How Many Fragment Pipelines Does G70 Really Have?

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

  1. dizietsma

    Banned

    Joined:
    Mar 1, 2004
    Messages:
    1,172
    Likes Received:
    13
    Unwinder always tells it straight, no bs. If you want to make him really mad ask an easy question about rivatuner when it is in rivatuners online help :)

    rivatuner is a great bit of software for overclockers, so thanks.
     
  2. Carl B

    Carl B Friends call me xbd
    Moderator Legend

    Joined:
    Feb 20, 2005
    Messages:
    6,266
    Likes Received:
    63
    The RivaTuner situation is good evidence to the contrary, but still there is a large die size discrepency between what 302 million transistors should equal and what 334m^2 in fact implies. I have to go with the 'hidden goodies' theory a little bit longer I'm afreaid. :)
     
  3. Geo

    Geo Mostly Harmless
    Legend

    Joined:
    Apr 22, 2002
    Messages:
    9,116
    Likes Received:
    213
    Location:
    Uffda-land
    Conspiracy theories are great entertainment, of course. But reading Unwinder I'm convinced he gave it his all with the best tools & experience available to us for an assault from the direction he attacked. In my mind, the ball is in the court of those who still want to pursue this to bring new facts to the table.
     
  4. _xxx_

    Banned

    Joined:
    Aug 3, 2004
    Messages:
    5,008
    Likes Received:
    86
    Location:
    Stuttgart, Germany
    On the other side, with only one major manufacturer we'd have games 100% optimized for that manufacturer and more "exclusive" features used.
     
  5. martrox

    martrox Old Fart
    Veteran

    Joined:
    Feb 6, 2002
    Messages:
    2,065
    Likes Received:
    16
    Location:
    Jacksonville, Florida USA
    And we would pay through the wazoo......... :roll: :wink:
     
  6. Bouncing Zabaglione Bros.

    Legend

    Joined:
    Jun 24, 2003
    Messages:
    6,363
    Likes Received:
    82
    You'd also see the same slow architechture eeked out for years at a time with minimal improvements at maximum cost to the customer. Developers would not have new tech to use, and when they did, wouldn't use it much because only very few could afford the latest cards.

    Do you think NV40/G70 would be so good with the competition from ATI and the shame of NV30 to drive Nvidia on? Don't you think R520 and R600 are going to have to be very special for ATI to compete with Nvidia?
     
  7. Geo

    Geo Mostly Harmless
    Legend

    Joined:
    Apr 22, 2002
    Messages:
    9,116
    Likes Received:
    213
    Location:
    Uffda-land
    Slightly OT (or at least tangential), I went back and read the original post. . .which basically said yields could be as low as 10-40 dies per wafer if all parts of the chip have to work. And no "7800nu" in sight for those other 130-150 dies. Really, the lack of a 7800nu at launch is as much a bit of evidence as the "secret warehouse where the Ultras live" part of it. If they exist, what are they planning to do with them? If they don't. . .well, that's just not possible unless. . .etc.

    So I see where the suspicion comes from at least.
     
  8. Dave Baumann

    Dave Baumann Gamerscore Wh...
    Moderator Legend

    Joined:
    Jan 29, 2002
    Messages:
    14,079
    Likes Received:
    648
    Location:
    O Canada!
    Well, does Unwinders solution take into account of hardware locking of Quads?
     
  9. Dave Baumann

    Dave Baumann Gamerscore Wh...
    Moderator Legend

    Joined:
    Jan 29, 2002
    Messages:
    14,079
    Likes Received:
    648
    Location:
    O Canada!
    Something to bear in mind that the defects talked of here are going to be random in nature, so they are going to be scattered across wafers in different locations, so generally speaking a wafer would be pretty damned unlucky to get a defect distributed to hit ever core! This has to be taken into account, but with these defects rate yeilds of full core parts could be in the 9-30% range (and thats not taking into account clocks, however it seems as though the clocks may be conservative).
     
  10. Carl B

    Carl B Friends call me xbd
    Moderator Legend

    Joined:
    Feb 20, 2005
    Messages:
    6,266
    Likes Received:
    63
    I think they're going to stick with the single G70 part for a bit, and roll out the full line around the time that ATI brings their R5xx line out. There will be more than one 7000-series part this gen, of this I am sure. The entire pricing structure of the industry is built upon this.
     
  11. Geo

    Geo Mostly Harmless
    Legend

    Joined:
    Apr 22, 2002
    Messages:
    9,116
    Likes Received:
    213
    Location:
    Uffda-land
    Well, let's be generous and give them 40%. That leaves 100 dies per wafer on the order of 6800U performance. . .being thrown away?

    Re "hardware locking". . .what would that look like, I wonder? My baseline understanding is that somewhere in the chip there has to be a "control area" that knows the addressing of all the functional units, and this is what Unwinder is uncovering. It has to exist or you can't talk to the functional units.

    Maybe there is hardware locking going on at a level down from there, and possibly it is more redundancy based? The redundancy explanation has to be taken seriously as well, with possibly just one ps quad and one vs extra. Rys seemed to be going that direction, as I find it easier to believe he saw Michael Jackson in that die photo (finally got it to download) than he really thot he saw one more ps. I would think that financially, there is greater reward to NV from adding another 30% yield (to pick a number out of the air) of 7800gtx than there is in selling Ultras.
     
  12. Carl B

    Carl B Friends call me xbd
    Moderator Legend

    Joined:
    Feb 20, 2005
    Messages:
    6,266
    Likes Received:
    63
    I'm not sure where you're going with that Geo, but if 30% of those processors were indeed 'ultras,' then it would certainly be worth it to launch them as such rather than as GTX's. To begin, the price-point would certainly be higher - giving a better margin to the GTX. On top of that, it would of course have performance ramifications - very important for mindshare.

    'Redundancy' in terms of GPU's means no more or no less than it means for Cell - die aspects that otherwise could be used, planned to be deactivated for yield reasons. But if yields improved there would be no reason to continue in the vein of 'redundant' systems.
     
  13. Geo

    Geo Mostly Harmless
    Legend

    Joined:
    Apr 22, 2002
    Messages:
    9,116
    Likes Received:
    213
    Location:
    Uffda-land
    Well, we all know that digital systems aren't as comfortable with odd numbers. A 7 quad part, while easier to make than an 8 quad part, still greatly increases your yield of 6 quad parts. That's what I meant by the 30%. . .increasing yields of GTX by having a single spare quad & ps at a redundant level (maybe it can have 7 and only address 6 because of the way the redundancy is managed? I'm still struggling for an explanation that integrates Unwinder's findings) is, I'm guessing, more financially rewarding than whatever the market is for a high-priced tiny-yield Ultra.

    Besides, we all look at the clocks/cooling/mem speed on the GTX and we know darn well there's worthy Ultra capability there *without more quads*. If they had spare quads to pull out of their pocket for that differentiation, wouldn't you expect they might have been a little more aggressive with GTX clocks?
     
  14. digitalwanderer

    digitalwanderer Dangerously Mirthful
    Legend

    Joined:
    Feb 19, 2002
    Messages:
    17,230
    Likes Received:
    1,752
    Location:
    Winfield, IN USA
    I like Unwinder too, he doesn't do bullshit. He doesn't give it out and he hates to take it.

    I can respect that. :)
     
  15. Carl B

    Carl B Friends call me xbd
    Moderator Legend

    Joined:
    Feb 20, 2005
    Messages:
    6,266
    Likes Received:
    63
    Hmmm... a single extra quad and vertex eh? Well, I could buy that. It doesn't seem like it would account for all those 'mystery' transistors, but then again doubling down seems like it could consume too many. Definitely a reasonable way of looking at it.

    Well, one way or another, we'll know in the next couple of months. :)
     
  16. Geo

    Geo Mostly Harmless
    Legend

    Joined:
    Apr 22, 2002
    Messages:
    9,116
    Likes Received:
    213
    Location:
    Uffda-land
    Yes, that's some comfort! I've always been irked that we never did get to the bottom of the discrepancy in the counts/sizes between R420 and NV40. :? It seems much more likely this mystery will be a short-lived, one way or the other.
     
  17. Blastman

    Newcomer

    Joined:
    Jul 15, 2003
    Messages:
    176
    Likes Received:
    2
    NV can always down-bin the GTX’s that don’t make it to a 6800U (if it’s pin compatible) or the 7600, whatever that card will look like.
     
  18. trinibwoy

    trinibwoy Meh
    Legend

    Joined:
    Mar 17, 2004
    Messages:
    10,429
    Likes Received:
    430
    Location:
    New York
    Yeah I'm quite sure there will be a 16 pipe G70 based SKU from Nvidia. I wonder what the odds are of defects on a particular chip only hitting the pipes. Would they ever use a 8 pipe G70 on one of their lower SKU's or would they just dump it?
     
  19. dizietsma

    Banned

    Joined:
    Mar 1, 2004
    Messages:
    1,172
    Likes Received:
    13
    It does, well at least on the nv43. Both nv43 rev a2 core and rev a4 core show two pixel units ( quads ), if you try to activate the 2nd for a4 core in rivatuner nothing happens because the quad is hardware locked. I have had both a2 and a4 cores and this is how it behaves, you can still see the quad listed in rivatuner but just cannot do anything about it.
     
  20. Carl B

    Carl B Friends call me xbd
    Moderator Legend

    Joined:
    Feb 20, 2005
    Messages:
    6,266
    Likes Received:
    63
    The odds of it hitting the pipes are pretty good - since the pipes comprise the majority of transistor useage. Of course this thread really centers around the speculation that the GTX itself is already a bineed chip, with defects bringing it from xx-pipes down to 24.

    They'll have to have yet further binned chips down the line though, 0r just a G70 based chip with smaller die and fewer pipes, a la G73.
     
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...