What is Nvidias problem?

[Gamingphreek]

Everyone knows how strenuous it is to run AA and AF on a graphics card. However ATI seems to take much less of a performance hit than Nvidia does when these are enabled. What is the main problem(s) that are holding Nvidia back from the title. And what is ATI doing that is making it so efficient?

-Kevin

 

[Dman877]

Actually, nvidia cards seem to do AA with less of a hit, while ATi does AF with less of a hit IIRC.

 

[Gamingphreek]

-_- well when enabling both 1 or the other Nvidia seems to take a big performance hit. Not sure whichone it is but i know for sure that when both are on they take a much larger performance hit that ATI, why is this?

-Kevin

 

[SickBeast]

Well, many of the NV30 cards only had a 128-bit memory bus which really held them back with AA/AF. Aside from that, the "brilinear" really helps ATi with AF, along with their "performance" mode. That said, the 9700/9800 cards don't have "brilinear" and they still have superior AA/AF, so who knows. They both use different methods for AA/AF. ATi now even has the TAA feature which "doubles" AA performance for free provided that a minimum of 60FPS is sustained.

 

[Gamingphreek]

I know about temporal AA, Nvidia should really get that, but i thought Nvidia just updated the way that they use AA. Dont they use rotated grid now or is that what ATI uses. I mean every benchmark you see Nvidia hardware begins to fail when you turn on AA and or AF.

-Kevin

 

[SickBeast]

Maybe ATi has perfected their AA performance in drivers. They have been doing rotated grid since the 9700PRO. nVidia probably still has to program their drivers to the metal. Both the 6800U and the X800XT use the same memory and both run a 256-bit bus, so it really does come down to drivers and/or implementation.

 

[reever]

Both companies don't have the same bandwidth saving algorithms, so don't expect either one of them to get similar hits with different effects, ever.

 

[Gamingphreek]

But doesn't Nvidia have much more bandwidth on the 6800UE than the X800XTPE? So once Nvidia new architecture matures along with the drivers for this architecture will we see less of a performance hit.

-Kevin

 

[VIAN]

probably.

though, part is hardware and part is software.

 

[coldpower27]

Concerning, AA Nvidia's solution is faster as long as their doing the same type and amount of AA.

It's ATI's AF which gives them a huge advantage, the difference in performance for ATI between 2xAF and 16XAF is quite little while the same difference for nVidia is quite large.

 

[Gamingphreek]

Ok then lets change this question. What is wrong with Nvidias AF. What is costing them so much.

-Kevin

 

[CaiNaM]

Originally posted by: Gamingphreek
Ok then lets change this question. What is wrong with Nvidias AF. What is costing them so much.

-Kevin

because filtering takes one of nvidia's alu's away from doing other processing... ati's af implementation doesn't rob processing cycles from something else.

 

[Gamingphreek]

Ok if we can figure all this crap out why cant Nvidia get this right. Why dont in the fall refresh they fix this little performance hampering feature. Are they stupid or something???

-Kevin

 

[CaiNaM]

i certainly doubt it's related to intellegence (or lack thereof) in any way.. it's just a compromise which had to be made in the design decision. with all the extra features and etc, i doubt they had the luxury to add even more resistors, etc.

 

[Gamingphreek]

Well then why dont they move to a Low-K Dielectric and stick fix the AF. Then after all that i would think that they will have the lead back. Or is it a lack of physical space?

-Kevin

 

[CaiNaM]

Originally posted by: Gamingphreek
Well then why dont they move to a Low-K Dielectric and stick fix the AF. Then after all that i would think that they will have the lead back. Or is it a lack of physical space?

-Kevin

i would think it's more of real estate, power consumption isn't really an issue.. but you can only fit so many transisters onto a die...

 

[Gamingphreek]

And their chip is already massive. Wouldn't going to a Low-K decrease the die size a bit? Also how much more space could it take, why dont they just make it a little bit bigger.

-Kevin

 

[Matthias99]

Originally posted by: Gamingphreek
And their chip is already massive. Wouldn't going to a Low-K decrease the die size a bit? Also how much more space could it take, why dont they just make it a little bit bigger.

-Kevin

LOL. Not an electrical engineering major, are we?

Going to Low-K doesn't decrease the die size; only shrinking the process size helps there. Low-K helps with electrical characteristics of the circuit.

And I have no clue how much more space 'fixing' their AF implementation would take, but I'm guessing it would be significant (at least 5-10% bigger), and would require major changes to their drivers as well. They're already pushing it at 220M transistors; going up to 250M (for instance) would likely be a nightmare, and would tank their yields even further.

 

[jiffylube1024]

Plus it's not like Nvidia's performance hit is terrible or anything; it's bearable, just compared to ATI's solutions they have a larger performance hit.

There are articles on ATI and Nvidia's AF algorithms all over the net (including here at Anandtech). Read a few of them and they will give you some good info on this issue.

 

[Gamingphreek]

Originally posted by: Matthias99

Originally posted by: Gamingphreek
And their chip is already massive. Wouldn't going to a Low-K decrease the die size a bit? Also how much more space could it take, why dont they just make it a little bit bigger.

-Kevin

LOL. Not an electrical engineering major, are we?

Going to Low-K doesn't decrease the die size; only shrinking the process size helps there. Low-K helps with electrical characteristics of the circuit.

And I have no clue how much more space 'fixing' their AF implementation would take, but I'm guessing it would be significant (at least 5-10% bigger), and would require major changes to their drivers as well. They're already pushing it at 220M transistors; going up to 250M (for instance) would likely be a nightmare, and would tank their yields even further.

No.. Lol i dont know what i was thinking when i wrote that. But hey it would let them use less power and all that...
Anyways what is there die size now, is it 90nm or is it still at 130nm. If its at 130 why not just shrink it to 90. If its at 90 well i guess the next step down would be 65nm.
Why would it take so much space to fix the AF problem. What all would they have to do?

-Kevin

 

[Genx87]

A few things I believe are going on here.

1. ATIs brilinear appears to give them a nice boost in AF performance
2. I believe ATIs arch looks like a 32X0 arch when using AA.
3. I believe it is only AF + AA that causes the performance hit for Nvidia. Again I think this goes back to the Brilinear that ATI does.

 

[vshah]

Originally posted by: Gamingphreek

Originally posted by: Matthias99

Originally posted by: Gamingphreek
And their chip is already massive. Wouldn't going to a Low-K decrease the die size a bit? Also how much more space could it take, why dont they just make it a little bit bigger.

-Kevin

LOL. Not an electrical engineering major, are we?

Going to Low-K doesn't decrease the die size; only shrinking the process size helps there. Low-K helps with electrical characteristics of the circuit.

And I have no clue how much more space 'fixing' their AF implementation would take, but I'm guessing it would be significant (at least 5-10% bigger), and would require major changes to their drivers as well. They're already pushing it at 220M transistors; going up to 250M (for instance) would likely be a nightmare, and would tank their yields even further.

No.. Lol i dont know what i was thinking when i wrote that. But hey it would let them use less power and all that...
Anyways what is there die size now, is it 90nm or is it still at 130nm. If its at 130 why not just shrink it to 90. If its at 90 well i guess the next step down would be 65nm.
Why would it take so much space to fix the AF problem. What all would they have to do?

-Kevin

you can't just shrink die sizes at the snap of a finger, you have to wait till the process has matured enough to get sufficient yields. also, the low-k process is more expensive, and hurts profit margins on gpus.

-Vivan

 

[Gamingphreek]

But the benefits of Low-K are enough that going to it would override the downsides.

I understand about shrinking the process but it cant be that hard. 90nm is pretty mature considering processors are being released with it,albeit GPU's are much different from processors but they really do seem to need it.

-Kevin

 

[nRollo]

Originally posted by: Gamingphreek
Ok if we can figure all this crap out why cant Nvidia get this right. Why dont in the fall refresh they fix this little performance hampering feature. Are they stupid or something???

-Kevin

Kevin, what do you do for a living and what are your degrees in?

What you are asking them to do is engineer something that would make them the best in the world at something, rather than second best. Either puts their engineers a long way above "stupid" and probably you as well?

 

[Gamingphreek]

What are you talking about!
Just for the record i have no degrees (yeah im still in high school)

Anyways, im not calling them stupid, but why dont they fix this!?!? This has been a problem since AA and AF came out. Nvidia has always had trouble implementing these two as effectively as ATI. If all they have to do is shrink the die add some stuff what is possibly so hard? Granted this is out of my league as i dont know what is involved in die shrinks and stuff (getting into the highly technical forum there).
I know im asking for more or less an almost perfect Video Card but what is so hard about all the stuff listed?

-Kevin