G70, G71, R520, R580 die size discrepancy
- Thread starter atomt
- Start date
There's no API interface for supersampling AA with HDR.Jaws said:
When you're talking about selective supersampling, which is accessed by requesting MSAA while transparency AA is enabled, then the supersampling that is performed is the same sample pattern as the MSAA, i.e., rotated grid.
But when you're attempting to do FSAA on an FP16 render target on nVidia hardware, your only recourse is to draw to a larger buffer than the final output (or multiple separate buffers), and recombine these manually-created samples when inputting this FP16 texture (or textures) for the tonemapping pass. So this will be ordered grid in the simplest case (simple downsampling of a larger buffer), or rotated/sparse grid if the developer so desires (when rendering to multiple buffers and blending between them, likely at a performance cost, as geometry would have to be resubmitted, and texture caches would be less effective).
Mintmaster
Veteran
Chalnoth said it twice already, but I'll try to simplify it for you a bit. When writing a 3D game with AA, developers can either specify the multisample amount, or render to a higher resolution and shrink it themselves.Jaws said:
NVidia has never shown the ability to use anything but a regular grid for supersampling via their control panel, whether we are looking at their 4xS, 8xS, or 16xS AA modes. If that's not confirmation for you, then I don't know what is. 2x SLI AA may be the best thing you can do, assuming they can figure out how to support it with HDR rendering and all the off-screen buffers.
Any pattern is possible with supersampling if you're willing to render the scene n times with jittered geometry. That is possible on any card with render to texture. But "real hardware-supported RGSSAA" wouldn't need multiple geometry passes or any effort from the developers.Jaws said:
Jawed said:
FGCG was initially introduced for low(er) power applications. Given the initial problems with getting clock skew right and the corresponding reduction in maximal clock speed, it's a possibility that they postponed using this technique for the high-end products.
Jawed said:
Yes, that seems very likely. But note that laptop/mobile components don't have to run at the breakneck speeds of desktop parts, so you have more margin to introduce new techniques.
Jawed said:
All I'm saying that this may explain the missing 25 mm2 that everybody loves to speculate about.
BTW, even if they only now start using FGCG, it doesn't mean they'll abandon the coarse grained gating: that's still an extra bonus for when a block is really completely idle.
Jawed said:
I didn't really say that they're lagging: I don't even know if Intel and AMD are using fine grained clock gating for their maximum speed processors. (Although I can't imagine that they're not using it for their mobile chips.)
It's a totally different ballgame anyway: Intel processors still use a lot of custom logic, where they can use every dirty trick in the book. As far as I know, ATI en Nvidia use standard cell technologies, where the amount of games to play is much more limited.
kemosabe said:
This one is larger than the one on dailytech. The dailytech one also have 2 rows of unused
solder pad all round the package. The die is larger and the package (substrate) is larger as well. This is identical to the G71 in the nvnews.net down to the chipcaps. Major difference is laser marking G71-U-N-A2 vs G71-GT-N-A2
One is the Ultra (GTX) and the other GT.
serenity said:
Thanks for providing me a pictures from a photographer who can't shoot straight.
The empty video decoder Philips SAA7115 to the left of G71 is a LQFP100. Body size
14mm x 14mm. Tilt makes comparison difficult.
http://www.semiconductors.philips.com/package/SOT407-1.html
chavvdarrr
Veteran
well, wild guess, but G71 seems to be like 15x15 or 14.5x14.5 , ie slightly bigger
Note, the larger die images are using the same rules for the later scans (which includes all the R5xx and new 90nm G7x series), so they should be comparable.
overclocked
Veteran
Did you get 287 right or is it 278million transistors, wonder if you accidently swapped that numbers?
Edit, when you got the time can you uppdate the NV30 era cards with diesize also..
Edit, when you got the time can you uppdate the NV30 era cards with diesize also..
Last edited by a moderator:
geo said:
I have not seen any accurate measurements on G71. It still look bigger than 196mm^2.
Anyway, the one that really mattered to me was G73 where it would have been really obvious whether it was 150 or 125. I got screwed by another fake die at dailytech. The so-called G73 that have been floating around vr-zone and others is another fake. G73 is now smaller than GDDR3 memory package.
So I was wrong. It is OK to measure my wang with Nvidia ruler because it would not measure
smaller than actual. Does ATI have the other kind of ruler?