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:
    8,141
    Likes Received:
    1,830
    Location:
    Finland
    IIRC it's the "Apple chip" aka Vega with 20 CUs
     
  2. DmitryKo

    Regular

    Joined:
    Feb 26, 2002
    Messages:
    642
    Likes Received:
    485
    Location:
    55°38′33″ N, 37°28′37″ E
    Sure, but did anyone ever try to identify and label out the functional blocks in a Compute Unit on the actual die shot? Things like LDS/registers/TMUs etc.

    Looks almost impossible to me without some inside information, since logic blocks are spread across all silicon-copper layers to maintain power and thermal limits, and the visible layer only reveals parts of these blocks but there is probably not enough information to reconstruct the entire CU .


    The single official photo of a GCN die I could find is a functional diagram in the AMD Polaris whitepaper, with a grid of superimposed green squares standing in for CUs, but these don't even try to match the real Polaris10 die - though it does match what's visible of the real silicon like memory channels and I/O (for the most part, as the right edge is differrent).
     
    #1242 DmitryKo, Jul 16, 2019
    Last edited: Jul 16, 2019
    digitalwanderer, DavidGraham and BRiT like this.
  3. rSkip

    Newcomer

    Joined:
    Jan 10, 2012
    Messages:
    9
    Likes Received:
    17
    Location:
    Shanghai
    http://chipsleuth.com/tahiti.html#annotated-rams-overview
    This guy is great. He's got a Tahiti article focused on CU and a Fiji one about the whole chip.
     
    DmitryKo, digitalwanderer and iamw like this.
  4. ToTTenTranz

    Legend Veteran Subscriber

    Joined:
    Jul 7, 2008
    Messages:
    9,780
    Likes Received:
    4,431
    So apparently the RX 5700XT does ~2150MHz average on 250W by tuning the power tables and placing the vcore at 1.25V (up from 1.2V default):




    In the meantime, with undervolt to 0.966V the card consumes slightly above 130W while clocking 40MHz lower.



    Plus, the drivers still don't allow any memory overclock so there might be some headroom there too.

    It will be interesting to see what the third parties can do with open air coolers and custom voltage tables.
     
  5. fehu

    Veteran Regular

    Joined:
    Nov 15, 2006
    Messages:
    1,436
    Likes Received:
    373
    Location:
    Somewhere over the ocean
     
  6. ToTTenTranz

    Legend Veteran Subscriber

    Joined:
    Jul 7, 2008
    Messages:
    9,780
    Likes Received:
    4,431
    AMD really likes their tradition of needlessly overvolting all their graphics chips out-of-the-box, I guess.

    ¯\_(ツ)_/¯

    At least now they all have the automatic undervolting option in the drivers which when working with Navi should provide a very decent reduction in power consumption.
    Though unless they enable it by default for Navi cards, I wonder how many of the average Joes are ever going to tick that box.
     
    AstuteCobra, BRiT and digitalwanderer like this.
  7. PizzaKoma

    Newcomer

    Joined:
    Apr 29, 2019
    Messages:
    38
    Likes Received:
    59
  8. PizzaKoma

    Newcomer

    Joined:
    Apr 29, 2019
    Messages:
    38
    Likes Received:
    59


    Some driver bugs and the driver doesn't play well with Handbreak (nor x264 GPU accelerated encoding, AMD is working on it since a few days back) atm but other then that it's trading blows with a Nvidia 2080 for Resolve, Premiere, Da Vinci Resolve & Handbrake

    Value for performance is emphasized a lot.
     
    Tarkin1977 and Lightman like this.
  9. DmitryKo

    Regular

    Joined:
    Feb 26, 2002
    Messages:
    642
    Likes Received:
    485
    Location:
    55°38′33″ N, 37°28′37″ E
  10. snarfbot

    Regular Newcomer

    Joined:
    Apr 23, 2007
    Messages:
    532
    Likes Received:
    188
    Ive had a gcn card since 2012 and never even knew this was a thing.
     
    #1250 snarfbot, Jul 17, 2019
    Last edited: Jul 18, 2019
    PizzaKoma likes this.
  11. PizzaKoma

    Newcomer

    Joined:
    Apr 29, 2019
    Messages:
    38
    Likes Received:
    59
    Same :runaway: tested it out though, pretty good for animation, especially anime and fast moving scenes or sports
     
    snarfbot likes this.
  12. 3dilettante

    Legend Alpha

    Joined:
    Sep 15, 2003
    Messages:
    8,118
    Likes Received:
    2,860
    Location:
    Well within 3d
    I've seen various attempts over the years (including the site linked in a later post). I think reverse-engineering companies like Chipworks (now part of Techinsights?) would have more certain breakdowns--but those don't come cheap.
    My running assumption for something like Polaris, based on various attempts at dissecting the CU:
    Starting from the midline of an SE and going out:
    L1, Load/Store, Filtering, SIMD 0 and SIMD1, LDS and scalar, shared front end (scalar cache, instruction cache, etc.), SIMD 2 and 3.
    The supposed branch/message unit might be somewhere in the area the scalar and front end, I'm not sure. There are also things like wavefront instruction buffers whose position isn't clear to me.

    I think the LDS is past the first set of SIMD blocks and before the front end because there's SRAM roughly in the shape and capacity as a register file, but not part of the SIMD blocks.

    Generally a unit of some kind is going to try to have a relatively contiguous 2D footprint (don't want more distance for signal travel if not necessary), although with automated synthesis the exact boundaries get pretty blobby. There's only one layer of active transistors, with the rest being interconnect and elements like power planes and inductors/capacitors. While there are concepts with more than one transistor layer, at least traditionally the electrical properties of layers besides the single layer that was cut from the mono-crystaline ingot have been too poor to be justified with most chips and workflows.

    Manually designed blocks or blocks with some manual guidance or customization tended to be more common in older chips, and so their floorplans and units tended to have much more visible boundaries and paths. It's more difficult now because the paths and units are much smaller relative to the images we have, and without more domain-specific knowledge there are fewer obvious landmarks in the seemingly chaotic synthesized blocks.
     
  13. snarfbot

    Regular Newcomer

    Joined:
    Apr 23, 2007
    Messages:
    532
    Likes Received:
    188
    So out of the box this thing does not ramp the fan over 43% so for my case which is pretty well ventilated i was getting 91 degrees and 110! Junction temp at 100% load. After some tweaking its staying below 90, but imagine it was in a shitty dell pc or something no bueno.
     
    Lightman likes this.
  14. RedVi

    Regular

    Joined:
    Sep 12, 2010
    Messages:
    387
    Likes Received:
    39
    Location:
    Australia
    https://www.tomshardware.com/news/asrock-radeon-rx-5700-challenger-amd-graphics,39967.html

    Asrock show off their custom 5700 cards.

    My main concern with custom cooled cards is whether the heatsink/heatpipes make full contact with the die and aren't just carry-overs designed for nvidias larger dies. This happened with many 290/X cards from vendors that sold both AMD and Nvidia.

    I guess one way to hopefully be sure before reviews are out is to buy Sapphire and the like, AIBs that only sell AMD.
     
  15. CarstenS

    Veteran Subscriber

    Joined:
    May 31, 2002
    Messages:
    4,796
    Likes Received:
    2,054
    Location:
    Germany
    As long as the die is roughly at the center between the mounting screws, I cannot understand how a heatsink's contact area would not make full contact with a smaller die than it was designed for. Contrarily, there would be a larger margin for error on the sides.
     
  16. Psycho

    Regular

    Joined:
    Jun 7, 2008
    Messages:
    745
    Likes Received:
    39
    Location:
    Copenhagen
    no-X likes this.
  17. CarstenS

    Veteran Subscriber

    Joined:
    May 31, 2002
    Messages:
    4,796
    Likes Received:
    2,054
    Location:
    Germany
    Yeah, that's excess heatpipes (the die even seems to touch the three larger heatpipes), so the cooling could be cheaper, but not better with a tailored heatsink. I was under the impression, RedVi was talking about bad cooling performance and thus higher noise levels.
     
  18. Psycho

    Regular

    Joined:
    Jun 7, 2008
    Messages:
    745
    Likes Received:
    39
    Location:
    Copenhagen
    Cooling performance on that card was pretty bad compared to the competition. Not surprising since the thermal connection to the two outer pipes - and corresponding parts of the fins - must be pretty bad. But I agree that if it was all encapsulated in a copper block instead of being direct touch it shouldn't matter too much.
     
  19. RedVi

    Regular

    Joined:
    Sep 12, 2010
    Messages:
    387
    Likes Received:
    39
    Location:
    Australia
    That's the example I was thinking of Psycho. I Hope those designs are a thing of the past as from memory the thermal performance of that Asus 290X was quite bad compared to other designs. I was reminded of it because I have a 290X and because of there again being significant die area difference between AMD and Nvidia for the same tier parts.
     
  20. Kaotik

    Kaotik Drunk Member
    Legend

    Joined:
    Apr 16, 2003
    Messages:
    8,141
    Likes Received:
    1,830
    Location:
    Finland
    Not sure if this should be in here, some Radeon Software thread or where, but this little snippet has been dug out from 19.7.2 drivers:
     
    gamervivek and Lightman like 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...