[B3D Analysis] R600 has been unleashed upon an unsuspecting enthusiast community

Care to flesh out that link--er, post with what makes it worth highlighting?

More interesting to my feeble understanding is their see-sawing R600 performance in X3. The game relies heavily on stencil rate due to shadows, right? So how does that mesh with R600--with a far lower theoretical and synthetic Z-/stencil-rate than G80--doing so well without AA yet having its legs cut out from under it with AA? Does this relate to the ROPs/RBEs passing some work to the shaders? If so, does that mean the benchmark isn't stencil-, and so ROP-limited, but shader-limited?

/mental flailing

BTW, anyone else still scratching their heads about that Hexus bit pairing R600 and a 2GHz target?

 

Compared to box both tent AA's show a considerable from drop, look at hothardware review for evidence of this. Indeed 2xAA tent is similar frame rate to 4xAA box, 4xAA tent is similar to 8xAA box. Still, after all these years there is no such thing as a free lunch ..which is rather reassuring

 

I have been wondering why 3dmarks vertex and pixel shader results were so different for G80 and R600. R600 massively leads on vertex, G80 on pixel shading.

Xbitlabs suggest it is due to texture limitations

http://www.xbitlabs.com/articles/video/display/r600-architecture_19.html

and

http://www.xbitlabs.com/articles/video/display/r600-architecture_20.html

Would the tessellation transistor budget been better spent elsewhere if this is true ? Only time will tell I guess.

 

Xbit took a detailed look at FSAA quality and perfromace.

EDIT: Damnation, dizietsma beat me to it.

 

I have to admit I do not understand those xbit graphs, seems to indicate that the tent AA has more FPS even though it is taking more samples? Also they do not compare narrow to wide at 4,6 or 8x anywhere ??? It's horribly confusing....

The hothardware graph to me seems more logical


http://www.hothardware.com/articles/ATI_Radeon_HD_2900_XT__R600_Has_Arrived/?page=12


Ill have to read the techreport piece again, does not seem to be much consistency at present amongst the tables I have seen.

 

Thanks, taking a look at the links now. Hopefully xbitlabs used the 8.638 drivers.

Unfortunately, once again they didn't bother to test CFAA IQ while in motion, so their IQ analysis is a bit useless to me.

Regards,
SB

 

I think these benches were done with newer ATI drivers that's why they look a bit better. I think the next set of drivers should improve them more too.

On a side note, I really like the bench's format the way you roll the mouse over the card you're looking at and then it shows you the +/- % of all the other cards compared to it. I wish there was an English site that had this exact same setup. Is there? Anyone?

 

I tried sorting out the Xbit results at 1600x1200 (my own resolution of interest)

No AA: 173.8

2x: 122.3
2x + NT (4xCFAA) :103.9
2x + WT (6xCFAA): 96.8

4x: 97.3
4x + NT (6xCFAA): 64.2
4x + WT (8xCFAA): 62.5

8x: 52.5
8x + NT (12xCFAA): 31.1
8x + WT (16xCFAA): 28.7

 

Aye the whole shibang from Xbitlabs. Done in Half-Life 2.

1280 - 191.4

(2 samples) 2x - 160.9
(6 samples) 2x + NT - 140.1
(10 samples) 2x + WT - 132.4

(4 samples) 4x - 129.8
(8 samples) 4x + NT - 90.4
(12 samples) 4x + WT - 86.4

(8 samples) 8x - 72.9
(12 samples) 8x + NT - 43.9
(16 samples) 8x + WT - 41.4

1600 - 173.8

(2 samples) 2x - 122.3
(6 samples) 2x + NT - 103.9
(10 samples) 2x + WT - 96.8

(4 samples) 4x - 97.3
(8 samples) 4x + NT - 64.2
(12 samples) 4x + WT - 62.5

(8 samples) 8x - 52.5
(12 samples) 8x + NT - 31.1
(16 samples) 8x + WT - 28.7

1920 - 158.8

(2 samples) 2x - 103.6
(6 samples) 2x + NT - 87.2
(10 samples) 2x + WT - 82.4

(4 samples) 4x - 80.7
(8 samples) 4x + NT - 62.4
(12 samples) 4x + WT - 52.6

(8 samples) 8x - 44.4
(12 samples) 8x + NT - 24.9
(16 samples) 8x + WT - 23.4

So it appears that taking the initial 4 samples for a Narrow Tent incur a fairly hefty hit, however taking an additional 4 samples for a Wide tent incurs a very small hit.

Taking 8 samples using CFAA is a smaller hit than taking 8 samples for MSAA from withing the box. Notice the 8xMSAA numbers vs the 4xMSAA+Wide Tent numbers.

This is also verified somewhat by the 8xMSAA+Narrow Tent numbers compared to the 4xMSAA+Wide Tent numbers. The discrepency probably due more to the differences between 4xMSAA and 8xMSAA.

So it appears that pulling samples for CFAA from surrounding pixels is cheaper than pulling samples from within the originating pixel.

Some interesting things to think about in comparison to the TR piece. It'd be nice to see more data points from a greater assortment of games.

Or I may just be pondering on nothing here.

Regards,
SB

Edit - I'm wondering if Narrow Tent and Wide tent only sample from 2 surrounding pixels instead of all 4 surrounding pixes when using 2xMSAA and 4xMSAA. Anyone know the answer? And if so, wouldn't this mean that 4x+Wide Tent will be significantly less blurry than 8x+Wide Tent?

Edit 2 - No, it can't be as the drop due to NT or WT is roughly the same regardless of whether it's 2x, 4x, or 8x MSAA.

 

You mean the title originally developed with/for Xenos?

 

X-Bit Labs has a detailed comparison of IQ now.

http://www.xbitlabs.com/articles/video/display/r600-architecture_16.html

Their analysis of R600 power consumption is interesting, too. Plus they measure how loud the stock cooler is under load. ~63 dBa has got to be a record of some sort. And probably not a good kind of record.

 

The tent filter is not a form of AA nor should it be described like that. It's just a post-filter.

 

S_B, you sample numbers seem strange. If wide tent samples from two neighboring pixels, how does 2x+Wide Tent take 10 samples?

 

To further prove the point that it's just a pimped-up blur filter of sorts, look at these pics from XBit:

ATI CFAA 16x:
http://www.xbitlabs.com/misc/pictur...comparison/r600/hl2/HL2_R600_16xWCFAA.jpg&1=1

nV CSAA 16xQ:
http://www.xbitlabs.com/images/video/radeonhd2000/fsaa_comparison/g80/hl2/HL2_G80_16xQ_thmb.jpg

What's up with stuff missing in the foggy background in the ATI shots? Also, the fence on the right got eaten by the tent filter, it seems.

Bad, really bad.

 

OK, I got beaten real badly to previous post.

I can answer a question though, X-Bit Labs used driver 8.37.4.

 

Yes, I do mean the title originally developed with/for Xenos, however R600 isn't xenos, and the shader units are quite a bit different let alone rest of the chip, so I don't really think it would give R600 any advantage; PC version is part of TWIMTBP program

 

Well 63dBA assuming you play with the card in the open without a case and only 3cm from your ear.

The number's at 1m are a lot more telling. At that distance it's roughly the same as a 7950 GX2. And I'm willing to be my X1800XT is louder at load. Not to mention the current world record holder the NV30 that tested at 58 dBA at greater than 1m while in a case. I believe X-bit Labs does all their sound metering testing in an open air configuration.

It's still too loud, but not the Jet Engine that some seem to like implying.

Regards,
SB

 

Narrow Tent takes 4 samples. 1 sample from each surrounding pixel.

Wide Tent takes 8 samples. 2 samples from each surrounding pixel.

At least that is my understanding from the reviews I've been reading including the one here at B3d.

Regards,
SB

 

Xbit did say R600's showing the same problem with fog that G80 had at launch, so I wouldn't give those pics much weight. You could check the G80 launch review to see if its defects were similar, I guess, if you really wanted to use that pic as a basis for comparison.

 

Just like many other games which often ran better on ATI cards. Moot argument, or do you expect the game to be better "tuned" for nV chips than ATI chips? I don't think so until proven otherwise.