AMD Ryzen Threadripper 2990WX Gaming Benchmarks Stunted by Faulty Nvidia Driver

Discussion in '[H]ard|OCP Front Page News' started by cageymaru, Aug 15, 2018.

  1. cageymaru

    cageymaru [H]ard|News

    Messages:
    18,631
    Joined:
    Apr 10, 2003
    Golem.de has made a claim that AMD Ryzen Threadripper 2990WX benchmarks are held back by a faulty Nvidia driver. It seems that some games run at half speed on the 32 core beast compared to its 16 core 2950X sibling when running an Nvidia Geforce GTX 1080 Ti GPU. When the Nvidia GPU was switched out for an AMD Radeon RX Vega 64, the frame rate increased substantially and in some cases the frame rate doubled.

    After switching from the Geforce to a Radeon RX Vega 64 (test) , the frame rate rose dramatically in four of the five games tested, sometimes doubling in size. Only Assassin's Creed Origins runs on the Threadripper 2990WX with the AMD card as slow as before. We asked Nvidia if the driver problem is known.
     
  2. IdiotInCharge

    IdiotInCharge Not the Idiot YOU are Looking for

    Messages:
    6,916
    Joined:
    Jun 13, 2003
    Wild- would love to see some locally sourced results ;)
     
  3. Shmee

    Shmee [H]ard|Gawd

    Messages:
    1,103
    Joined:
    Sep 12, 2014
    I am not sure who could do such a thing...
     
  4. notarat

    notarat [H]ard|Gawd

    Messages:
    1,508
    Joined:
    Mar 28, 2010
    FTFY
     
    GDI Lord and John721 like this.
  5. MrDeaf

    MrDeaf Limp Gawd

    Messages:
    337
    Joined:
    Jun 9, 2017
    Interesting.

    I recall there used to be a bug in certain games with Six core Phenom that never occurred under quad and eight cores.
    All was well after the bug got patched.
     
  6. dreadcthulhu

    dreadcthulhu [H]Lite

    Messages:
    73
    Joined:
    Apr 10, 2017
    I am guessing that Nvidia's driver have some sort of bug that shows up with very-high core count CPU's. The Tech Report's recent review of the 2950X also tested a 2990WX locked down to 16 cores; for the games they tested, in 16 core mode the 2990WX was a bit faster than the natively 16 core 2950X, and a lot faster than the 2990WX was with all its cores on.

    https://techreport.com/review/33987/amd-ryzen-threadripper-2950x-cpu-reviewed
     
    Lakados and Vercinaigh like this.
  7. Verado

    Verado [H]Lite

    Messages:
    70
    Joined:
    May 16, 2017
    I think Kyle_Bennett will need more than one 1250W PSU for a 2990WX and a Vega 64 :LOL:
     
  8. Elf_Boy

    Elf_Boy 2[H]4U

    Messages:
    2,133
    Joined:
    Nov 16, 2007
    Did they set the nv driver to multithread in the nv control panel? I am guessing so but still.....
     
  9. gamerk2

    gamerk2 [H]ard|Gawd

    Messages:
    1,293
    Joined:
    Jul 9, 2012
    I know the current driver has some serious issues anyways. Still, shouldn't it be standard practice at this point to test at least one GPU from each vendor when testing game performance, just so things like this get spotted early?
     
  10. dgz

    dgz [H]ardness Supreme

    Messages:
    4,468
    Joined:
    Feb 15, 2010
    What Kyle needs to be doing right now

     
  11. LstOfTheBrunnenG

    LstOfTheBrunnenG [H]ardness Supreme

    Messages:
    7,974
    Joined:
    Jun 3, 2003
    Faulty? Or just sensitive to NUMA?
     
  12. Peter2k

    Peter2k Limp Gawd

    Messages:
    304
    Joined:
    Apr 25, 2016
    Sooo, ist this the "big news" nvidia was advertising?
    Half the speed with an AMD CPU :D
     
  13. renz496

    renz496 Limp Gawd

    Messages:
    185
    Joined:
    Jul 28, 2013
    It seems nvidia really going to need a CPU that work best the way their GPU are designed....they need their own x86 CPU.
     
  14. Wolf_Tech

    Wolf_Tech Limp Gawd

    Messages:
    217
    Joined:
    Sep 19, 2010
    Or AMD has put code in there boards and cpus to slow down nvidia and make amd graphics cards faster on there own platform. Has happened before in this business many times.
     
  15. Brackle

    Brackle Old Timer

    Messages:
    7,288
    Joined:
    Jun 19, 2003
    If that was the case that would make their CPU reviews perform worse. Even AMD knows they aren't top dog when it comes to graphics. It would make no sense for AMD to do that....why you ask?

    When they compare it to their own 1950X with the same Nvidia card, it shows better gaming performance on the 1950X.....Then that would take sales away from the 2990WX.....
     
  16. lostin3d

    lostin3d [H]ard|Gawd

    Messages:
    1,338
    Joined:
    Oct 13, 2016
    Generally Poor Performance. . . .hehe
     
  17. Lakados

    Lakados Gawd

    Messages:
    654
    Joined:
    Feb 3, 2014
    I’m not sure if this is a bug or a selling feature..... Maybe an intended limiter to keep their consumer cards out of servers.
     
  18. mnewxcv

    mnewxcv [H]ardness Supreme

    Messages:
    6,001
    Joined:
    Mar 4, 2007
    I can't see how that would be achieved legally if the card is advertised as performing a certain way.
     
  19. power666

    power666 n00bie

    Messages:
    27
    Joined:
    Jun 23, 2018
    My first guess is that the driver has problems dispatching more than 32 threads simultaneously. There is an old limitation within Windows where an application could not have more than 32 threads active in an application. This has since been hammered out for newer applications and versions of Windows. Unsure of how this works for low level Windows drivers.

    Second guess would be with NUMA and that two of the four dies don't have memory attached. This could be emulated on the 16 core parts: only put memory on channels connected to a single die.
     
  20. Gottfried Leibnizzle

    Gottfried Leibnizzle Limp Gawd

    Messages:
    200
    Joined:
    Apr 29, 2015
    Oi see wot u did thar.
     
    alxlwson and lostin3d like this.
  21. KazeoHin

    KazeoHin [H]ardness Supreme

    Messages:
    7,453
    Joined:
    Sep 7, 2011
    I remember on my old Xeon (28 threads) GTAV wouldn't even load unless I disabled cores. Essentially if the application saw over 16 threads it's index would overflow and not work at all. I'm guessing that Nvidia has that thread # index set to be 5 bits (32 unique numbers) on Geforce drivers.
     
    TitaniumLizzard likes this.
  22. Creig

    Creig Gawd

    Messages:
    750
    Joined:
    Sep 24, 2004
    Reminds me of a few years back with Batmangate. Nvidia tried to claim copyright on standard MSAA coding in Batman : Arkham Asylum. Nvidia had Eidos put in a Hardware Vendor ID lockout that would only allow MSAA to be viewed on Nvidia cards. Funnily enough, there was an "error" in the coding that forced AMD cards to perform part of the MSAA processing anyhow. So even though AMD cards weren't allowed to select that mode, this "Nvidia only" MSAA implementation was actually slowing down AMD cards.
     
  23. Nightfire

    Nightfire Gawd

    Messages:
    848
    Joined:
    Sep 7, 2017

    Wow. Just wow. Most reviewers have been using a flagship nVidia GPU for cpu gaming benchmarks since forever and you think AMD would purposely make their cpus look worse.

    They also have been using overclocked Intel cpus to compare various gpu performance.

    You really didn't think that one through, did you?
     
    DedEmbryonicCe11 and Darth Kyrie like this.
  24. tangoseal

    tangoseal [H]ardness Supreme

    Messages:
    6,255
    Joined:
    Dec 18, 2010
    I knew something was wrong with those benches. There was no way an 80mb cache 32 core 4.0 ghz chip should have been that pathetic in games . Just didn't add up.
     
  25. tangoseal

    tangoseal [H]ardness Supreme

    Messages:
    6,255
    Joined:
    Dec 18, 2010
    Still happened on threadripper. If you set the affinity then problem was solved.
     
  26. IdiotInCharge

    IdiotInCharge Not the Idiot YOU are Looking for

    Messages:
    6,916
    Joined:
    Jun 13, 2003
    A lot of these issues just sound like a programmer took shortcuts, like those that relied on clockspeed (consoles) or framerate (physics in Bethesda games).

    It's the kind of stuff that AMD should have found in testing and brought up to vendors while also providing guidance to reviewers; their lack of communication (if not also engineering effort) has resulted in a poor initial showing, when they could have passed on responsibility to the vendors that can actually fix the problems and let reviewers focus on the stuff that works as intended on release.
     
  27. mnewxcv

    mnewxcv [H]ardness Supreme

    Messages:
    6,001
    Joined:
    Mar 4, 2007
    I think it will work itself out in the coming months. The Ryzen 1st gen launch had a ton of problems as well if you recall.
     
  28. IdiotInCharge

    IdiotInCharge Not the Idiot YOU are Looking for

    Messages:
    6,916
    Joined:
    Jun 13, 2003
    Oh I expect it to- none of these are new problems themselves. Just quite visibly untested (or uncommunicated) issues that at the very least should have been included in review materials, not just to ensure that TR2 was given the fairest shake possible, but to give something for reviewers to test for a baseline that fixes may be compared to when they arrive.

    And at the very least, while these issues shouldn't be unexpected, they certainly could have been already addressed (as they mostly seem to have been in Linux), they do present an opportunity to ensure that codebases are optimized for n-core scaling while accounting for local vs. remote memory access, and that should hold us over until the current computing paradigm changes.
     
  29. Meeho

    Meeho 2[H]4U

    Messages:
    3,893
    Joined:
    Aug 16, 2010
    And you know they haven't?
     
  30. IdiotInCharge

    IdiotInCharge Not the Idiot YOU are Looking for

    Messages:
    6,916
    Joined:
    Jun 13, 2003
    You can point to reviewers that have quoted AMD on the subject in their reviews published when the NDA lifted?

    Seriously, these issues took investigation by the [H] and others- which means that AMD either didn't do the work, or they didn't communicate their results, and that resulted in their product being shown in a lesser light with lesser results that lacked an accompanying 'why'.
     
  31. Meeho

    Meeho 2[H]4U

    Messages:
    3,893
    Joined:
    Aug 16, 2010
    I can't point to anything, I'm asking you to link something that made you state AMD didn't do the work you mentioned.

    They have been doing some mind boggling stupid things that it wouldn't be unlike them, but I would still like a confirmation that they haven't tried to address any of the mentioned issues before the launch.
     
    Darth Kyrie likes this.
  32. Darth Kyrie

    Darth Kyrie n00bie

    Messages:
    14
    Joined:
    Jan 3, 2018
    I didn't realize AMD was supposed to make sure nVidia's drivers work on nVidia GPU's. Silly me, I was under the impression that drivers should be tested by the company that writes the drivers.
     
    Creig likes this.
  33. IdiotInCharge

    IdiotInCharge Not the Idiot YOU are Looking for

    Messages:
    6,916
    Joined:
    Jun 13, 2003
    I stated that either they didn't do the work, or they didn't communicate with reviewers. I'm just as willing to accept AMD being their typical selves when it comes to communication.

    I'd love to see confirmation that they're dialing it in!

    The delta is AMD's CPU. Yeah, they should test it with stuff that it's likely to be used with, and if they ship with discrepancies, they should communicate that so that they're not explicitly or implicitly left with the blame.

    If Nvidia's drivers are a problem, that information should have been part of the reviewers materials.
     
  34. Meeho

    Meeho 2[H]4U

    Messages:
    3,893
    Joined:
    Aug 16, 2010
    True. Sometimes it seems AMD is their own worst enemy.
     
  35. Darth Kyrie

    Darth Kyrie n00bie

    Messages:
    14
    Joined:
    Jan 3, 2018
    According to this post, it's also a problem on Intel CPU's with 64 threads as well. There is something about nVidia drivers that don't like 64 threads.
     
    Brackle likes this.
  36. IdiotInCharge

    IdiotInCharge Not the Idiot YOU are Looking for

    Messages:
    6,916
    Joined:
    Jun 13, 2003
    And?

    Did Intel just release a 'consumer' 64-thread CPU?
     
  37. Darth Kyrie

    Darth Kyrie n00bie

    Messages:
    14
    Joined:
    Jan 3, 2018
    No, they didn't but that doesn't mean shit to you based on your posting history (I have been coming here for almost 3 years now, don't let the fact that I actually registered in January fool you), any opportunity you get to shit all over AMD you take it. Please explain to me how this is AMD's fault that nVidia's consumer GPU drivers don't work on 64 threads but their professional drivers do.
     
    Creig likes this.
  38. Brackle

    Brackle Old Timer

    Messages:
    7,288
    Joined:
    Jun 19, 2003
    So it looks like an Nvidia drivers problem. So much for all the conspiracy people coming in laugh
     
    Creig and Darth Kyrie like this.
  39. IdiotInCharge

    IdiotInCharge Not the Idiot YOU are Looking for

    Messages:
    6,916
    Joined:
    Jun 13, 2003
    Nice, way to get personal :ROFLMAO:

    I run what gets the job done. I've run everything. I have an AMD GPU right next to my 1080Ti, right now. I'd be running Ryzen if AMD were able to keep up with Intel for games.

    And the change here is that AMD released a 64-thread CPU for (very high end) consumer use, and they didn't inform reviewers that Nvidia's driver might not be optimized for that use case, either because they didn't know, or they failed to communicate.
     
  40. Elf_Boy

    Elf_Boy 2[H]4U

    Messages:
    2,133
    Joined:
    Nov 16, 2007
    NV has no reason to make amd look bad, right?

    I don't see it being practical or reasonable to expect amd to have checked each and every hardware and software configuration possible.

    NV released a new driver last week. Should they not have tested?

    It is really easy to point the finger and blame. I would have thought everyone here understands how complex the machines and systems involved are and understands these things happen.

    Was testing some virtual networking software a couple weeks ago so the game master for the 5e group I play with can use fantasy grounds while traveling and in places he can't get a port open.

    After some hours it turned out the reason one of my friends could not get it to work was his c drive was almost full and he installed on another drive.

    Wouldn't work unless on c:

    I really don't see the difference.

    NV driver shortcoming is not amd's fault any more then the virtual tunnel not working was the fault of windows or FG.

    Someone assumed 640k was more then enough at one time too.

    Really don't understand why so many people seem to need to blame and accuse.

    /end soap box rant
     
    Darth Kyrie likes this.