Theoretical question

sdf

Distinguished
Aug 5, 2007
231
0
18,690
Just a question, and excuse my ignorance if I may be leaving something out or missing the mark. Something I have been wondering for a while now as new cards are coming out and all the spec talk on them. Spec for spec if you take a 3870 and a 8800 GT side by side the 3870 pounds the 8800 with the exception of texture fill rate and the texture units where the 8800 has a large advantage. If you took a 3870 and only changed the texture units to the same as the 8800, ie 56, would that improve performance? If so any reason why they wouldn't want to have added more for the 38xx series? Seems strange from a competitive standpoint.
 

homerdog

Distinguished
Apr 16, 2007
1,700
0
19,780
The 320 "stream processors" on the R600(HD2900)/RV670(HD3800) are actually just 64 processors each capable of 5 operations per clock cycle under ideal conditions. If data is not fed to them properly they lose great deal of efficiency. The same can be said of Nvidia's 8 series, but to a lesser extent. Also, the SPs of the HD3800s are clocked at the same speed as the core, whereas the 8800s' SPs operate at a multiple of the core clock (>2x).

The texture unit advantage of the 8 series isn't all that important. It's the ROP count and design that gives 8 series cards the advantage when AA is applied. The 8 series has hardware in its ROPs specifically designed for AA resolve. The R600/RV670 has to perform AA calculations with its SPs, a more flexible but currently slower process.

Sorry if this seems a bit complex, but you asked a rather complex question :)
 

spoonboy

Distinguished
Oct 31, 2007
1,053
0
19,280
not quite, if the texture unit count wasn't soo important then ati wouldn't (or so its rumoured) be doubling the texture unit count for its next high end chip (rv770), going from 16 in the current rv670 with 320 shaders to 32 (tmu's) and 480 shaders in the rv770.

With some quick and rough maths the ratio of texturing units to shaders is 1/3 higher. If implemented in the 3850 & 3870 (and 2900xt and pro too) that would mean something like 21.33 tmu's instead of just 16. Of course you can't have 0.33 tmu's but this is just hypothetical after all. (note ati's texture units are a bit more beefy than nvidias, so less are required for equivalent performance, although nvidia still has a big lead here)

"The 320 "stream processors" on the R600(HD2900)/RV670(HD3800) are actually just 64 processors each capable of 5 operations per clock cycle under ideal conditions."

Thats not quite true. There are 64 more capable shaders, each with their own group of 4 less capable ones. Another unit attempts to feed them with instructions cued in the best possible (most efficient) order looking 'forward' 5 instructions. This is complicated, and was often highlighted as a possible issue when the r600 was released but driver revisions (poor early performance and some plain broken games performance) I would speculate have gone a long way to preventing chronic stalling here.

"The texture unit advantage of the 8 series isn't all that important. It's the ROP count and design that gives 8 series cards the advantage when AA is applied."

The 8 series generally has an advantage when AA (and even AF) is NOT applied. Indicating that either shaders or texture units are lacking. Given r600 and above only have 16 TMUs and an apparent horde of shaders, it would appear that more texture units are lacking. Having the shaders clocked seperatley from the main core shouldn't be an advantage for Nvidia when ati have more but slower clocked shaders. 3Dmark06 feature test results do show some very capable pure shading power from the r600 & rv670 cores. So on that point it's pretty moot.

As far as ROPs go, its not quite true that the 8 series has an advantage in having more of them, rather that they are used to implement AA, whereas for DX10 ati cards they 'just' shuttle pixels out to the monitor. So comparing ROPs here is a bit irrelevant.

However, the poor price/performance of the ati 2900 & 2600 xt & pro, and no competitor to the 8800gtx last year really helped 'the way it's meant to be played' take off. Games are developed with nvidia cards in mind, make of that what you will, but in my opinion that matters more than lacking texturing power.


 

homerdog

Distinguished
Apr 16, 2007
1,700
0
19,780
Well you went into more detail than I was going to, but that's cool.

Also, having more slower clocked shaders may actually be worse than having fewer higher clocked ones. More parallelism means more dependency on the dispatch processor to break up the threads nicely.
 
@ sdf
Long story short
While the guys are doing a great job trying to explain it in simple terms (thats never gona be easy) :)
The simple truth of it is that the two cards work so differantly and are built so differantly from each other that its pretty pointless trying to compare them side by side.
As has been said the next gen is rumoured to be upping the count so when the reviews come out we can tell what differance they do make.
Mactronix
 

rallyimprezive

Distinguished
Jul 18, 2007
470
0
18,780
Personally, I happy to see that ATI and nVidia are taking such different approaches to the ultimate goal of high FPS and good IQ. I think that one of the issues is that nVidia has been working more closely with game devs to optimize engines for nVidia's particular architecture.

Its good though because it gives the consumer more options when making a purchase.
 

FrozenGpu

Distinguished
Dec 8, 2007
986
0
18,990

like harboring terrorists, like saudia arabia? :lol: :whistle: :pfff:
 

SpinachEater

Distinguished
Oct 10, 2007
1,769
0
19,810



TWIMTBP is actually a genius program...I wasn't trying to insinuate that anything cynical was happening on NVidia's behalf... :whistle:
 


Oh i have to agree that coming at it from differant angles is good as the chances of a big break through are essentially doubled. Im not sure too many people actually take it into account when making a purchase though, and by that i mean for example you wouldnt generally hear someone say "well actually im getting the Nvidia card as i like the way it does Tansparancy AA and not Adaptive AA like the ATI card does". So looking at it like that its hard to say that the differance in archetecture is the reason for the choice being there. the X1*** and the 7 series were alot closer in that respect but the choice was still there.