[long]I've found the GFFX AA IQ bug! ( hopefully )

[Bambers]

It may well be a capturing problem although its a little odd that it only affects 2x and QC. They do work fine in need for speed though. FXs 2x is the same as 9700s (aside from the fog) in that game and the quincux screenshot even shows the blurring too.

 

[Tahir2]

LOL.. for some reason I didn't think so.

Walking over from Canada to Santa Clara (that is where the rivals are based right?) would be a mean feat though!

Now you are gonna diss me and tell me that you are not based in Canada at all but have offices just outside of NVIDIA.. d'oh!

 

[OpenGL guy]

Walking over from Canada to Santa Clara (that is where the rivals are based right?) would be a mean feat though!

Hint #1: I work in Santa Clara.

Hint #2: ATi an office in Santa Clara.

Now you are gonna diss me and tell me that you are not based in Canada at all but have offices just outside of NVIDIA.. d'oh!

Bingo!

 

[Tahir2]

I consider myself dissed then?
Don't answer that!

 

[PSarge]

I thought it was a known fact that GF4 did the final blend for some of it's AA mode (QC in particular) in the DAC stage, thereby making it impossible to screencap it properly. Is that true, or have I been reading too much propaganda

If so, then isn't it a reasonable conclusion that they're doing the same on GFFX?. It would be ideal for 2x (and they'd continue doing it for QC). Assuming they can only blend 2 samples in the DAC stage then anything above 2 would need to be blended in the frame buffer, allowing it to be screencapped (although it might be only a partial result).

Real proof is....Do the screencaps look the same as the monitor output? Does anybody have a card to look at?

 

[Dave Baumann]

I think thats the case PSarge, however I think there is some question wether this would only be applicable to the 2X modes or all modes or not.

As to having a card - there only appears to be about 4 in existance at the moment!

 

[PSarge]

Hmmm, sounds like I might be playing catchup

Put it this way...if it was my chip, and I had a cheep way of blending two sample buffers together, I'd use it for all the AA modes I could. I'd vote for all screen caps are possibly invalid.

Could really do with a screen capture device that sits on the end of a DVI cable, but I think the MPAA (i.e. Hollywood) might not be to happy with that.

 

[nggalai]

Regarding 2xMSAA and the "hardly any anti-aliasing at all" comments by other reviewers, this is what Thilo (PCGH) had to say:

kann ich so nicht bestätigen [Wie bei Anandtech erwiesen, ist die 2xAA Qualität doch deutlich schlechter als die 2xAA Qualität der 9700'er Karten.]. es scheint zwar so, als ob bei der fx keine gammakorrektur erfolgt. aber ansonsten sind die unterschiede nicht großartig.

"I can't confirm Anand's findings that 2xAA quality with GFFX is clearly inferior to 2xAA quality with a Radeon9700. Apparently the GFFX doesn't do gamma correction, but apart from that, the differences are nothing to write home about." (paraphrased)

ich hab keine ahnung, was der kollege da gemacht hat. [Anand: "The first thing you should notice is that neither of the 2-sample AA algorithms (2X or Quincunx) seem to be doing much of anything in the way of actually anti-aliasing. "] ich für meinen teil habe alle benchmarks, qualitätsvergleiche und screenshots (fsaa-tester usw.) mit "balanced" gemacht und keine der vom kollegen beschriebenen probleme mit 2x gehabt.

"I don't have the foggiest idea what my colleague did do wrong. [i.e. "The first thing you should notice is that neither of the 2-sample AA algorithms (2X or Quincunx) seem to be doing much of anything in the way of actually anti-aliasing."]. As far as I'm concerned, I made all screenshots, benchmarks and quality comparisons with "balanced" settings and had none of the issues with 2xMSAA as described by Anand." (paraphrased).

Really seems to be the framebuffer thing.

ta,
-Sascha.rb

 

[nggalai]

I think thats the case PSarge, however I think there is some question wether this would only be applicable to the 2X modes or all modes or not.

I wonder. Hmm. Well, can a RAMDAC blend more than two multisample buffers? I am not too sure about it, but I'd say "no." Blending with 4xMSAA thus has to be done in the framebuffer, I'd say.

ta,
-Sascha.rb

 

[PSarge]

Hmm. Well, can a RAMDAC blend more than two multisample buffers? I am not too sure about it, but I'd say "no." Blending with 4xMSAA thus has to be done in the framebuffer, I'd say.

Their RAMDAC can do whatever they want it to, but assuming that they only designed it to do two...

You might blend three samples in the frame buffer and merge the forth in the DAC. Thinking about it I don't think it actually buy you anything to do that, but....

(you're only moving the 4th sample read from the core to the DAC after all. Possibly slightly less work for the core, but it depends on the architechture.)

 

[Dave Baumann]

"I can't confirm Anand's findings that 2xAA quality with GFFX is clearly inferior to 2xAA quality with a Radeon9700. Apparently the GFFX doesn't do gamma correction, but apart from that, the differences are nothing to write home about." (paraphrased)

I wonder. Hmm. Well, can a RAMDAC blend more than two multisample buffers? I am not too sure about it, but I'd say "no." Blending with 4xMSAA thus has to be done in the framebuffer, I'd say.

AFAIK, GF's do not use 'Multisample buffers', as in separate buffer for separate samples - I still think they define a single Multisample buffer scaled to the size of the AA level. But that not the issue, the issue is the number of samples the RAMDAC can combine at once (by the time its reached the RAMDAC it's not caring if its come from separate or multiple buffers).

Two opposing things here. 1.) this came from 3dfx, and their V5 RAMDAC's could only cope with two samples (buffers) per chips, so perhaps, if they just lifted the 3dfx RAMDAC, it might be limited to two samples. 2.) On the other side, if this works for Quincunx as well you'd think it can handle more than two samples at once because Quincunx is dealing with at least 5 colour values at any one time.

 

[Tahir2]

It doesnt do Gamma correction? Can I say 'poo'?
Oh ok then..I wont

 

[Arun]

nVidia insisted a lot about Gamma Correction in their PDFs
Since those drivers don't support DX9 features of the card according to rumors, I guess it doesn't support Gamma Correction either but will soon...


Uttar

 

[Dave Baumann]

Gamma correct won't be a driver / DX9 thing, its a hardware function. However, the documentation actually only specified Gamma correct in the Pixel Shader, it didn't single out Gamma Correct FSAA, so I was always a little hazy on that point.

 

[LeStoffer]

Since those drivers don't support DX9 features of the card according to rumors, I guess it doesn't support Gamma Correction either but will soon...

How will you then explain that R300 had working Gamma Correction from the start without DX9? I'm starting to think that the drivers has been rushed beyond what nVidia would normally accept.

 

[Luminescent]

Just wanted someone to confirm whether or not Nvidia's 2X MSAA is indeed rotated grid. Is it?

 

[Bambers]

Yes it is and has been since the gf3. Just look at some 2x screenshots or nvidias pdfs on AA on gf3/4

 

[Luminescent]

So the 2X mode should give reasonable edge IQ. What I don't understand is why blends would be carried out in the ramdac. Aren't samples taken and blended within the AA pipeline the the processor?

 

[Bambers]

It does seem a little odd but it could explain the result on face3 from brents review

Look how the FX has had to resort to agp texturing at 1280x1024 with 2xAA yet the 9700 apparently still has room in its memory.

If the FX is blending 2x in the ramdac and keeping a double sized buffer in memory...

 

[ERP]

So the 2X mode should give reasonable edge IQ. What I don't understand is why blends would be carried out in the ramdac. Aren't samples taken and blended within the AA pipeline the the processor?

It saves doing a copy forwards, on GF3 the copy forwards can take as much as 2+ms for Quincunx mode.

The tradeoff is that your framebuffer is as big as your backbuffer.