How Many Fragment Pipelines Does G70 Really Have?

Discussion in 'Beyond3D News' started by Dave Baumann, Jun 24, 2005.

  1. Dave Baumann

    Dave Baumann Gamerscore Wh...
    Moderator Legend

    Joined:
    Jan 29, 2002
    Messages:
    14,079
    Likes Received:
    648
    Location:
    O Canada!
    With <a target="_b3dout" href="http://www.beyond3d.com/previews/nvidia/g70/">G70 now fully dissected</a> we know have a little greater understanding of the part. Of course one element that characterises G70 is that of die size: its big.

    Estimates of the die size put it in the order of around 330mm<sup><font size="2">2</font></sup>, which on a 12" wafer would give in the order of around 150-170 full die to the wafer. Now, the issues here are one of yield - the fewer die there are on a wafer the more likely they are prone to defects. An analyst has suggested to us that there may be as many as 130-160 defects per wafer for earlier 110nm lines from TSMC (although the process is now more mature so that is likely to have decreased) which could give could potentially give very low yields for the chips should everyone need to work without any defects at all.

    Now, one thing that we have been used to over the past few years releases is that of utilising defect cores in as lower performance products. The pixel pipelines, followed by the vertex pipelines, are taking the majority of the die space on products such as these and so a natural process defect is most likely to manifest itself in one of these areas; by disabling the pixel quad or vertex shader the defect has manifested itself in the IHV is then still able to recoup much of the cost of that particular piece of silicon, rather than wasting the entire die area.

    The difference with G70 is that it is even larger than previous die sizes, which means that it is quite possible that full product yields may be quite small. Could it be that, in order to alleviate this, the defect rate is already built in to the current GeForce 7800 GTX and 24 fragment shaders is not the full number that G70 supports? Whilst many have speculated on an Ultra variant still to come and have been looking at the overclocks of the GTX’s to speculate where it may hit, perhaps it could be the case that there are 32 fragment pipelines in G70 and the fully working die are quietly being collected for use at a later date?
     
  2. tEd

    tEd Casual Member
    Veteran

    Joined:
    Feb 6, 2002
    Messages:
    2,094
    Likes Received:
    58
    Location:
    switzerland
    wouldn't then rsx also have 32 fragment pipelines and being it in 90nm wouldn't nvidia or better sony want to have all of those enabled
     
  3. cho

    cho
    Regular

    Joined:
    Feb 9, 2002
    Messages:
    416
    Likes Received:
    2
    :lol: so, the g70 is in short supply ?
     
  4. Rockster

    Regular

    Joined:
    Nov 5, 2003
    Messages:
    926
    Likes Received:
    39
    Location:
    On my rock
    That would suggest they also have a pair of vertex shaders disabled as well. Minus a few things here and there that wouldn't need to be duplicated you are essentially talking about 2 NV40's on a single die, which was already over 200 million transistors. ROP's, video encoder, TMDS, 2d areas would have stayed the same but pretty much everything else practically doubled. I doubt 300 million is enough to get that done. Especially considering the additional add units packed in to every new pipe.
     
  5. BrynS

    Regular

    Joined:
    Jul 21, 2003
    Messages:
    406
    Likes Received:
    18
    If additional redundancy (i.e. extra quads that have been disabled) has been built into current 7800GTX products to improve yields, what would the likely Quad/ROP layout comprise? 8/32, 8/24, 8/16 or 6/24? Would the first two even be possible within 300 million transistors?

    How does one test/look for additional quads beyond X-raying the chip or similar.

    Demirug, have your driver/software investigations revealed anything?

    Cheers,

    BrynS
     
  6. phenix

    Regular

    Joined:
    Feb 22, 2003
    Messages:
    620
    Likes Received:
    1
    Location:
    Cambridge, MA
    NV40
    16 ROPs
    16 Fragment Shader pipes
    6 Vertex Shader pipes

    222 millions transistors
    287 mm2
    130nm

    G70
    16 ROPs
    24 Fragment Shader pipes
    8 Vertex Shader pipes

    300 millions transistors (if true)
    334 mm2
    110nm

    G70-NV40= 78M trans. = 8 FS + 2 VS pipes

    I think it should not be all that difficult (at least for ATI) to figure how many pipes there should be on G70.


    PS. Can it be that the 300 million figure is not true? I feel that this size of chip should acommodate at least 350M transistors considering the optical shrink and die size increase together..
     
  7. Dave Baumann

    Dave Baumann Gamerscore Wh...
    Moderator Legend

    Joined:
    Jan 29, 2002
    Messages:
    14,079
    Likes Received:
    648
    Location:
    O Canada!
    Did you read the post?
     
  8. pocketmoon66

    Newcomer

    Joined:
    Mar 31, 2004
    Messages:
    163
    Likes Received:
    9
    No, only the 32 pipe G70 :)

    Can you image that in SLI with some OC'ing :)
     
  9. cninja

    Newcomer

    Joined:
    Jun 24, 2005
    Messages:
    1
    Likes Received:
    0
    To add some credence to the speculation: when the NV40 came out, there were die lithograph photos that were passed around, but not any for the G70.

    ..At least I think so... A quick Google search brought up this article:
    www.hothardware.com/hh_files/S&amp;V/nv40_debut.shtml

    If they did have some super-secret extra shaders lying under the heat spreader, that would explain the lack of photos, since they would be easy to count.
     
  10. digitalwanderer

    digitalwanderer Dangerously Mirthful
    Legend

    Joined:
    Feb 19, 2002
    Messages:
    17,217
    Likes Received:
    1,736
    Location:
    Winfield, IN USA
    I think your theory is bang-on Dave, it does explain why they came out with their mid-tier high-end card first as well as why they have such an abundant supply of them.

    I expect the G70 Ultra will be a whole lot more than just a speed bumped G70.
     
  11. phenix

    Regular

    Joined:
    Feb 22, 2003
    Messages:
    620
    Likes Received:
    1
    Location:
    Cambridge, MA
    My question is who are they fooling if this is the case? ATI would be the first one to spot such a trick no doubt. Are they intending to sell faulty G70s for 600 bucks while waiting for the process to improve until they get good yields of real G70s?
     
  12. digitalwanderer

    digitalwanderer Dangerously Mirthful
    Legend

    Joined:
    Feb 19, 2002
    Messages:
    17,217
    Likes Received:
    1,736
    Location:
    Winfield, IN USA
    Who says they're fooling anyone? They're selling a product and if they're describing what they're selling accurately then there ain't no "fooling" going on.

    If other G70 chips are more capable doesn't make these "crippled" per se, it's just how binning works.
     
  13. phenix

    Regular

    Joined:
    Feb 22, 2003
    Messages:
    620
    Likes Received:
    1
    Location:
    Cambridge, MA
    I didn't mean that they are cheating customers or something. I meant that maybe they are planning a surprise attack on R520 when finally it launches. That kind of fooling. :wink:
     
  14. Tim Murray

    Tim Murray the Windom Earle of mobile SOCs
    Veteran

    Joined:
    May 25, 2003
    Messages:
    3,278
    Likes Received:
    66
    Location:
    Mountain View, CA
    If this is true, you're an IDIOT to buy a 7800 GTX now, since it will drop to $400 or so (I HOPE, oh God I hope) when the 32-pipe version comes out...
     
  15. Maintank

    Regular

    Joined:
    Apr 13, 2004
    Messages:
    463
    Likes Received:
    2
    It wouldnt surprise me they would do this for a couple of reasons.

    For the size of the die 8 pipelines are probably not going to make or break it. It allows them a margin of error is a quad or two didnt make it out of the fab. So this can help increase yields a little.

    Secondly by having that margin of error it allowed them to ensure a supply of chips. This also forces ATI to show their hand with the R520.

    If the R520 comes out with 32 pipes Nvidia has a quick backup plan.
     
  16. Bouncing Zabaglione Bros.

    Legend

    Joined:
    Jun 24, 2003
    Messages:
    6,363
    Likes Received:
    82
    Won't that upset a lot of cutting edge early adopters? Especially as Nvidia have already said "there is no ultra". Unless Nvidia think that R520 will be really late and can use a 32 pipe ultra as a refresh in Spring 2006 rather than a response to R520 in a few months.

    However, I think it's unlikely given the .11 process and the way the G70's 24 pipes have been counted - will 32 pipes really be 32 pipes?

    Personally I think it's the memory bottleneck that will really make the difference. If one company or the other can overcome that problem, it will give a clear winner.
     
  17. Mindriot

    Newcomer

    Joined:
    Jun 3, 2003
    Messages:
    73
    Likes Received:
    0
    There may be no Ultra, but whats stopping them from labeling it a 7900?

    afterthought:

    With 32 pipes, would they be creating a bottleneck with only 16 ROP's? Or is that still sufficient to keep things moving?
     
  18. trinibwoy

    trinibwoy Meh
    Legend

    Joined:
    Mar 17, 2004
    Messages:
    10,428
    Likes Received:
    425
    Location:
    New York
    Well there is no "ultra" until there is an "ultra" so Nvidia technically isn't lying. I don't think they ever said that there will never be an ultra.

    Agreed. It would be nice to know what yields on the G70 look like though. What are the odds of a straight 24-pipe 300M+ transistor chip on TSMC's .11nm process having "good" yields? Wasn't there a patent out there about adding a single extra pipeline for redundancy or something like that - maybe they don't have 2 full extra quads but maybe just a few extra pipes for backup.
     
  19. wireframe

    Veteran

    Joined:
    Jul 14, 2004
    Messages:
    1,347
    Likes Received:
    33
    I don't see how you work this one out. If future products are stronger, buying one today (top of the line, no less) is a bad idea? If this is true now then it must have been true before and everyone who owns a top of the line card is an idiot because there will always be some better product in the future. Does not compute. And of course the 7800 GTX is likely to drop in price in the future. This has also always been true (with some exceptions).

    However, if it indeed is the case that the G70 does have 32 fragment pipes, it is a brilliant move by Nvidia. Good yields now and time to stock up on some good cores for the "ultra" while they wait for faster memory. Sounds a bit toogood to be true, if you ask me. Especially if you bake in the rumored problems ATI have been having getting good yields on R520. If there is any truth to this situation, ATI could be in for a really nasty battle. One they may do well on in terms of performance, but that may cost them dearly.
     
  20. digitalwanderer

    digitalwanderer Dangerously Mirthful
    Legend

    Joined:
    Feb 19, 2002
    Messages:
    17,217
    Likes Received:
    1,736
    Location:
    Winfield, IN USA
    [​IMG][​IMG][​IMG]
    No, no...the 7800 GTXU! :lol:

    I agree with you Baron, but I was thinking of more of a $300-350 pricepoint for some reason.

    Also, if Wavey is right isn't there a chance that some bright puppy out there might figure out a way to enable the disabled bits? Me loves 9500/9700 X800 pro/X800 XT tricks and such... ;)
     
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...