Could next gen consoles focus mainly on CPU?

Discussion in 'Console Technology' started by Theeoo, Jan 8, 2018.

  1. Mobius1aic

    Mobius1aic Quo vadis?
    Veteran

    Joined:
    Oct 30, 2007
    Messages:
    1,683
    Likes Received:
    259
    Function, do you have a link to those developer docs?

    Trying to prime myself up on 10 year old Terascale architecture and it's descendants. It's well past 2am (got off work at 12am), and my brain is pretty much frazzled now trying to find direct facts on the R700 architecture........ Based on the generally vertical and horizontally mirrored memory layout of each block, I can definitely see it as 4 x 5-ALU shader units, with what I assume is 2KB of local store per cell since R700 shaders have 16KB per 5-ALU shader (64KB per block)? Plenty of articles I've read never provided a real set number of ALUs per block, often resorting to describing the SIMD arrays only, which of course can vary highly depending on GPU and generation. This in turn I guess led to my confusion of thinking the SIMD blocks were always 8 x 5 ALU such as the RV630 being 3 arrays of 40 ALUs each, but R600 being 4 arrays of 80 ALUs. I figured 40 was the minimum per block. Low and behold I just read up on the RV610 being 2 SIMD arrays of 20 ALUs :razz:

    But, I read up on a lengthy NeoGAF thread talking about how the size and layout of the shader cores doesn't easily correspond to other Radeon shader cores either. There's also the claim from the Project Cars developer (he said she said) stating the GPU has up to 192 threads which itself is confusing because R700 is what, either 64 or 128 threads max? Not sure, but ideas from all over the place in terms of customization, including far fetched ones like the shader cores or texture units having embedded GC/Wii TEVs.

    Really just wish someone official would list the specs and be done with it, and of course please continue to grill me with facts so my ignorance is less (I'm not killing, I want the knowledge!). Bed time :sleeping:
     
  2. phoenix_chipset

    Regular Newcomer

    Joined:
    Aug 26, 2016
    Messages:
    546
    Likes Received:
    246
    176 gflops is the only thing that makes sense unless Nintendo says otherwise. Even so I think it's akin to a 768mb fermi @ 176 gflops vs. a 256mb nvidia 8xxx series @ 204 gflops for Xenos. (on paper it's 240 but something about it makes it 204 in practice. Can't quite recall but it was developer comment)

    Looking at latte's die shot it does appear as though the Hollywood gpu is incorporated in there. Doesn't seem far fetched to me ; some part or all of Hollywood has to be in there somewhere.:yep2:

    I don't know how much stock i'd put in the project cars devs comments as their efforts on Wii U are shady at best with their kickstarter shenanigans.
     
    #302 phoenix_chipset, Apr 17, 2018
    Last edited: Apr 17, 2018
  3. Esrever

    Regular Newcomer

    Joined:
    Feb 6, 2013
    Messages:
    759
    Likes Received:
    523
    But then they would have to tell everyone how weak their hardware really is. It's 40 ALUs per block, people who claim otherwise are delusional.
     
    BRiT likes this.
  4. Mobius1aic

    Mobius1aic Quo vadis?
    Veteran

    Joined:
    Oct 30, 2007
    Messages:
    1,683
    Likes Received:
    259
    Xenos is 64 threads max, so the 204 GFLOPS number might come from using all the Vec4 ALUs + a limited number of the Scalar ALUs. 192 GFLOPS is the general "real usage" amount I've heard of however, sticking with the Vec4 ALUs I presume.
    I think that's a given, and the idea of incorporating it's technology into the Wii U's actual graphics hardware is kind of stupid. It's likely just the full Broadway core, re-laid out in a manner we don't readily recognize since we're talking a pretty big chip with so many other components. Not to mention all the custom design work involved would get expensive, and perhaps create issues for porting code from other systems.
     
  5. HBRU

    Regular Newcomer

    Joined:
    Apr 6, 2017
    Messages:
    570
    Likes Received:
    97
    Dont know where to post as beeing an hereticus Jaguar lover I've been banned in the main next gen thread. But wow new chinese amd SOC 400 mm2 vs 232 mm2 of the PRO having for the rest more and less the same teraflops... And this had just 4 Ryzen cores. And nobody says its really a bad result. Ryzens are too big... Waste of silicon.
     
  6. Gubbi

    Veteran

    Joined:
    Feb 8, 2002
    Messages:
    3,580
    Likes Received:
    980
    The quad core complex is 40 mm².

    Cheers
     
    ToTTenTranz likes this.
  7. Pixel

    Veteran Regular

    Joined:
    Sep 16, 2013
    Messages:
    1,000
    Likes Received:
    467
    I wonder if Sony is going to have to be more aggressive with clock speeds next generation to match MS on performance. Has Microsoft given a glimpse of the cooling solution (vapor chamber) and more a aggressive stance on clock speeds they might implement for their mainstream X2 Scarlett console?

    Also Given that the Subour is 1300mhz I wonder if they will employ a vapour chamber cooling solution like the 1172mhz Scorpio GPU.
     
    HBRU likes this.
  8. HBRU

    Regular Newcomer

    Joined:
    Apr 6, 2017
    Messages:
    570
    Likes Received:
    97
    Ok. But then how to they need 400 mm2 ?!? Its huge.... And according to this the price 700 dollars... Curious to see head to head comparision with Ps4 pro
     
    #308 HBRU, Aug 9, 2018
    Last edited: Aug 9, 2018
  9. BRiT

    BRiT Verified (╯°□°)╯
    Moderator Legend Alpha

    Joined:
    Feb 7, 2002
    Messages:
    14,921
    Likes Received:
    13,096
    Location:
    Cleveland
    The GPU takes the rest.
     
  10. HBRU

    Regular Newcomer

    Joined:
    Apr 6, 2017
    Messages:
    570
    Likes Received:
    97
    For having 4 Teraflops on a Vega basis ? Same bus size as pro (256 bits) I guess you are wrong with your estimate Ryzen size...
     
  11. HBRU

    Regular Newcomer

    Joined:
    Apr 6, 2017
    Messages:
    570
    Likes Received:
    97
  12. Gubbi

    Veteran

    Joined:
    Feb 8, 2002
    Messages:
    3,580
    Likes Received:
    980
    But a CCX, a quad core complex, is just 44mm². In the Raven Ridge SKUs the CCX is just 40mm² because of the reduced L3 cache.

    Cheers
     
    BRiT likes this.
  13. Gubbi

    Veteran

    Joined:
    Feb 8, 2002
    Messages:
    3,580
    Likes Received:
    980
    Anyway, where are people getting the 400mm² size from? If I look at the image at AT, the main SOC is between 1.75 to 2.2 times the size of the GDDR packages (depending on perspective shortening). Assuming these are 14x10mm FBGA I get 245 to 280 mm².

    Not unreasonable at all, IMO.

    Cheers
     
    BRiT and TheAlSpark like this.
  14. HBRU

    Regular Newcomer

    Joined:
    Apr 6, 2017
    Messages:
    570
    Likes Received:
    97
    Anadatech sayd that
     
  15. Lalaland

    Regular

    Joined:
    Feb 24, 2013
    Messages:
    596
    Likes Received:
    266
    Despite the focal length of the camera, we can make some estimates about the die size of the silicon. Because of the focal length, we can tell the far memory module is about 74% the size of the near memory, so doing some pixel counting and taking that variance into account, we get a upper limit on die size of 21.31mm x 18.65mm and a die area of 397.5mm2.

    From the AT article, a CCX may be just 40 mm2 but that excludes the uncore and you can't build a CPU or APU without the uncore. Do we have any dieshots for Ryzen that can reliably estimate the uncore area? This is 400mm2 on the current process but iirc someone on here pointed out that there isn't that much of an area shrink from the TSMC 7nano process
     
    BRiT likes this.
  16. Shifty Geezer

    Shifty Geezer uber-Troll!
    Moderator Legend

    Joined:
    Dec 7, 2004
    Messages:
    42,991
    Likes Received:
    15,156
    Location:
    Under my bridge
    That's retail, not BOM. We don't know the business model but, seeing as the PC uses Windows, it seems a give that Subor aren't going to control the software sales, meaning people who buy this machine will be buying games on Steam or paying subscriptions through existing WoW/LoL type portals. If so, the business model is profit on the hardware as a PC company, meaning a notable markup. Cost for the silicon is possibly more than for consoles too because they may not be able to guarantee significant sales, meaning having to negotiate a higher design price and the recoup that through unit sales.

    Basically, lots of variables so comparisons can't be drawn.
     
    Lalaland likes this.
  17. HBRU

    Regular Newcomer

    Joined:
    Apr 6, 2017
    Messages:
    570
    Likes Received:
    97
    Yes of course, not only the die size matters in the cost of 700 dollars... But a 400 mm2 sized die is quite expensive anyway I suppose at 12 nm or 14 nm... And of course beeing it a PC you cannot have Jaguars inside... But really courious to see FPS comparisions with ps4 pro and more with OneX (more software comparable).... I guess will understand why OneX has not Ryzens inside.
     
  18. Michellstar

    Regular Newcomer

    Joined:
    Mar 5, 2013
    Messages:
    652
    Likes Received:
    362
    The ps4pro die is not 232, it's about 320 mm2 (being aprox 350 the X)
     
    Mobius1aic, HBRU and BRiT like this.
  19. Pixel

    Veteran Regular

    Joined:
    Sep 16, 2013
    Messages:
    1,000
    Likes Received:
    467
    Apparently the console version is using an IoT version of Windows. Does the IoT version of windows allow vendors or whatever to lock down the OS and only allow them to install approved software? Maybe they could block admin access to users and remove the browser. Having a custom bios and drivers might mean that even if you formatted and tried to install windows you wouldn't be able to since you don't have access to the custom drivers?
     
  20. Mobius1aic

    Mobius1aic Quo vadis?
    Veteran

    Joined:
    Oct 30, 2007
    Messages:
    1,683
    Likes Received:
    259
    Hope Fenghuang makes it into systems here in the Western world. But I'm gonna bet AMD and Intel have some sort of agreement to not compete with Kabylake-G in certain markets and segments. A version on interposer with HBM would be a KL-G murderer, and even cheaper in total to manufacture. More desireable would be drop-in AM4 socket, which I think is doable but even a BGA config would directly negate KL-G and huge segment of lower end gaming laptops with GTX 1050s.

    If Fenghuang does come to west though, I would want systems with 16 GB. 8 GB isn't really alot when you have Window's crap in the background and games can eat 3 to 4 GB VRAM easy.
     
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...