AMD Could Do DLSS Alternative with Radeon VII through DirectML API

Discussion in 'Frontpage news' started by Hilbert Hagedoorn, Jan 17, 2019.

  1. Hilbert Hagedoorn

    Hilbert Hagedoorn Don Vito Corleone Staff Member

    Messages:
    48,398
    Likes Received:
    18,573
    GPU:
    AMD | NVIDIA
  2. fantaskarsef

    fantaskarsef Ancient Guru

    Messages:
    15,698
    Likes Received:
    9,575
    GPU:
    4090@H2O
    Well... same as with DLSS: make it work, dear driver teams / devs. Before we see it working, this is nice to know but... not much more besides that.
     
    -Tj- likes this.
  3. sverek

    sverek Guest

    Messages:
    6,069
    Likes Received:
    2,975
    GPU:
    NOVIDIA -0.5GB
    It's Gsync and Freesync scenario all over again.
     
    Neo Cyrus likes this.
  4. Valken

    Valken Ancient Guru

    Messages:
    2,922
    Likes Received:
    903
    GPU:
    Forsa 1060 3GB Temp GPU
    Thanks for posting this Hilbert and especially for that hires die shot!

    Can I offer a suggestion? Can you embed the guru3d.com watermark like an embossed engraving in your future pictures? I think it would look much classier since I do use your pictures as wallpaper.
     

  5. fantaskarsef

    fantaskarsef Ancient Guru

    Messages:
    15,698
    Likes Received:
    9,575
    GPU:
    4090@H2O
    That picture in the article even has a watermark :D
     
  6. 999Anticlock9wiSe

    999Anticlock9wiSe Guest

    Messages:
    21
    Likes Received:
    6
    GPU:
    AMD
    problem with this tech is you have to bake it so procedurally generated graphics may not work.
    I do think its a step backwards not forwards because of its limitations .
     
  7. Hilbert Hagedoorn

    Hilbert Hagedoorn Don Vito Corleone Staff Member

    Messages:
    48,398
    Likes Received:
    18,573
    GPU:
    AMD | NVIDIA

    I can look into that sure, btw for some high-res VII wallpaper, click here.
     
    Keitosha, Embra, -Tj- and 3 others like this.
  8. labidas

    labidas Guest

    Messages:
    328
    Likes Received:
    119
    GPU:
    R9 290 / RX5700XT
    Nvidia be like: "You need our PhysX chip, CUDA cores, tensor cores, Gsync module for those things!"
    AMD be like: "Hold my beer..."
     
  9. Valken

    Valken Ancient Guru

    Messages:
    2,922
    Likes Received:
    903
    GPU:
    Forsa 1060 3GB Temp GPU
    Thanks Hilbert! That nerd pr0n just made my day! :p
     
  10. fantaskarsef

    fantaskarsef Ancient Guru

    Messages:
    15,698
    Likes Received:
    9,575
    GPU:
    4090@H2O
    I'm not sure I can follow you, what do you mean by "bake" it?


    AMD can do it with brute force, Nvidia tries to be "smarter" here and run it via specialised hardware. If you read Hilbert's article, it might do just the same on AMD's hardware, just might need more horsepower to do so.
     

  11. cryohellinc

    cryohellinc Ancient Guru

    Messages:
    3,535
    Likes Received:
    2,974
    GPU:
    RX 6750XT/ MAC M1
    Those are some great news, and that's a sexy shot right there. "whistles" ;)

    Thank you for sharing @Hilbert Hagedoorn
     
  12. RooiKreef

    RooiKreef Guest

    Messages:
    410
    Likes Received:
    51
    GPU:
    MSI RTX3080 Ventus
    Lol! That’s a good one.
     
  13. Fediuld

    Fediuld Master Guru

    Messages:
    773
    Likes Received:
    452
    GPU:
    AMD 5700XT AE
    Which Vega (both 10 & 20) has that horsepower to do DirectML it via GPGPU as per AMD when asked about this. Also on the same answer AMD said that RVII is 62-65% faster than the RTX2080 in Luxmark ray tracing benchmark, Which is using OpenCL based ray tracing. (Vega 64 is also 3% faster than RTX2080 on this benchmark).
     
  14. HWgeek

    HWgeek Guest

    Messages:
    441
    Likes Received:
    315
    GPU:
    Gigabyte 6200 Turbo Fotce @500/600 8x1p
  15. Denial

    Denial Ancient Guru

    Messages:
    14,206
    Likes Received:
    4,118
    GPU:
    EVGA RTX 3080
    DLSS is free. Turing still has an advantage because of the tensors, which DirectML can leverage.
     
    yasamoka likes this.

  16. dr_rus

    dr_rus Ancient Guru

    Messages:
    3,886
    Likes Received:
    1,015
    GPU:
    RTX 4090
    DirectML is an API. DLSS can work via DirectML as well once it will actually be available. The bulk of effort however is in developing DLSS itself, not in choosing an API which is will run through. DLSS running via NV's own NGX right now doesn't cost anything to end user.

    What AMD will actually need though to make something like DLSS work are tensor cores. Which none of their GPUs have right now.

    All Turing cards support double rate FP16 on main SIMDs.
     
  17. BlackZero

    BlackZero Guest

    Similar to how A.I. centric game engines need set boundaries, that would be a limitation of hardware and not necessarily ray tracing implementations. Of course, procedurally generated content could it self contain specific boundaries, negating any disadvantages.

    However, I do think, it could then become a question of artistic style vs bland environments.
     
    Last edited by a moderator: Jan 17, 2019
  18. fantaskarsef

    fantaskarsef Ancient Guru

    Messages:
    15,698
    Likes Received:
    9,575
    GPU:
    4090@H2O
    None of what you wrote is wrong. Only that it's not what it's about... especially not in this thread. We're talking about AI algorythm execution, DLSS has nothing to do with ray tracing, so your benchmark shows little in terms of comparable performance or anything in general about DLSS.
    Also, why should any 2080 user run OpenCL ray tracing in games when they have DXR / RTX... that simply doesn't make any sense as of now. Sure, if at some point OpenCL RT is the thing, Nvidia will have a problem (as it has happened with some things in the past between red and green too), but right now, this benchmark is useless besides being able to brag about it, or in professional environments (where I guess it will be more of a matchup between Nvidia's and AMD's professional cards, in which you'd have to compare Vega2's daddy to a Titan, not the 2080).
    Like @dr_rus said, AMD has to "invent" a way to do DLSS first, this probably takes quite some time, then hast to work together with the game devs to test it, then hast the gamedevs to submit their game, they push it through it's AI training parcours on the big computers, then AMD is where Nvidia says they're right now. So... like I said, right now, nice to know, but this news article probably only shows it's significance in half a year from now or later.
     
    BlackZero likes this.
  19. Wrinkly

    Wrinkly Active Member

    Messages:
    73
    Likes Received:
    44
    GPU:
    Radeon 290 / 4GB
    DLSS is free, Tensor cores however are not :p
     
    holler likes this.
  20. xrodney

    xrodney Master Guru

    Messages:
    368
    Likes Received:
    68
    GPU:
    Saphire 7900 XTX
    That's questionable... because when not using DLSS or RTX, RT cores are worthless but at same time taking A LOT of space on die.
    There are not many ways to test DLSS so far but from what GN tested in FF XV, to be honest DLSS looked considerably worse than native 4k with TAA.
     

Share This Page