon cards: As far as I tested the W7000, it performed as one would expect from a similar sized Radeon card.

    Is GPU-Z reading the card parameters (CU count, clock, etc.) as they are expected? Also, core utilization during DC jobs would be interesting to keep an eye on, I guess.
     
  17. TennesseeTony

    TennesseeTony Elite Member

    Joined:
    Aug 2, 2003
    Messages:
    2,010
    Likes Received:
    399
    [​IMG]

    GPU-Z isn't able to see the real-time clocks. MSI AfterBurner doesn't show anything at all for it. Any suggestions on other monitoring tools?



    These results are with HT off for now. The card required that I step down to Windows 8.1 for drivers (boo). Running CPU tasks also (Asteroids), as this will give a 'real world' result, but enough CPU cores are free to avoid hitting 100% CPU utilization. The results are......weird.

    3 tasks=108 68 seconds each=100 159 tasks per hour
    6 tasks=130 seconds each=166 tasks per hour
    9 tasks=198 seconds each=163 tasks per hour
    12 task=260 seconds each=166 tasks per hour
    15 task=325 seconds each=166 tasks per hour
    18 task=395 seconds each=164 tasks per hour

    One hour is 3600 seconds, divided by number of seconds per task, multiplied by the number of tasks, right? How'd so many of them end up at 166.15?????? Lol. 2nd cup of coffee may be in order to fix my calculator...or maybe it's right and I just can't believe what I'm seeing.
     
    #92 TennesseeTony, Mar 31, 2017
    Last edited: Mar 31, 2017
  18. crashtech

    crashtech Diamond Member

    Joined:
    Jan 4, 2013
    Messages:
    6,918
    Likes Received:
    320
    Looks like 6 tasks is optimal. My R9 Fury just dies at anything more than 2, so it's cool that your card doesn't.
     
  19. StefanR5R

    StefanR5R Senior member

    Joined:
    Dec 10, 2016
    Messages:
    589
    Likes Received:
    231
    For 0 MHz core clock, the results look already much better. :tonguewink:

    (edit) Would HWiNFO64 be able to show GPU clock?

    I don't know anything about OpenCL or AMD streams programming, but I imagine that those GPGPU feeder processes do a lot of memory copies, for example. So depending on how memory intensive Asteroids is, these CPU tasks could take away more resources than task manager would have one believe (i.e. other resources that task manage simply doesn't monitor).
     
  20. Orange Kid

    Orange Kid Elite Member

    Joined:
    Oct 9, 1999
    Messages:
    3,156
    Likes Received:
    193
    I have a silly question.
    Why 3 tasks or multiples there of? I guess why not 2, 4, 8, or 16 as the shaders and clocks are all even numbers when divided by these. I know I'm missing something here but don't know what. I hope you are understanding my strange reasoning. I don't ever run more then one task at a time, so don't quite get the multiples as the outcome is still about the same.
     
  21. iwajabitw

    iwajabitw Senior member

    Joined:
    Aug 19, 2014
    Messages:
    632
    Likes Received:
    92
    When I built that new system with the dual 280's I had a horrible time with the drivers, then once I got it straight, Win 10 auto updated them for me and screwed it all up again. Had 0 in afterburner and gpuz as well. This seems like a driver issue to me, but I am not familiar enough to lend more support.
     
  22. TennesseeTony

    TennesseeTony Elite Member

    Joined:
    Aug 2, 2003
    Messages:
    2,010
    Likes Received:
    399
    Orange Kid, 3 tasks was always a good balance on the R9-280X, giving good results and still letting me move the mouse cursor smoothly. This card has no video output, so it doesn't make the system sluggish.

    Ok, round two, going up one at a time instead of 3. The first round was just to get an idea of a sweet spot.

    I made a mistake above, "one minute 8 seconds", and I wrote 108. :D Let's make that 68.

    3 tasks=68 seconds each=159 tasks per hour
    4 tasks=82 seconds each=175 tasks per hour
    5 tasks=110 seconds each=163 tasks per hour
    6 tasks=130 seconds each=166 tasks per hour
    7 tasks=149 seconds each=169 tasks per hour
    8 tasks=180 seconds each=160 tasks per hour

    I'll play around with it some more, but so far this card is supposed to be 2.5x more double precision capable than the R9-280X it was supposed to replace, but so far it's only proven to be marginally better. I will likely ask the seller if I can return it.
     
    #97 TennesseeTony, Mar 31, 2017
    Last edited: Mar 31, 2017
  23. iwajabitw

    iwajabitw Senior member

    Joined:
    Aug 19, 2014
    Messages:
    632
    Likes Received:
    92
    Ah, 68 instead of 108, typo of 40secs. Good catch!. I noticed its GFLOPS is 343 currently looking at your MW app details, my 280x's are doing 287GFLOPS currently. So does that translated to the 2.5x more DP from that chip? It could still go higher as the RAC comes up and levels out. I do hate that, that is calculated on a 30 day avg.
     
  24. StefanR5R

    StefanR5R Senior member

    Joined:
    Dec 10, 2016
    Messages:
    589
    Likes Received:
    231
    Did you try uninstalling the Radeon driver before installing the FirePro driver? (The FirePro driver, once installed, will also drive Radeon cards just fine, IME.)

    Also, you could use the DP benchmarks in FAHBench for additional sanity checks.
     
  25. StefanR5R

    StefanR5R Senior member

    Joined:
    Dec 10, 2016
    Messages:
    589
    Likes Received:
    231
    The release notes mention several implementation steps of Hawaii support, but S9150 or W9100 are not explicitly listed.