Question about double-Z output

[AlNom]

Why has ATi shied away from implementing this in their GPUs (except for RV530)? Is the transistor cost that significant?

 

[Graham]

AFAIK xenos does this, as do the x1x00s when antialiasing is used.

 

[TurnDragoZeroV2G]

RV530 and Xenos both do double-Z, but the other ATI parts don't (not in this context--they all double z samples with AA, but not without it, which is what Alstrong is talking about here). When AA is enabled, Z output is doubled again for RV530 like other parts (which puts it at 16 z samples with 4 ROPs - 4*2*2AA) and quadrupled for Xenos (64 z samples with 8 ROPs - 8*2*4AA) per clock.

Bandwidth might very well be an issue for this. Xenos obviously has exactly enough bandwidth for all those reads and writes, and the X1600 only has 4 ROPs sitting on however much bandwidth it has (excuse my ignorance =P)

Of course, a totally uneducated guess, so feel free to ignore it.

//I think

 

[Dave Baumann]

Something to bear in mind that double Z, outside of MSAA operations, isn't a pureplay ROP operation - the necessary data has to get there in order for it to work.

 

[KimB]

But since the z-buffer is typically very compressable, I don't understand why bandwidth should be too much of a concern (increasing the resolution of a z-buffer won't hugely impact the required bandwidth to store or read data).

 

[Xmas]

Bandwidth is not a concern. If you can read and write Z and write color, you can also do double Z. But you also need to double the number of quads coming from the rasterizer.

 

[AlNom]

Thanks for the replies guys.

So basically, it just comes down to a transistor trade-off that ATi just didn't want to make. It just seems like they "could" get a pretty big performance boost in all those games that make use of double Z with and without MSAA. As TDZ mentioned, doubled output again with MSAA for Xenos and RV530. Wouldn't we see significant performance gains

I guess they're just considering all the games that are even making use of double Z too... But I'd have thought they'd want something *crazy* for the high end to definitively put it at the best of the best. (SOrry, my own mindless speculation )

Are the patents available for both companies' implementations? (or a hint on where I should be searching in the first place would be nice. )

 

[sireric]

Thanks for the replies guys.

So basically, it just comes down to a transistor trade-off that ATi just didn't want to make. It just seems like they "could" get a pretty big performance boost in all those games that make use of double Z with and without MSAA. As TDZ mentioned, doubled output again with MSAA for Xenos and RV530. Wouldn't we see significant performance gains

I guess they're just considering all the games that are even making use of double Z too... But I'd have thought they'd want something *crazy* for the high end to definitively put it at the best of the best. (SOrry, my own mindless speculation )

Are the patents available for both companies' implementations? (or a hint on where I should be searching in the first place would be nice. )

There's a significant difference in doing Z compares for AA and for non-AA. In the AA case, you have at least 2 z checks required per cycle, per pixel anyway. So all you need to do is double up the amount of "Z Rops" and then you can double your Z rate. However, for non-AA cases, you effectively need 2x more pixels around to do 2x the Z rate. That requires wider datapaths on top of more "Z Rops", as well as more scan generation, etc... Also, there are some good compression algorithms for AA Z which don't work as well for non-AA Z. Consequently, you effectively need more memory BW for non-AA Z if you increase that rate.

So, it's generaly much more costly to accelerate non-AA Z versus AA Z. Yes, you'd get performance gains with more non-AA Z. But it would cost a lot of HW.

 

[AlNom]

enlightening, thank you