NVIDIA Announces 16x AA For SLI

In response to ATI's CrossFire launch, NVIDIA has revisited a few points about their own solution, as well as revealed some sneak peeks into the future of their SLI technology.

Of course, one of the key points NVIDIA wanted to make was that their game support is no less than ATI's. Initially, either games needed profiles setup in order to run, or users had to know how to hack the NVIDIA XML file. NVIDIA is now offering the ability to enable user selected SLI modes for games that do not have profiles. Profiles will take precedence over user selected modes, but even games whose profiles disable SLI will allow the user to force it on.

Their other real point of contention with ATI is their claim that they add quality options where NVIDIA does not. As we know, ATI is enabling 10xAA and 14xAA options for games that don't see any real benefit from SLI otherwise. In order to top the announcement that ATI made, NVIDIA has revealed that they are planning on bringing out a 16xAA mode via SLI in a driver to be launched in early July.

We haven't gotten as much detail about this implementation as we currently have on ATI's AA modes. We don't know what the final sample point patter will look like, but NVIDIA has said that they will provide this detail when they finalize it themselves. We do know that, regardless of what NVIDIA decides, their 16x mode will be a combination of 4x multisampling and 4x supersampling. The debate currently is whether or not to implement supersample AA via an increased resolution or by rendering the scene 4 times with each rendering being slightly shifted. The advantage of the latter method is that rotated grid SSAA can be used, but the disadvantage is that the geometry load would be increased. NVIDIA has told us that they can do either method but haven't decided which to settle on.

Why is 4xSS plus 4xMS equal to 16xAA? Because each supersample point contains 4 multisample points giving us 4 times the multisample points. The other advantage is that SSAA applies to the entire scene, so we get 4xSSAA applied to parts of the scene that would see no benefit from multisampling (the interior of polygons and textures).

This mode will not be a simple combination of two scenes rendered with the current 8xAA, but rather each card will render 4xSS + 4xMS. For alternate and split frame rendering, each card will be doing full 16xAA. This may also give us a glimpse into the future as each generation graphics cards continue to increase in power. Doing full 16xAA on each card means we could see this order of AA running on a single card in a generation or so.

We are definitely interested in testing this mode when it comes along early next month.



View All Comments

  • Crucify - Monday, June 20, 2005 - link

    I am starting to wonder whether SLi was "created" due to bandwidth needs, or the need to sell two cards instead of one... Why don't they just spend money on GPU research, or memory speed..? Reply
  • patrick0 - Tuesday, June 14, 2005 - link

    Doss that mean I need a SLI setup to use 16x AA? I just ordered a 7800GTX, are they saying those 24-pipelines aren't suficient? Reply
  • Sabresiberian - Tuesday, June 14, 2005 - link

    #4 Just because you are a gaming addict doesn't mean everyone else is :) Congratulations on being in recovery!

    Also, Everquest 2 can benefit from SLI or Crossfire. No games 'require' such of course, but it's not about 'require' for enthusiasts.

  • JarredWalton - Monday, June 13, 2005 - link

    14 - If you figure that one out, Derek, write it as a plugin for Photoshop and you can probably make a decent amount of money selling it! ;) Reply
  • DerekWilson - Monday, June 13, 2005 - link

    Ahh ... if quality were the only concern :-)

    Personally, I can't wait for realtime bicubic interpolation when scaling down for SSAA ...

    How about if they let us pick our own convolution kernel for custom realtime filtering?

    If only all photoshop/premeir functions were fully accelerated on GPUs ...

    But I digress ... these guys have enough to do that would actually enhance the end user's visual experience in a significant way before the tiny little details are tackled ...

    Hmm ... I wonder if it's possible to write a shader program to do bicubic interpolation to resample a hires image to fit the display ...

    /me runs off to read his orange book
  • ViRGE - Monday, June 13, 2005 - link

    Frankly, I'm really hopeful that with the big two fighting over SSAA, that we'll see a return to SSAA being an option in the high-end market; while MSAA was a necessity due to speed, I still pine for SSAA on the days I'm not video card limited. Reply
  • DerekWilson - Sunday, June 12, 2005 - link

    #11 --

    We'll have to wait till we get the hardware to really see if it's all fluff, but I suspect that we will be able to see a good bit of difference in places multisampling doesn't help.

    You get two extra benefits from 4xMS + 4xSS AA -- the 4xSS AA applies to the entire scene -- all textures. This improves the look of textures that alias easily. Granted, this is not as much a problem as mipmapping and anisotropic filtering vertually eliminate the problem.

    But (oddly, as ATI pointed out to us) we also see a benefit of sharper anisotropic filtering with increased levels of SSAA. This is because textures at SS subtexels can have AF applied to them. Using 16x AF and 16x AA we would end up with each pixel inside a polygon the result of 4 subpixels that have had 16x AF applied to them. As ATI has told us that their implimentation essentially double the number of AF sample point, could we call NVIDIA's solution 64x AF? Honestly, I doubt that the result of 16x AF and 16x AA (in terms of AF) will result in the same image as 64 AF, but quality will certainly be increased.

    But the final measure will be in visual impact. We just want to wait to test it to make the final call.

    By the way -- If NVIDIA named their AA modes the same way that ATI is naming theirs, current 8xAA would be called 10xAA, and their 16xAA would be refered to as 20xAA ... Just to make that clear.
  • Gerbil333 - Sunday, June 12, 2005 - link

    That is retarded. AA and AF sort of work like an exponential function; once you get so far, the increased image detail is so sharp that you're not going to notice any further increases in sharpness. 4x AA looks pretty sharp. 8x hardly looks any better. 16x is going to be very hard to distinguish between against 8x, and it's hard enough seeing a difference between 8x and 4x. Therefore, this is merely marchitecture at its finest. Who cares?! Reply
  • tombman - Friday, June 10, 2005 - link

    Higher AA modes than 8x have been available for nvidia cards since a very long time by using external tools like ATUNER.I´ve been playing with 16xAA, 2x2 SSAA and even 4x4 SSAA (that´s 16 supersampling samples!!!) and they all work! For 4x4 SSAA SLI is recommended, but it looks absolutely fabulous.

    Just check out www.3dcenter.de, there you can find atuner.
  • johnsonx - Friday, June 10, 2005 - link

    In related news, ATI announces 17xAA for CrossFire, which is followed by nVidia's 18.5xAA announcement for SLI. Questioned on how they will acheive 18.5xAA, nVidia said 'We'll let you know when we figure it out". ATI quickly counters with a 21.99113xAA announcement, though they admit it won't really be available until they add their PI-based rendering engine in late 2005. NVidia ups the ante almost immediately with their Multi-Super-Quincunx 96x AA method that will require 4 dual-core GeForce 8900 Ultra cards. Questioned about frame rate, nVidia said 'Frame Rate? We don't need no stinking Frame Rate! We've got 96xAA!". ATI is left strangely silent...

Log in

Don't have an account? Sign up now