DC performance for Ryzen 3000

Discussion in 'Distributed Computing' started by pututu, Jul 8, 2019.

  1. pututu

    pututu [H]ard DCOTM x2

    Messages:
    1,642
    Joined:
    Dec 27, 2015
    Please post any results, findings, etc pertaining to Ryzen 3000 performance in regards to distributed computing here.

    Would love to see the actual AVX-256 (aka AVX2 in intel chips) performance in DC projects (e.g. Primegrid LLR sub-projects, asteroid) where currently intel chip dominates. There are synthetic benchmark like prime95 but in the reviews that I read they only talk about power consumption, temperature and torture test which is what prime95 is designed for.
     
    Pocatello likes this.
  2. pututu

    pututu [H]ard DCOTM x2

    Messages:
    1,642
    Joined:
    Dec 27, 2015
  3. motqalden

    motqalden [H]ard|DCOTM Feb 2018, June 2019

    Messages:
    1,225
    Joined:
    Jun 22, 2009
    3900x stock profile 2933 32GB ram cas 14

    Milkyway
    13 Jul 2019, 20:32:47 UTC 14 Jul 2019, 2:01:59 UTC Completed and validated 3,144.98 3,135.59 227.52 Milkyway@home Separation v1.46
    windows_x86_64

    Thats about an average result at 52.5 - 53 minutes

    Compared to 1950x stock 32GB cas 14

    13 Jul 2019, 18:14:45 UTC 14 Jul 2019, 1:47:44 UTC Completed and validated 5,168.44 5,139.75 227.51 Milkyway@home Separation v1.46
    windows_x86_64
    Task times around 86 minutes

    thats a pretty big generational improvement.


    will have more results as here soon hopefully
     
    pututu likes this.
  4. motqalden

    motqalden [H]ard|DCOTM Feb 2018, June 2019

    Messages:
    1,225
    Joined:
    Jun 22, 2009
    more to add once i figured out the math here

    3900x 227 points / 53 min = 4.28ppm x24 thread =102.8PPM total rig
    1950x 227 point / 86 min = 2.64ppm x 32 thread =84PPM total rig.
    so the 12 core is 21% faster than the 32 core with HT enabled.
    one more variable is the 1950x is pulling 178w and the 3900x is pulling only 104w as observed by HW monitor. (package power)

    edited for clarity+
     
    Last edited: Jul 13, 2019
  5. VietOZ

    VietOZ n00b

    Messages:
    27
    Joined:
    Apr 12, 2018
    Mine does NFS in 38 minutes vs the 1700 did 50 minutes ... both on Win10
    Same everything, just swapped out the CPU
    about 30% faster runtime and almost double the points (more threads)
    One thing I didn't like it was the temp. Seems to run hot, but I think I know what it is. I'll post again if it's true.
     
  6. motqalden

    motqalden [H]ard|DCOTM Feb 2018, June 2019

    Messages:
    1,225
    Joined:
    Jun 22, 2009
    3900x with HT turned off
    Results for Primegrid
    Proth Prime Search LLR (PPS)

    14 Jul 2019 | 4:07:44 UTC 14 Jul 2019 | 4:38:50 UTC Completed and validated 1,581.02 1,575.42 167.30 PPS (LLR) v8.03
    So 26 min roughly per task.. I have not run this before so i don't know if that's good
     
  7. motqalden

    motqalden [H]ard|DCOTM Feb 2018, June 2019

    Messages:
    1,225
    Joined:
    Jun 22, 2009
    So i was hitting 80c a push pull AIO and i was seeing temps at 80c running prime which is probably worse case scenario. Yes warm and i think it would be too much for the stock cooler or it would throttle some i'm guessing. It does supposedly have a 95c safe temp as listed in Ryzen master. so maybe its not as temp sensitive as earlier gens
     
  8. motqalden

    motqalden [H]ard|DCOTM Feb 2018, June 2019

    Messages:
    1,225
    Joined:
    Jun 22, 2009
    that temp is with -.1 core volt offset btw and the cooler is Arctic Liquid Freezer 240
     
  9. Bill1024

    Bill1024 Gawd

    Messages:
    994
    Joined:
    Apr 26, 2013

    Did you run one task per core and run as many at once? What memory speed and timings?
    If running one task per core, 26m is not really to bad at all.

    My i7-8086k at 4.5ghz HT off no MT, one task per core, has been doing them in 23m36s to 24m30s
    Running 80c-83c with a cust water loop 2 rads, 240mm and a 360mm
    Memory was DDR4 3600 @ 3900 19.20.20.40 T2 Faster memory makes a difference
     
  10. EXT64

    EXT64 [H]ard DCOTM x2

    Messages:
    469
    Joined:
    Mar 27, 2013
    Any primegrid power estimates mot? That is impressive speed. Also, what clocks is it running at during prime?
     
  11. motqalden

    motqalden [H]ard|DCOTM Feb 2018, June 2019

    Messages:
    1,225
    Joined:
    Jun 22, 2009
    So i was running all 12 threads with HT disabled 1 task per thread, and not running any gpu or anything else. Memory speed 2933 cas 14,
    Clocks were hitting around 4.0GHZ but sometimes going slightly below that very briefly. 132 w package power as read by HW monitor.
    This was stock clocks with core voltage offset of -0.075 hitting temps around 81c
     
  12. motqalden

    motqalden [H]ard|DCOTM Feb 2018, June 2019

    Messages:
    1,225
    Joined:
    Jun 22, 2009
    I might be able to do better clocks with the memory as my kit is rated higher, but with 2x 16GB Dual rank sticks 2933 was the supported speed for my mobo. I tried to go 3200 but so far not luck on x470 taichi.
    I plan to try to mess with subtimings and system voltage to see if i can get this higher but i am not holding my breath. It would boot at 3200 cas 14 but would freeze just after getting in bios.
     
  13. motqalden

    motqalden [H]ard|DCOTM Feb 2018, June 2019

    Messages:
    1,225
    Joined:
    Jun 22, 2009
    Thomas Brada Experimental Grid
    3900x
    Average over a page of results. 111 minutes per task windows 10 64

    1950x average 145 minutes per tasks windows 10 64

    3900x = 12.47 Points per minute all threads
    1950x = 12.75 PPM all threads
    Very close victory for the TR
     
  14. RFGuy_KCCO

    RFGuy_KCCO [H]ard DCOTM x3

    Messages:
    862
    Joined:
    Sep 23, 2006
    Linux or Windows? That time is slightly faster than my 9900K@4.7 on Win10.
     
  15. motqalden

    motqalden [H]ard|DCOTM Feb 2018, June 2019

    Messages:
    1,225
    Joined:
    Jun 22, 2009
    Windows
     
  16. motqalden

    motqalden [H]ard|DCOTM Feb 2018, June 2019

    Messages:
    1,225
    Joined:
    Jun 22, 2009
    Seti@home
    Windows 10 64
    hyperthreading on, also running gpu tasks
    Lunatics optimized app for seti
    Stock clocks, good cooling both 32GB ram


    3900x
    Ram 3600 cas 15

    7914839166 3588487697 1 Aug 2019, 18:31:22 UTC 1 Aug 2019, 22:59:57 UTC Completed and validated 3,607.47 3,595.31 73.54 SETI@home v8
    Anonymous platform (CPU)
    3600 seconds or roughly 60 minutes on average.

    1950x
    2933 ram cas 14
    7914593934 3588371632 1 Aug 2019, 16:23:42 UTC 1 Aug 2019, 21:33:19 UTC Completed and validated 6,365.34 5,752.14 69.17 SETI@home v8
    Anonymous platform (CPU)
    Roughly 1 hour 46 minutes per task
     
    pututu likes this.
  17. motqalden

    motqalden [H]ard|DCOTM Feb 2018, June 2019

    Messages:
    1,225
    Joined:
    Jun 22, 2009
    tweaking my ram further and letting things settle in and i am seeing alot of V8 tasks finish in low 50 minutes on the 3900x so i think the average might be closer to 53 minutes
     
    pututu likes this.
  18. motqalden

    motqalden [H]ard|DCOTM Feb 2018, June 2019

    Messages:
    1,225
    Joined:
    Jun 22, 2009
    3900x is pretty much destroying my 1950x on WCG SCC. Here are my recent stats where both rigs were running SCC. the low scoring days were days i was running other projects so ignore that, but where the data points cross is relevant.
    Also if anything my 3900x had more downtime as i was messing around with it alot more.

    Recent production days for 3900x
    3900xWCG.PNG


    VS recent production days for TR 1950x Keep in mind that both machines are only running Stop Childhood Cancer sub project:
    1950xWCG.PNG


    and yes the recent scores have been better with upping my ram clocks. Sure some of this has to do with the way the points are awarded. If you look at the number of results returned they are pretty close.
    So on paper the 3900x is completing about the same number of tasks (which is amazing since its 8 threads less) but because the work units are returned faster i am also getting point bonuses.

    Again this is both systems on windows 10 64 bit running stock clocks.
     
    pututu likes this.
  19. pututu

    pututu [H]ard DCOTM x2

    Messages:
    1,642
    Joined:
    Dec 27, 2015
    SCC will run faster in linux coz it is using Autodock Vina application. 3900x probably consumes a lot less power. Overall this is a great DC chip. Wonder what the 3950x will bring.