R200/8500 and Multisampling AA

[swaaye]

I was rereading the old Anandtech preview of the Radeon 8500 today and saw that they had benchmarks of Porche Unleashed (I believe) running with multisampling AA on the 8500. They said the benchmarks were very hard to get to complete, but the scores showed the 8500 having a sizable lead over the GF3 in SSAA mode.

edit: Here's a link: http://www.anandtech.com/showdoc.aspx?i=1517&p=16

So, what happened to this MSAA on the 8500? Obviously something must have been terribly wrong with it for them not to develop it in the drivers and offer is as a useful feature and marketing bullet. (Though their Smoothvision labeling was deceptive. Many thought the 8500's AA was better somehow, but it was just SSAA in the end.)

 

[AlNom]

I never knew the R200 was even hinted at having MSAA.....

Good find. Hope someone here knows....

 

[ChrisRay]

Ya there were rumors the 8500 had multisampling AA on release too.

 

[swaaye]

Yeah, MSAA could have made the 8500 one amazing board for the time. However, with the 8500 only supporting SSAA in the end, and the fact that HyperZ gets crippled with SSAA, 8500's performance is decimated when you turn on AA.

Hearing about what happened would be an interesting story for sure. I tried searching the forums here a bit but didn't really see anything...

 

[Pete]

Earlier on in this review we mentioned that SMOOTHVISION (ATI's Anti Aliasing function) would reduce some of the blurring artifacts associated with other multi sampling based AA algorithms such as NVIDIA's Quincunx.

This may be Anand bungling terminology or just being fed some more bad info (again). You can tell it's not MSAA by looking at the improvement in texture quality between the two ATi zoomed-in screenshots (second and fourth: 0xAA and 2xAA, respectively). The road markings and sidewalk at 1 show marked improvement, and the shutters on the leftmost building also show some differences. That's SSAA talking, not MSAA. The performance difference is indeed surprising, considering MSAA vs. SSAA, but perhaps clock speed differences are the main culprit (8500 at 250/275 in this preview, GF3 at only 200/230)?

Weren't the rumors about stochastic sampling patterns, which Anand touches on?

 

[swaaye]

I suppose it could be clock speeds. In general, a 8500 is quicker than even a GF3 Ti500, so a regular GF3 should be getting shoved around. With AA on perhaps the GF3's memory controller is particularly inefficient or overwhelmed and perhaps we are seeing the 8500's raw advantage even though it had awful drivers at this point....

I wish Anandtech had looked into this more thoroughly and updated his preview. I bet ATI liked the false PR though.

 

[Xmas]

Well, the GF3 has no framebuffer compression which means it's extremely bandwidth limited with AA on. And it doesn't have the scanout optimization of the GF4 so if Anand really tested Quincunx (though the screenshot doesn't show Quincunx) that's another performance hit.

 

[Hyp-X]

With AA on perhaps the GF3's memory controller is particularly inefficient

2x AA effectively disables the Z-buffer compression on GF3, because it doesn't compansate for the fact that the samples are not ordered.