NVIDIA GF100 (Fermi) Technology preview - Accelerated Jittered Sampling | 32x Anti Aliasing

Image quality improvement

No Sir, no new Anisotropic filtering changes for once, hey ever since GT200 that's been already really good. However NVIDIA did put focus on better AA shadows and a new AA mode.

Accelerated Jittered Sampling - Improving Anti Aliased shadows
You guys must recognize this, you play a game, have 8xAA on and everything looks nice and dandy. Then you look at the shadow and notice it's all blocky and messed up. Here's where a new Sampling method kicks in, it's called Accelerated Jittered Sampling. Now the geek explanations (and really I had to look this up) is this: Jittered Sampling is a stirred process in which values are sampled equally over a rectilinear subspace. The exact position of the respective sample in each sub rectangle is thereby varied randomly.

Take a look at the first example below (lower left image). You'll notice the weird blocky shadows. This is in 3DMark Vantage with AA levels turned up. So in layman's terms, what does it do then ? Well, that banding is removed and replaced by noise. The result are much more smooth non-blocky shadows,

Let's have a peek at a scene again, where now Accelerated Jittered Sampling is applied:

Traditional	Accelerated Jittered Sampling

So there you go, to tackle this issue GF100 has accelerated Jittering Sampling which improves the quality of AA shadows, and sure will likely also bring some additional AA performance.

A new Anti Aliasing mode - 32x CSAA
Yes yes, a new AA mode is introduced. Now I'll be honest with you here. After 8xAA I gave up on all the other AA modes, 8xAA is plenty enough for me without the risk of a large performance hit. NVIDIA of course will always seek new features to present to you as an end user, and sure .. 0,008 percent of you are really into uber high AA modes. Well good news for you then. NVIDIA designed a custom sample anti-aliasing mode where you can get your freak on with 32x AA samples at very little performance loss, it's of course done with a trick or .. well two actually ;)

It's almost like cooking, you take the right ingredients stir and shake it up a little and boom ... there you go. What NVIDIA is doing with the 32x CSAA mode is relatively simple to explain, they take eight color samples and then 24 coverage samples to define your AA pixel color value. The method was actually already introduced on G80 and GT200 if memory serves me right, but has been improved. The GT200 however managed 16xCSAA with eight color samples and eight coverage samples.

Here's what's NVIDIA is doing with their 32x CSAA filter:

And for the end result (and image always look a little crappy when edited and blown up, but you'll get the idea.

So to the left the GeForce GTX 280 with 16xCSAA and to the right the new GF100 (Fermi) with 32xCSAA. Now what I need to tell you here is that the above example comes from an NVIDIA PDF presentation. We can not verify this whatsoever until we have the final product in our hands. 32xCSAA should be possible with a minimum performance hit, in fact it should be as fast as say the GT200 at 16xCSAA. Actually, if we may believe NVIDIA, 32xCSAA takes roughly the same hit as normal 8xAA, which admittedly would be really good.