Why does FSAA drop framerate?

I'm no expert, but it would seem that AA has to be applied after everything is drawn and so it takes extra time.

Well if you're fillrate limited it would because it takes more fillrate (not as much as supersampling, but still some). It also takes more memory bandwidth if that's your limiting factor.

I noticed that FSAA2,4,6 don't differ that much in framerate on my system. I have a 9800Pro.

Quote:

Original post by redcliffe Why does FSAA drop framerate? Shouldn't it be possible for video chipsets to do that seperately to other functions in a way that doesn't slow down the rest of the render? Do any chips do that? Thanks,

huh, no. FSAA is not something that can be done separately it is an integral part of the rendering.

Reasons it might drop framerate (depending on where you bottleneck is of course) :
- multisampling, the most common implementation nowadays has to allocate much larger buffers to be able to apply AA. 4xAA means a four time larger buffer. Larger buffer means more memory consumption and more memory consumption means less space for texture or other objects in video memory. If they have to spill out to host memory (agp or pcie) then rendering will be much slower.
- the activation of multisampling can disable some optimizations, or make them less efficient. For example, double fill rate for Z-only rendering is disabled on Geforces when multisampling is on.
- Resolve of multisampling is something that can be done in the DAC or by doing an extra operation just before present. In the first case, the DAC will consume extra bandwidth, in the second case that's another operation that will take some CPU time and gpu time. Usually first is faster but can be made worse for memory consumption reasons.
- Since the buffers are much larger, much more information has to transit on the memory interface for rendering: of course this is mitigated with compression, that can actively maintain bandwidth on 90% of the frame. Also some hardware implementations have an additional parallel cost for outputting more than one sample per pixel, but this cost can be hidden if every other calculation takes more than the time of writing those extra samples.

Your mileage may vary. High end hardware of today are good at reducing the impact of multisampling and allow for other high quality forms (transparency AA or flexible supersampling).

nb : compression nowadays only save bandwidth it doesn't reduce memory footprint (talking about non destructive compression).