NVIDIA G80: Image Quality Analysis

Discussion in 'Beyond3D Articles' started by Geo, Dec 12, 2006.

  1. Geo

    Geo Mostly Harmless
    Legend

    Joined:
    Apr 22, 2002
    Messages:
    9,116
    Likes Received:
    213
    Location:
    Uffda-land
    <a href="http://www.beyond3d.com/content/reviews/3"><img border="1" src="http://www.beyond3d.com/reviews/nvidia/g80-iq/focus.png" align="right" width="75" height="75"></a>With every new hardware generation comes an expectation of better image quality and more performance, and with G80 NVIDIA deliver that in pretty big spades. Outwardly the hardware is capable of all the big feature checkboxes, such as "HDR+AA" and very high quality filtering. Digging deeper you find true 8x multisample AA, coverage sample AA, filtering of data in non-linear colour spaces and more.

    Our NVIDIA G80 Architecture and GPU Analysis hinted at what's possible from a filtering and antialiasing point of view, and it's <a href="http://www.beyond3d.com/content/reviews/3">here that we'll flesh it out a bit further. </a>
     
  2. Cypher

    Newcomer

    Joined:
    Jun 28, 2005
    Messages:
    85
    Likes Received:
    1
    Is it me or does the image linked in this body of text:
    not have transparency AA on? All of the trees have aliased leaves...

    (specifically, this image: http://www.beyond3d.com/reviews/nvi...n/Oblivion-2560x1600-16xQAA-TSAA-16xAF-HQ.png )
     
    #2 Cypher, Dec 12, 2006
    Last edited by a moderator: Dec 12, 2006
  3. Sxotty

    Veteran

    Joined:
    Dec 11, 2002
    Messages:
    4,875
    Likes Received:
    332
    Location:
    PA USA
    Its you :)

    Actually the top left looks a bit like that.
     
  4. Skrying

    Skrying S K R Y I N G
    Veteran

    Joined:
    Jul 8, 2005
    Messages:
    4,815
    Likes Received:
    61
    Very interesting GFFX filtering result. :lol:
     
  5. 3dcgi

    Veteran Subscriber

    Joined:
    Feb 7, 2002
    Messages:
    2,435
    Likes Received:
    263
    What's up with this passage in the conclusion?
    I doubt 16x CSAA costs about as much as 16x MSAA as I suspect it doesn't require more than 4 depth tests, 8x for 16xQ. If I'm correct then the die area is likely not large either.
     
  6. Rolf N

    Rolf N Recurring Membmare
    Veteran

    Joined:
    Aug 18, 2003
    Messages:
    2,494
    Likes Received:
    55
    Location:
    yes
    Eggcellent!

    One thing that strikes me as odd is on page 2:
    "Coverage sampling fallback happens on a per-pixel basis depending on the coverage testing, which has the potential to introduce an image quality reduction in terms of 'popping' pixels, where the final colour changes on a high frequency basis because of the coverage mask and test fallback scenarios. However, the cases where that's likely to occur are most likely pathological."

    I don't think there's anything wrong at all with calling it a high frequency thing, it really is aliasing proper, if it happens. But wouldn't the amplitude of the error be limited to the error margin of coverage sampling? Defining an error margin for anti-aliasing is a bit fickle and I won't really try to be precise now. But all CSAA modes have at least four multisamples, so the difference to the "true" pixel color should be less than, uh, 33% for normal edges (alpha test etc will kill us here but stay with me).

    CSAA brings pixel color closer to the "true" pixel color.

    The amplitude of the artifact you're describing there is the difference between "less than, uh, 33%" and "closer". And that's a bad case (not the worst ...).

    So maybe it's not the frequency of the artifact that is worrisome or needs pathological circumstances. It looks more like the artifact should be pretty frequent but of small amplitude, and that you need pathological circumstances to push it over the threshold to become visible. I think the frequency is rather constantly high in all this.

    Disclaimer: This isn't a correction. I don't know jack what I'm talking about anyway. I'm just discussing an aspect of CSAA, which I think might mark an interesting tradeoff that was taken with the technique.
     
  7. rendezvous

    Regular

    Joined:
    Jun 12, 2002
    Messages:
    347
    Likes Received:
    12
    Location:
    Lund, Sweden
    I would like to see more screenshots like in the AF tests whith high frequency textures for more cards.
    While it won't be a replacement for animations to test shimmering, the Moiré patterns should give som indication of how well the cards handle the filtering of "difficult" textures.
     
  8. Rys

    Rys PowerVR
    Moderator Veteran Alpha

    Joined:
    Oct 9, 2003
    Messages:
    4,156
    Likes Received:
    1,433
    Location:
    Beyond3D HQ
    I think you're actually very right (IMO). When I thought about the conditions per-pixel, I was thinking about the fallback being no AA at all, rather than what actually happens :oops: I've been thinking about a test that renders minified thin geometry against a high contrast background, and rotates and jitters those triangles very subtly to see what happens on edges. Maybe this weekend.

    Will speculatively correct the piece regardless.
     
  9. Rys

    Rys PowerVR
    Moderator Veteran Alpha

    Joined:
    Oct 9, 2003
    Messages:
    4,156
    Likes Received:
    1,433
    Location:
    Beyond3D HQ
  10. Arun

    Arun Unknown.
    Moderator Legend Veteran

    Joined:
    Aug 28, 2002
    Messages:
    5,023
    Likes Received:
    299
    Location:
    UK
    16x CSAA takes 12 "ROP cycles", which is fricking nice, but 8x CSAA takes 8 (that's OK, I guess) and 16xQ CSAA takes 20 cycles. Personally, I think 16x is awesome, but 8x and 16xQ really feel like weak gimmicks at best.

    The sometimes absurd performance drops you see between 8xQ and 16xQ are primarily because of that; the former takes 8 ROP cycles, the latter takes 20. (or, if you want to be technically correct, you should divide that by four, given they're still quad-ROPs...)

    I think Rys really wasn't nice to that poor CSAA algorithm, given that I believe it's really an awesome thing. Heck, it's certainly one of the biggest actual innovations I've seen this year in the GPU industry. It's not because I don't 16xQ from a performance perspective that the algorithm isn'tnice at 16x, and I think there's a lot of potential for something like 32x CSAA, if the number of ROP cycles taken isn't abusive. (which I think it wouldn't be, thinking a bit about how the algorithm works, but I'm not entirely sure tbh...)


    Uttar
     
  11. ChrisRay

    ChrisRay <span style="color: rgb(124, 197, 0)">R.I.P. 1983-
    Veteran

    Joined:
    Nov 25, 2002
    Messages:
    2,234
    Likes Received:
    26
    16xQ though IMO is really there too fix worse case scenerios of 16x. Not replace it as a "High Quality" 16x. Supporting it as a means to end to give those who care about the weaknesses of 16x CSAA an alternative,
     
    #11 ChrisRay, Dec 12, 2006
    Last edited by a moderator: Dec 12, 2006
  12. Jawed

    Legend

    Joined:
    Oct 2, 2004
    Messages:
    10,873
    Likes Received:
    767
    Location:
    London
    It seems to me that G80's MSAA sample positions look like sparse samples, not rotated grid samples.

    Jawed
     
  13. Arun

    Arun Unknown.
    Moderator Legend Veteran

    Joined:
    Aug 28, 2002
    Messages:
    5,023
    Likes Received:
    299
    Location:
    UK
    4x is fundamentally rotated, other modes are definitely sparse. I guess I just suck at proofreading and didn't notice that :) Rys?


    Uttar
     
  14. Tokelil

    Regular

    Joined:
    Mar 27, 2002
    Messages:
    329
    Likes Received:
    0
    Location:
    Denmark
    Can someone explain what orthogonal filtering and orthogonal ROPs means? (I know what orthogonal means in mathematics, but Im not sure I understand what it means in this case)
     
  15. Vadi

    Regular

    Joined:
    Apr 15, 2004
    Messages:
    446
    Likes Received:
    2
    Location:
    Austria
    That's a cool review. :)

    Why is there no zoomed 2xAA shot of Oblivion? Also the colours of the sampling patterns should be more distinguishable from each other.
     
  16. Xmas

    Xmas Porous
    Veteran Subscriber

    Joined:
    Feb 6, 2002
    Messages:
    3,298
    Likes Received:
    137
    Location:
    On the path to wisdom
    It means there is no cross-dependency between certain features. All filter modes can be used with all texture formats, and all AA modes can be used with all framebuffer formats.
     
  17. Tokelil

    Regular

    Joined:
    Mar 27, 2002
    Messages:
    329
    Likes Received:
    0
    Location:
    Denmark
    Thanks for explaining! That made it much more clear.
     
  18. Cypher

    Newcomer

    Joined:
    Jun 28, 2005
    Messages:
    85
    Likes Received:
    1
    Btw, aside from the lack of TRAA in that one hi-fi screenshot, very great article. Took a bit longer than I liked, but it's certainly a lot more comprehensive than most of the more relatively amateurish reviews I've checked out. It looks like your AA demo might have to make the right edge of the white triangle be more vertical though. With the 16x AA modes, the gradient along the edge isn't long enough to have 16 seperate colours! When NV busts out the SLi-enhanced 32xQAA you won't be able to find the benefits given!

    Also, have you thought about having your AA demo include alpha-tested/texkill'd elements as well, to find differences that occur when enabling translucency AA? Like, have three ways of rendering the triangles:

    1) Solid fill, as it is right now.
    2) Filling it using fixed-function alpha test (e.g. apply a texture such that at the triangle edges it's solid, but internally it's translucent)
    3) If there's any difference (I doubt there would be; it's fairly likely that the microcode the driver generates using the FF alpha includes the microcode version of the texkill instruction, but you never know!) use a shader that includes the texkill instruction to cull pixels based on some arbitrary criteria.

    I ask because it could be interesting to see if/how TRAA uses things like the CSAA samples or something, y'know?

    Btw, I know that the performance article is yet to come, but I'm curious... in that ultra-high fidelity screenshot of Oblivion you took (16xAF, 16xQ CS/TRSSAA, HDR, max settings), about what speed did it run at? Was the game playable at all?
     
    #18 Cypher, Dec 12, 2006
    Last edited by a moderator: Dec 12, 2006
  19. dsdsdk

    Newcomer

    Joined:
    Dec 13, 2006
    Messages:
    1
    Likes Received:
    0
    i'm wondering about the conclution..

    "and NVIDIA's insistance that filtering optimisations still need to be applied as a default on hardware like GeForces 8800 GTS and 8800 GTX."

    in taking you know that ati have optimizations too? so this cant be the reason for you to mention it..
    Then you must really not like optimization..
    [Rant]
    first off, why - they are there for your good, not nvidia's..?
    secondly, do you know that they are nessesary? When do you define optimization, is Z-buffer/removal an optimization? yes it is.. would you like 10-20% performance drop just so the coumputer draws pixel you cant see due to others overlapping? great...
    [/rant]
    The optimization dont affect quality.. try making hl2 savegame, load, take ss. Then over again with optimization off.. I havent done it my self but i'm pretty confident the diff. will be 0.
     
  20. Arun

    Arun Unknown.
    Moderator Legend Veteran

    Joined:
    Aug 28, 2002
    Messages:
    5,023
    Likes Received:
    299
    Location:
    UK
    dsdsdk: The optimization Rys is talking about is brilinear. While the settings for it by default are not overly aggressive, it still causes a slight quality degradation, especially so in motion if you are very sensible to these things. Myself, I wouldn't notice it most of the time, but I'm pretty sure I can and will notice it with some specific texture types.

    Cypher: That AA demo isn't ours actually, we're using Demirug's from: http://www.nonatainment.de/web/Tools/AATester/tabid/60/Default.aspx - maybe we'll make our own a day or another to show other corner cases simultaneously, but that wasn't on today's agenda, heh :)


    Uttar
     
Loading...

Share This Page

  • About Us

    Beyond3D has been around for over a decade and prides itself on being the best place on the web for in-depth, technically-driven discussion and analysis of 3D graphics hardware. If you love pixels and transistors, you've come to the right place!

    Beyond3D is proudly published by GPU Tools Ltd.
Loading...