AMD: Navi Speculation, Rumours and Discussion [2019]

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

  1. Kaotik

    Kaotik Drunk Member
    Legend

    Joined:
    Apr 16, 2003
    Messages:
    8,254
    Likes Received:
    1,940
    Location:
    Finland
  2. Leovinus

    Newcomer

    Joined:
    May 31, 2019
    Messages:
    31
    Likes Received:
    12
    Location:
    Sweden
    https://www.anandtech.com/show/15104/amd-adds-radeon-rx-5300m-to-mobile-gpu-lineup

    Ryan made a nice writeup. It seems the PRO SKU's differ a little bit as per usual. 5300M PRO featuring 22 CU's instead of 20 but with a cut down memory bandwidth from 128 to 96 with a maximum of 3 GB of VRAM. For what purpose though? What's the benefit?

    I've been looking forward to this for a long time. Might be time to upgrade my MacBook at long last. My little trooper from 2011 needs retirement. I had been looking casually at the Vega 20 specced MacBook Pro but balked at the price for that option. The new 5300M PRO of the "standard" specification seems to be stronger still for a much more agreeable sum.
     
  3. no-X

    Veteran

    Joined:
    May 28, 2005
    Messages:
    2,300
    Likes Received:
    255
    Radeon Pro 5300M has 20 CUs + 128bit bus, while Radeon RX 5300M has 22 CUs + 96bit bus. And Radeon Pro 5500M has 24 CUs at up-to 1300 MHz, while Radeon RX 5500M has 22 CUs at up-to 1645 MHz. Pro SKUs are power/performance optimized, non-Pro SKUs are price/performance optimized. Different configurations allows to sell every manufactured Navi 14 GPU (there are models with disabled CUs, with disabled ROPs and with disabled memory channel, so many different defects can be "cured"). Apple seems to get exclusively fully functional GPUs with 24 CUs (like they got Polaris 11 with all 16 CUs while PC segment got only 14 CUs).
     
    BRiT and Leovinus like this.
  4. ToTTenTranz

    Legend Veteran Subscriber

    Joined:
    Jul 7, 2008
    Messages:
    10,074
    Likes Received:
    4,651
    Save money on a lower-end SKU by putting one less 8Gbit GDDR6 chip.
     
  5. Frenetic Pony

    Regular Newcomer

    Joined:
    Nov 12, 2011
    Messages:
    346
    Likes Received:
    89
    Huh, the ability to cut down on memory bandwidth is interesting, how do they do that exactly? The functional blocks were represented as 64bit buses apiece, but it always looked like disabling one would cut bandwidth to an entire block of WGPs.

    Obviously it's possible with current RDNA some way or another. Could mean we'll see further binned RDNA cards in the future. RX5600 with 192bit bus/6gb of ram for $200 or something. Or maybe they'll wait for RDNA 2 and 6XXX cards, 12gb of ram would probably be easily within min spec for next gen games (maybe?)
     
  6. Kaotik

    Kaotik Drunk Member
    Legend

    Joined:
    Apr 16, 2003
    Messages:
    8,254
    Likes Received:
    1,940
    Location:
    Finland
    Even though the simplified block diagrams talk about "64-bit memory controllers", they're actually 16-bit if I'm not mistaken, as seen in this multilevel cache hierarchy slide, there's 16 memory controllers which translates to 16bit controllers, since 16x16 = 256bit
    upload_2019-11-15_11-45-9.png

    edit:
    Isn't this dictated by GDDR6 anyway? Each GDDR6 chip has two independent 16-bit channels
    (late edit: fixed stupid mistake, slide of course, not die :D )
     
    #1506 Kaotik, Nov 15, 2019
    Last edited: Nov 16, 2019
  7. ToTTenTranz

    Legend Veteran Subscriber

    Joined:
    Jul 7, 2008
    Messages:
    10,074
    Likes Received:
    4,651
    Can they decouple the 2*16bit channels in a single GDDR6 module though? I'd assume they split the 32bit channel for electrical reasons.

    I'm just wondering if they could achieve a bandwidth granularity below 32bit.
     
    Frenetic Pony likes this.
  8. Kaotik

    Kaotik Drunk Member
    Legend

    Joined:
    Apr 16, 2003
    Messages:
    8,254
    Likes Received:
    1,940
    Location:
    Finland
    As far as I know (and a quick google run), they are truly completely independent channels with their own command and address & data buses, as in, "2 memory chips in 1"
     
    Frenetic Pony likes this.
  9. Frenetic Pony

    Regular Newcomer

    Joined:
    Nov 12, 2011
    Messages:
    346
    Likes Received:
    89
    GDDR6 specifies 16bit channels so it certainly seems possible from that perspective.
     
  10. Tofu

    Newcomer

    Joined:
    Mar 23, 2012
    Messages:
    18
    Likes Received:
    3
  11. TheAlSpark

    TheAlSpark Moderator
    Moderator Legend

    Joined:
    Feb 29, 2004
    Messages:
    20,814
    Likes Received:
    5,916
    Location:
    ಠ_ಠ
  12. Kaotik

    Kaotik Drunk Member
    Legend

    Joined:
    Apr 16, 2003
    Messages:
    8,254
    Likes Received:
    1,940
    Location:
    Finland
  13. Leovinus

    Newcomer

    Joined:
    May 31, 2019
    Messages:
    31
    Likes Received:
    12
    Location:
    Sweden
    Pardon my ignorance, clamshell design?
     
  14. Kaotik

    Kaotik Drunk Member
    Legend

    Joined:
    Apr 16, 2003
    Messages:
    8,254
    Likes Received:
    1,940
    Location:
    Finland
    In clamshell mode two memory chips use half width datapath (so 8bit in case of gddr6) and shared address/command path connected to same 16bit memory controller. (So you get twice the capacity with same bandwidth)
     
    TheAlSpark and Leovinus like this.
  15. TheAlSpark

    TheAlSpark Moderator
    Moderator Legend

    Joined:
    Feb 29, 2004
    Messages:
    20,814
    Likes Received:
    5,916
    Location:
    ಠ_ಠ
    Or they just have a really big Navi. ¯\_(ツ)_/¯
     
  16. Leovinus

    Newcomer

    Joined:
    May 31, 2019
    Messages:
    31
    Likes Received:
    12
    Location:
    Sweden
    If so I would assume its on the RDNA 2 architecture. I hope anyway. It just doesn't seem financially sound to release a high end, non RT, part if A) It's immediately hampered by lack of RT against the current competition and B) facing imminent internal competition with the release of RDNA 2 with RT and C) possibly a new nVidia arch with even stronger RT chops than Turing soon~ish.
     
    PSman1700 likes this.
  17. Kaotik

    Kaotik Drunk Member
    Legend

    Joined:
    Apr 16, 2003
    Messages:
    8,254
    Likes Received:
    1,940
    Location:
    Finland
    Supposedly it's 16GB version of Radeon Pro W5700 which just got announced as 8GB variant, but of course it could be "big Navi" too.
    https://www.amd.com/en/press-releas...7nm-professional-pc-workstation-graphics-card
     
    TheAlSpark likes this.
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...