AMD Ryzen Game Performance Fix?

Discussion in 'AMD Processors' started by SvenBent, Mar 4, 2017.

  1. Ocellaris

    Ocellaris Ginger Ale, an alcoholic's best friend.

    Messages:
    17,239
    Joined:
    Jan 1, 2008
    Where are some of your benchmarks showing any improvement using this? You made a tool and you want other people to experiment with it for you. You would be doing a lot better if you actually provided some performance data instead expecting others to do it for you.
     
    Last edited: Mar 19, 2017
  2. noko

    noko 2[H]4U

    Messages:
    2,125
    Joined:
    Apr 14, 2010
    Is that the correct arrangement for logical and virtual cores confirmed or was that used for a description on the reasoning?
     
  3. SvenBent

    SvenBent [H]ard|Gawd

    Messages:
    1,215
    Joined:
    Sep 13, 2008
    1: Look up
    2: Thats one of the reasons (The other, and more important, is CCX) we have this thread is to get benchmarks on this. So why don you try it out? The software is free or you can do it manually with affinity.


    This has been the correct layout for ALL Intel CPU's I've tested that every 2 logical cores are one physical core but please not i only have access to older generations of Core i7/I5/I3

    I believe it the same way for the CMT desing on the previous generations of AMD CPU's but im unsure on this but i can double check it.
    You would just have to replace Logical core with physical inter gear core. and physical cores with FPU
     
  4. SvenBent

    SvenBent [H]ard|Gawd

    Messages:
    1,215
    Joined:
    Sep 13, 2008
    It is just working on the threads of the main process as it is here most of the load is.
     
  5. noko

    noko 2[H]4U

    Messages:
    2,125
    Joined:
    Apr 14, 2010
    I could test like Cinebench with core 0 and 1 and then compare it to core 1 and core 2 using affinity to confirm. Do that in the morning.
     
  6. SvenBent

    SvenBent [H]ard|Gawd

    Messages:
    1,215
    Joined:
    Sep 13, 2008
    Thats pretty much how i've been doing it. which is why most my testing is with cinebench.
    1: easy to get
    2: distributed load very good among threades (7-zip is not that good at this)
    3: free :D
     
  7. noko

    noko 2[H]4U

    Messages:
    2,125
    Joined:
    Apr 14, 2010
    I got my core numbers wrong, test 0 and 1 (same core) and then 0 and 2 (two different cores).
    I could also compare 6 and 8 and compare it to the test for 0 and 2 to see if going across the two CCXs makes a difference (in this case I would expect a gain due to each having a dedicated L3 cache). But then most work in this case is probably done with the L1 and L2 cache, have to test to see.
     
  8. SvenBent

    SvenBent [H]ard|Gawd

    Messages:
    1,215
    Joined:
    Sep 13, 2008
    For CCX distribution &-zip would probably be best in 2 thread mode. because 7-zip worksin threads pairs. and the 2 threads in a pair are depending on data from thread 1 to thread 2 (one does the data modeling modelling the other does the entropy coding)
    So they should be hit hard once they are not on the same CCX due to all data has go between the threads


    Just remember to do any testing with core parking disabled if you are running windows 7
     
    Last edited: Mar 19, 2017
  9. noko

    noko 2[H]4U

    Messages:
    2,125
    Joined:
    Apr 14, 2010
    Win 10 and core parking in High Performance mode is off, so that is what I will use.
     
    SvenBent likes this.
  10. SvenBent

    SvenBent [H]ard|Gawd

    Messages:
    1,215
    Joined:
    Sep 13, 2008
    Just a small update.
    I've finnally made a more smarter "No SMT conflicts" method. now its no a dumb aprroches (enable or disable it)
    it should now automatically disable when software has enough thread to utilize more than physsical cores. and enable it self when the main process has to few CPU heavy thread.

    I am compiling and uploading tommorow
     
  11. noko

    noko 2[H]4U

    Messages:
    2,125
    Joined:
    Apr 14, 2010
    Well using Cinbench does not work, set Cinbench Affinity for 0 and 1, hit run and it uses all the cores/threads - go back into process affinity and all cores are checked. Looks like the program is automatically configuring the cores/threads. Did this several times, applying closing and reopening affinity which shows core 0 and 1 but as soon as you start to render it uses all the cores.

    Now when I do the single core render from the program - the Task manager shows four cores being used and not just one - weird - something is not right here. I would expect one core to be pegged not four at 20-50%. That is all I can do today.
     
  12. SvenBent

    SvenBent [H]ard|Gawd

    Messages:
    1,215
    Joined:
    Sep 13, 2008
    Sorry that s my fault i completle forgot about this.
    You have to set affinity AFTER you click start render. for some reason cinebench make sure to set affinity to all whenever you hit that start button.
    I had forgotten about this issue.
     
  13. mat9v

    mat9v n00bie

    Messages:
    61
    Joined:
    Dec 23, 2005
    Process Lasso has a function to constantly monitor and correct program affinity, you could try that.
    Forced Mode (continuously reapply settings) in Options.
     
  14. CSI_PC

    CSI_PC [H]ard|Gawd

    Messages:
    1,627
    Joined:
    Apr 3, 2016
    Is Lasso fully working-compatible with Ryzen?
    I read conflicting reports from Anandtech forum but maybe down to user settings *shrug*
    Thanks
     
  15. mat9v

    mat9v n00bie

    Messages:
    61
    Joined:
    Dec 23, 2005
    Worth a try anyway. I think they had problems with bios there so maybe a combination? There were some strange things happening like with affinity set programs were running on only 2 cores or something like that...
     
    CSI_PC likes this.
  16. noko

    noko 2[H]4U

    Messages:
    2,125
    Joined:
    Apr 14, 2010
    I am not confident in trying tweaking programs while still testing out, configuring my new system. Once I stabilized my OCs and settings I will have more leeway in exploring other options.

    I think maybe you might think in getting a RyZen system to really get down into the nit and gritty for your software development. The tool looks very cool but I am not at a point to test it out yet. My earliest estimates would be 2 weeks. I can do tests that are already part of the OS or other software I use, thinking of using maybe Aida64 with the affinity test to verify the core configurations, that gives very precise measurements or data.
     
  17. SvenBent

    SvenBent [H]ard|Gawd

    Messages:
    1,215
    Joined:
    Sep 13, 2008
    Ryzen is definatly going to be in my next build becuase must of my CPU power is beeing utilized for non gaming task. i do a severe among of data pattern recognition and brute force compression optimization. we are talking task thatcan scale to sevelrea hundred threads and easily takes weeks on my current system.
    Sadly my money situation is not for it.
    But i do now thee SMT fixing stuff for intel I3 I5 i7 and xeon are working i've tested that over several dozen systems

    Its more the CCX im interestde in or at leats just which (logical) cores are in which CCX
     
  18. noko

    noko 2[H]4U

    Messages:
    2,125
    Joined:
    Apr 14, 2010
    Well that need will probably grow when the server chips hit with multiple CCXs - so being ahead of game, even if a little could pay off big. At least it will help future developers maybe understand earlier what works better on threading.
     
  19. CSI_PC

    CSI_PC [H]ard|Gawd

    Messages:
    1,627
    Joined:
    Apr 3, 2016
    Can I suggest those looking to test this also consider AoTS.
    The game improves works with SMT in a positive/neutral way and also scales to work well beyond 4 Core CPUs.
    However it is a good example for this program because disabling SMT improves performance by around 10% on Ryzen and also does not scale performance to the 8 primary cores (sensitive to the CCX latency).

    So has a bit of everything to improve.

    Here shows negative performance with SMT and also negative performance scaling beyond 4 Cores (inter-CCX) relative to Intel.

    [​IMG]


    Shows here working positive/neutral with SMT on Intel for 7700K and 6900K

    [​IMG]


    Cheers
     
  20. CVNet1

    CVNet1 [H]ard|Gawd

    Messages:
    1,058
    Joined:
    Oct 10, 2002
    So SvenBent, how many people do you have testing your Ryzen SMT/CCX Affinity forcing program?

    After it is working as desired, do you think it will be mostly automated, or do you expect to have users set profiles for the specific processes that need to be targeted for modified processor affinity settings?
     
  21. toxsickcity

    toxsickcity n00bie

    Messages:
    27
    Joined:
    Mar 22, 2014
    I found a program recently that saves affinity settings and automatically reapplying it every time the app runs

    It simply adds a checkbox to task manager to save settings..

    I forget the name

    I just brought a ryzen and will trial this over next few days
     
  22. toxsickcity

    toxsickcity n00bie

    Messages:
    27
    Joined:
    Mar 22, 2014
  23. KD5ZXG

    KD5ZXG n00bie

    Messages:
    3
    Joined:
    Friday
    Your fixer is a praiseworthy effort.

    I'm having a tough time trying to observe where focused threads go.
    Perhaps when I put task manager in focus, that confuses things???
    I can barely switch fast enough to catch the tail end of what might
    have been happening earlier.

    Can Ryzen fixer be made to ignore task manager in focus? When
    might we experiment with your new Mercury?