G70, G71, R520, R580 die size discrepancy

trinibwoy said:
Hmmmm, interesting. If that was the case though, why not leave the extra quad in there for G71? It would still have ended up pretty small. And if you're dumping ~30M transistors, why not use it for something else - like HDR+AA or better AF?

Ripping stuff out doesn't take much engineering time. Adding new features in does.
 
geo said:
That would put them at about 58% increase in density moving 110nm to 90nm, taking both the 334mm2 and 196mm2 numbers seriously.

I'm tempted to go look at Wavey's "hidden quads" thread and see how many extra transistors he calculated G70 had. :D


But Isnt that with extra Quad added while you need another dispatch processor to keep 1:1 Ratio ????which I think 2XM is impossible to do so.
 
Last edited by a moderator:
geo said:
Ripping stuff out doesn't take much engineering time. Adding new features in does.

They've had just under a year. If this baby is really just a die-shrink with a redundant quad removed and absolutely no improvements architecturally, then Nvidia is either very lazy or has been putting in some serious time with G80. It's kind of analogous to ATi's squeezing of R300, while designing R580/Xenos.
 
trinibwoy said:
They've had just under a year. If this baby is really just a die-shrink with a redundant quad removed and absolutely no improvements architecturally, then Nvidia is either very lazy or has been putting in some serious time with G80. It's kind of analogous to ATi's squeezing of R300, while designing R580/Xenos.

I suspect you think pretty hard about the refresh at the same time as you think about it's older brother, so I'd say more than a year.

And, yeah, the R3xx comparisons looks pretty apt.

But I'd never accuse either IHV of having lazy engineers. I'm sure they've been working hard at something or other, probably multiple somethings, and hopefully G80 is well up that list.
 
trinibwoy said:
Hmmmm, interesting. If that was the case though, why not leave the extra quad in there for G71? It would still have ended up pretty small. And if you're dumping ~30M transistors, why not use it for something else - like HDR+AA or better AF?

Yeah this is a definite unexpected dump of transistors. Well if its a reduntant quad then maybe nV feels that cost of the extra quad vs. yields aren't going to be a factor in the g71 which might have been a factor in the g70.

About new features, lets wait and see ;) . HDR + AA shouldn't increase transistors much, I'm thinking is more of a modification of nV's ROPs which should do the trick.
 
Razor1 said:
HDR + AA shouldn't increase transistors much, I'm thinking is more of a modification of nV's ROPs which should do the trick.

Hmmmm I gathered it would as according to David Kirk:

David Kirk Bit-Tech interview, July 11, 2005


Quote:

"It would be expensive for us to try and do it in hardware, and it wouldn't really make sense - it doesn't make sense, going into the future, for us to keep applying AA at the hardware level. What will happen is that as games are created for HDR, AA will be done in-engine according to the specification of the developer."

And when I hear expensive in regaruds to talking about features in ICs it usally mean using too much of their transistor budget... but meh :)
 
jb said:
Hmmmm I gathered it would as according to David Kirk:



And when I hear expensive in regaruds to talking about features in ICs it usally mean using too much of their transistor budget... but meh :)

That could have just been a smoke screen, ATi sure didn't use that much transistors for this feature did they? We will know by tomorrow.

Edit, also nV already has SSAA+HDR so its kinda already there.
 
Last edited by a moderator:
dizietsma said:
I measured it at 180-220mm in the earlier thread with it likely to be the lower end of that range. I thought it too low then, but now I think it was spot on ..we shall see !

My initial measurements was 250mm^2 using diagonal heatsink mount holes. When I use
PCI-E, or horizontal references, I get like ~ 220mm^2. That is when I discover the aspect ratio of photos may not be 1:1 since B3D tends to be wider than tall while others are
square and that picture is taller than wider. (at angle as well) Compensating for this and assuming that actual package is square, I get between 200-220mm^2. Nvidia slide show it as 186mm^2 but it is fuzzy, so I use the more conservative 196mm^2.

It is easy to assume that Nvidia marketing would not mislead on die size since it is easily verified by others. However G73 (the one widely assumed to be G73 die), the comparison is much easier since a GDDR3 chip is nearby, so that should be 150mm^2.
Couldn't be wrong on that estimate unless that is not a G73 /some photoshop
image pasted onto card. I am inclined to believe that it is a G73 since 150mm^2 is the "sweet" die size for a $149 card. Both Nvidia and ATi target this size (NV43, RV410, RV530)

My measurement error on G71 could be high enough that 196mm^2 is the correct size but if Nvidia is misleading us on G73 125mm^2, what is another lie on same marketing
slide?
 
Razor1 said:
Edit, also nV already has SSAA+HDR so its kinda already there.
I wouldn't exactly call that hardware support. On everything except R5xx, you need to render at a higher resolution and then downsample for FSAA with an FP16 framebuffer.

More importantly, the lack of rotated grid will make SSAA's visual improvement rather minimal (except in cases of shader antialiasing), so the performance impact will be very unappealing to most people unless the application is pushing the video card very lightly, which is unlikely for FP16 HDR.
 
Jaws said:
The PR doesn't specify MSAA or SSAA. Do you have a link that specifies? Thanks...
If RGSSAA was supported on G70, you'd think somebody would have exposed it by now. G71 is highly unlikely to be different in this respect.
 
EasyRaider said:
If RGSSAA was supported on G70, you'd think somebody would have exposed it by now. G71 is highly unlikely to be different in this respect.

I've seen people make plenty of wrong assumptions on this forum. All I'm asking for is a link/ source for confirmation when there's ambiguity involved.
 
The MSAA sample grid on NVIDIA's G7x series are not sparse sampled, the rotated element of 4x MSAA appears to be fixed at that rotation - all you need to do is take a look at any of the > 4x sample patterns (8xS, 16xS, SLI8X, SLI16X, etc.) and you'll see they are all variations on the same basic MSAA pattern.
 
Dave Baumann said:
The MSAA sample grid on NVIDIA's G7x series are not sparse sampled, the rotated element of 4x MSAA appears to be fixed at that rotation - all you need to do is take a look at any of the > 4x sample patterns (8xS, 16xS, SLI8X, SLI16X, etc.) and you'll see they are all variations on the same basic MSAA pattern.

Okay thanks, but does this form of RG pattern also apply to SSAA or do you just get OGAA with it?
 
  • Like
Reactions: Geo
If you're talking about supersampling in conjunction with HDR, then that is entirely up to the developer. There is no API interface for hardware AA with HDR other than multisampling. So the developer can make the decision on whether or not to perform rotated-grid sampling, but doing so may come at some performance cost (depending on where the game is limited).
 
More good news for Nvidia (which I hate posting) there's links to a overclocked 3dmark06 score of 6700.

They overclocked it to 705mhz. If that's easy to do..wow.
 
Back
Top