AMD: Navi Speculation, Rumours and Discussion [2019]

Discussion in 'Architecture and Products' started by Kaotik, Jan 2, 2019.

  1. chris1515

    Veteran Regular

    Joined:
    Jul 24, 2005
    Messages:
    4,968
    Likes Received:
    4,053
    Location:
    Barcelona Spain


    Very interesting the benchmark Navi 21?

    One of the dev of openVR benchmark said it can't be the laptop with an AMD integrated GPU and a TRX 2060

    Other interesting tidbit
    Very interesting
     
    Lightman likes this.
  2. yuri

    Newcomer

    Joined:
    Jun 2, 2010
    Messages:
    232
    Likes Received:
    203
    This benchmark has been released 20 days ago. Uff...
     
  3. DavidGraham

    Veteran

    Joined:
    Dec 22, 2009
    Messages:
    3,359
    Likes Received:
    3,732
    We shall know soon enough, if big Navi is close by, there will be more relevant leaked benches for it than this.
     
  4. w0lfram

    Newcomer

    Joined:
    Aug 7, 2017
    Messages:
    237
    Likes Received:
    40
    I suspect the rumors & infobits we are getting are somewhat true.
    -505mm^2 rDNA2.0 die
    -7nm+
    -lower clocks
    -2080ti +15%~30% performance

    I would assume, the full die & binned is the 5900 Series, while the cut-downs become the 5800 Series. I think AMD (dr. su) will hit the market with these, at rock bottom prices and sell them enmasse like hotcakes. Leveraging her economy of scale of some 30 wafers @ TSMC on 7nm... and undercut the whole market with a massive price re-alignment. Under AMD's control.

    I don't think either GPU manufacturer is in a hurry, because even people with 1080ti would not bite on shaky RTX series. Turing wasn't enough and this "big navi" might be the price/performance ratio that lands in every gamer's rig. But it does seem, that rDNA2.0 is farther along than Ampere, as such we will see "big navi" 4-5 months before we see anything from Nvidia.

    The reason this is so fascinating is that AMD separated their gaming GPUs, from their enterprise GPUs. rDNA & Vega are going their separate ways and in doing this rdna is more focused on gaming industry and game design.


    With 40% more space, AMD can pack in 50% more performance... shedding GCN.. or as the rumor was heard..
     
  5. PSman1700

    Veteran Newcomer

    Joined:
    Mar 22, 2019
    Messages:
    3,139
    Likes Received:
    1,196
    That's less then Turings die size?
     
  6. ToTTenTranz

    Legend Veteran Subscriber

    Joined:
    Jul 7, 2008
    Messages:
    11,264
    Likes Received:
    5,865
    I don't think AMD will be selling a 500mm^2 GPU that performs 30% above the 2080 Ti at rock bottom prices.
    Not only would that would be detrimental to AMD's Radeon brand, but they now have enough mindshare to get closer to the profit margins practiced by nvidia. Which they already did with the Navi 10 cards.
     
  7. Kaotik

    Kaotik Drunk Member
    Legend

    Joined:
    Apr 16, 2003
    Messages:
    9,238
    Likes Received:
    3,182
    Location:
    Finland
    That's on quite a bit smaller process. For reference, TU106 has 10.8 billion transistors and is 445mm^2 on the TSMC 16/12nm process, while Navi10 has 10,3 billion transistors and 251mm^2 on TSMC 7nm (DUV) process
     
    Lightman, TheAlSpark and PSman1700 like this.
  8. PSman1700

    Veteran Newcomer

    Joined:
    Mar 22, 2019
    Messages:
    3,139
    Likes Received:
    1,196
    The Tensor cores also occupy some space i assume?
     
  9. naenrda

    Joined:
    May 21, 2019
    Messages:
    4
    Likes Received:
    5
    Rock bottom prices? Nope . . . why are people constantly delusional about AMD’s pricing? Selling such a huge N7+ die at rock bottom prices would, frankly speaking, be stupid, and the current AMD is anything but stupid. It wouldn’t work, in the first place.

    The fastest part will be significantly faster than the 2080ti, yes, but priced appropriately! And somehow, people will be shocked, just like after RDNA1.0 cards had their pricing revealed.
     
    nutball and Picao84 like this.
  10. no-X

    Veteran

    Joined:
    May 28, 2005
    Messages:
    2,345
    Likes Received:
    313
    Fractional. There were some analyses and (if I remember correctly) the difference was about 3 %.
     
    PSman1700 likes this.
  11. Benetanegia

    Regular Newcomer

    Joined:
    Sep 4, 2015
    Messages:
    343
    Likes Received:
    308
    No, that was for RT cores. iirc tensor cores was 7% or something like that.
     
  12. DegustatoR

    Veteran

    Joined:
    Mar 12, 2002
    Messages:
    1,664
    Likes Received:
    476
    Location:
    msk.ru/spb.ru
    7% compared to their cut down version in TU11x chips, likely way more than that when compared to their total absence. Tensor cores are kinda complex.
     
  13. iamw

    Newcomer

    Joined:
    Jul 20, 2010
    Messages:
    23
    Likes Received:
    46
  14. Benetanegia

    Regular Newcomer

    Joined:
    Sep 4, 2015
    Messages:
    343
    Likes Received:
    308
    There's no cut down version of Tensor cores in TU11x chips. It simply has FP16 ALUs instead. They should be extremely small and simple. And unlike Tensor cores, they are always going to be there in consumer cards anyways, there's no point in calculating how much area they take.
     
  15. DavidGraham

    Veteran

    Joined:
    Dec 22, 2009
    Messages:
    3,359
    Likes Received:
    3,732
    Wrong news about AFOX EEC database entries for the alleged Radeon RX 5800 (XT), 5900 (XT) and 5950 (XT) in circulation | Ask AFOX

    https://translate.google.com/transl...-xt-und-5950-xt-im-umlauf-nachfrage-bei-afox/
     
    Frenetic Pony likes this.
  16. Frenetic Pony

    Regular Newcomer

    Joined:
    Nov 12, 2011
    Messages:
    550
    Likes Received:
    248
    Hmm, 30% faster, or whatever, than a 2080ti.

    That'd put it about twice as fast a Rx 5700 xt. Just doubling that chip doesn't make a lot of sense, that powerdraw would be catastrophic. Wonder if it's a big RNDA 2.0. 6 functional blocks, 6 WGPs/12Cus a block, 72 Cus or etc. could get there, 24gbs of ram works for high end matching the new consoles too. How far ahead are testable GPUs from releasable GPUs? Less than a year? So big RDNA (unlikely?) or RDNA2 due this year, could well be.

    Also hardly surprised the "5900" doesn't exist. What would it be, a 20% boost over a 5700xt (and what happened to 5800)? Wouldn't be worth taping out and repping an entire new GPU. And the multiple models just doesn't sit with how slowly AMD's been doling out its different bins as of late. The 5600 has waited half a year till after the 5700 series despite being a bin of the same chip.
     
    #1656 Frenetic Pony, Jan 17, 2020
    Last edited: Jan 17, 2020
    DavidGraham likes this.
  17. DegustatoR

    Veteran

    Joined:
    Mar 12, 2002
    Messages:
    1,664
    Likes Received:
    476
    Location:
    msk.ru/spb.ru
    And what do you think tensor cores are in TU10x?
     
  18. Benetanegia

    Regular Newcomer

    Joined:
    Sep 4, 2015
    Messages:
    343
    Likes Received:
    308
    A lot more than just FP16 ALUs hanging from the same register structure as FP32.
     
  19. DegustatoR

    Veteran

    Joined:
    Mar 12, 2002
    Messages:
    1,664
    Likes Received:
    476
    Location:
    msk.ru/spb.ru
    A lot more of what?
     
  20. Benetanegia

    Regular Newcomer

    Joined:
    Sep 4, 2015
    Messages:
    343
    Likes Received:
    308
    Of everything, including many more FP16 ALUs and FP32 accumulator. The specifics are obviously unknown, but there has been discussion in these forums.

    Simply by logic, look at the comparison posted by iamw above. Getting rid of Tensor cores makes the SM go from 1.88mm2 to 1.38mm2 and there's an additional 0.15mm2 reduction on the scheduler. Considering that the 1.38mm2 TU11x SM still holds lots of registers, FP32, INT32, FP64 ALUs and SFUs in addition to the comparably tiny FP16 implementation, how much of that area do you honestly believe you can attribute to the FP16 ALUs? Do you really think it represents any significant* amount of those 1.38mm2 when there's so many other much larger SIMDs in there? And even then, however much you think they are, how much do you think it means compared to the full die? The SMs themselves are only 66mm2 (48 times 1.38) of the 284mm2 TU116 die!!!

    *Significant as in compared to the 0.5mm2 of the Tensor core.
     
Loading...

Share This Page

  • About Us

    Beyond3D has been around for over a decade and prides itself on being the best place on the web for in-depth, technically-driven discussion and analysis of 3D graphics hardware. If you love pixels and transistors, you've come to the right place!

    Beyond3D is proudly published by GPU Tools Ltd.
Loading...