[H]OCP updates GFfx review with correct 2xAA screenshots

Dave H

Regular
[url=http://hardocp.com/article.html?art=NDIxLDQ= said:
[H]ardOCP[/url]]It seems the only two Anti-Aliasing screenshots that were wrong originally are 2X and Quincunx which did cause the original screenshots not to output correctly with the built in screenshot method of F9 in UT2K3. Working with NVIDIA it seems we have a solution that should give you an actual representation of what the screen looks like during actual gameplay with Anti-Aliasing in 2X and Quincunx enabled.

Using Hypersnap DX we have taken the screenshots again at 2XAA and Quincunx in UT2K3 and NFSHP2 and have updated the pictures below in those two settings. You can now see that there is definitely AA happening at 2X and Quincunx. Our baseline conclusion still remains the same however; in-game AA quality still goes to the 9700 Pro, however they are much closer in comparison then we originally thought.

From glancing at the pics, real 2xMSAA is definitely being done. OTOH, the R300 indeed has noticably better IQ (on my screen). Considering that both are RGMS, my guess for the quality difference is simply that GFfx's AA is not gamma corrected. But perhaps there's more to it than that...
 
you really have to see it in-game and play on both to get a feel for the quality/performance on both...

i can upload to anyone who wants em the uncompressed pics in bmp format by email if you wanna zoom in see what you need to see etc...

the original pics are from 2.25MB to 3MB, i can rar em of course if you want any...
 
Brent said:
you really have to see it in-game and play on both to get a feel for the quality/performance on both...

Brent, so please tell us which card you'd use if you were going to spend some quality gaming time?
 
Dave H said:
From glancing at the pics, real 2xMSAA is definitely being done. OTOH, the R300 indeed has noticably better IQ (on my screen). Considering that both are RGMS, my guess for the quality difference is simply that GFfx's AA is not gamma corrected. But perhaps there's more to it than that...

I agree. I was about to say that the GF FX 4X FSAA wasn't that much better then the 2X FSAA on the R9700 after my first glance at the the Need for speed images. But then i had a look at the "map" on the left. Anyway, there's definitely a huge difference in FSAA quality between the two cards.

On my monitor (ViewSonic p225f 21") combined with my untrained eyes:

The R9700 6X FSAA looks WAY much better then the 8xS on the FX.

The R9700 4X FSAA looks better overall then the 8xS on the FX.

The difference between the 8xS and 4xS isn't that big.

Now, i must also add the the FX image is much darker (brown is much more brown :)) which might explain a bit of the huge difference. The R9700 images looks a bit washed out in comparision. Maybe just gamma settings, dunno. But they should have tried to made those setting as close to the same as possible before taking the screenshots.
 
How does Hypersnap DX manage to capture the real image?

I'm just wondering if it performing the math itself to simulate the post filter? If so, on what basis is it performing the math? Is it getting all of the buffer data from the card somehow?
 
demalion said:
How does Hypersnap DX manage to capture the real image?

I'm just wondering if it performing the math itself to simulate the post filter? If so, on what basis is it performing the math? Is it getting all of the buffer data from the card somehow?

Yep, that's it. It gets the entire framebuffer from the videocard memory. For 2x it simply averages the sample pairs together (there are 2 samples per pixel), and for Quincunx it averages together 5 samples. The sample in the middle gets a weight of 0.5, while the sample up left, up right, down left and down right get a weight of 0.125 each.
 
Dave H said:
Considering that both are RGMS, my guess for the quality difference is simply that GFfx's AA is not gamma corrected. But perhaps there's more to it than that...

I thought the R300 used sparse-sampled / pseudo-random sample patterns. It certainly allows programmable sample patterns, so it seems unlikely they would stick with a grid pattern.
 
I thought the R300 used sparse-sampled / pseudo-random sample patterns. It certainly allows programmable sample patterns, so it seems unlikely they would stick with a grid pattern.

R300 hardware allows programmable sample patterns. AFAIK this feature has never been turned on in the drivers. In the current drivers, I believe both 2x and 4x MSAA use a standard rotated grid pattern. (Not sure about the 6x pattern.)
 
Fuz said:
Brent said:
you really have to see it in-game and play on both to get a feel for the quality/performance on both...

Brent, so please tell us which card you'd use if you were going to spend some quality gaming time?

I have a 9700 Pro in my primary system, and its staying right there for now ;)
 
I guess that the HUGE thing that stands out to me are still 6xs and 8xs V.s. 6x on R300. Not only does the image on the 9700 look much better (especially the rails) but the GFFX loses by 20, and 40 FPS perspectively.

I just updated to the new UT patch today (and bonus pack). Performance accross the board is up for the 9700pro in D3D. It was good before but it is fantastically smooth and stable now, even at 6x FSAA+ 16X Perf Aniso. I just played for about 3 hours. I did not see the frame counter drop below 45 even one time. It was often hanging at about 65-70ish, with most combat situations around mid 50's. I had all the settngs maxed, with detail textures on at 1024x768 with 6x FSAA+16X Perf AF. I played all the new maps also. The performance was just really sweet, not super huge, but liquidy smooth and very consistant.
 
Brent said:
hehe, i still got it, still evaluating it...

What domestic cleaning role are you evaluating it in? :LOL:

MuFu.

P.S. I hope you guys are grabbing Gainward by the balls right now regarding getting a review sample of their "7dB" GFFX. Can't wait to find out what the hell that is...
 
Brent,

Pardon me but at least to me the xS shots don't look to be correct either.

Here's Hardocp's 8xS/FX shot:

10436208595cUSd31HIx_4_6_l.jpg


And here's a 8xS/NV25 shot (presupposition that what Rivatuner picks is actually 8xS and not just 8x; I used 41.09 and Rivatuner RC12.2). It's just the demo, since I haven't reinstalled the game yet.

8xS.jpg


Question: since you're still toying around with it, dropping to 16bpp colour depth/32bit textures should get in the 60's with 8xS in that scene right?

edit: screenshots captured with ingame F9.
 
The modes of 6xs and 8x that brent captured in both games seem to be 4xOGMS with 1.5x1 and 1x2 supersampling. The ones on the gf4 are 2xRGMS with 1.5x2 and 2x2 supersampling.
 
Bambers said:
The modes of 6xs and 8x that brent captured in both games seem to be 4xOGMS with 1.5x1 and 1x2 supersampling. The ones on the gf4 are 2xRGMS with 1.5x2 and 2x2 supersampling.

Meaning he actually used 8x and not 8xS on the FX? (could it be that he had actually 8xS in the driver panel enabled and it turned out to be just 8x in the end?)
 
After relooking at several interviews and tech briefs...

nVidia *never* talked about 8xS. Some reviewers did. Not nVidia.

When nVidia was asked about the new AA modes, they always said "6xS and 8x"
The difference between GF4 8xS and the GFFX 8x?
The GF4 8xS is non-ordered and D3D only.
The GFFX 8x is ordered and works in both D3D & OpenGL

Could it be nVidia is preparing a performance *and* IQ boost package to counter the R350? Or maybe they just messed up big time...


Uttar
 
Back
Top