GDC 2003 Jason Rubin ( he mentions something interesting ;)

[pcostabel]

Just some quick calc's- at 60 fps, that's 16 mil polygons per frame! On a 640x480 screen, that's literally 50 polys per pixel. Take it up to a basic HDTV resolution and that's still nearly 14 polys per pixel- or alternately, polys potentially 1/14 the size of a single screen pixel. I think the "n-gon" look and "cube finger" affliction will be pretty well decimated at that point.

...or you could look at this way- you could conjure up a graphics engine that works at a paltry 7% poly throughput efficiency and still come up with 1 poly per pixel (comparable to a 30 mil poly/sec game from today). That's a lot of resources left out to devote to lighting, effects, and filtering!

First of all, we're talking about HDTV for next gen, and that's 1 mil pixels per frame. Secondly, this polys per pixel computation doesn't make any sense since in reality there is no one to one relation between polys and pixels: some polys can cover half a screen while others can be so tiny to only cover a small part of a pixel, and you have to take into account overdraw.
Lastly, next generation machines will use high order primitives such as NURBS patches, so all this polys per second figures have little relevance. The FLOPS performances are much more important.

 

[V3]

some polys can cover half a screen while others can be so tiny to only cover a small part of a pixel, and you have to take into account overdraw.

I think that's why you should use LOD.

 

[marconelly!]

Lastly, next generation machines will use high order primitives such as NURBS patches, so all this polys per second figures have little relevance.

Still, polygons are used even in high end modeling for offline film renders. I don't know what is the reason for that, I've heard you can't model everything with NURBS, and perhaps they render too slow so replacing them with polygon object tesselated enough is always faster.

 

[pcostabel]

some polys can cover half a screen while others can be so tiny to only cover a small part of a pixel, and you have to take into account overdraw.

I think that's why you should use LOD.

Even with LOD, you can't get rid of sub-pixel polys or overdraw.
Saying that you only need 1 poly per pixel is just like saying you only need one color per pixel: why do we need 16mil colors when a 640x480 image only has 307,200 pixels?

 

[pcostabel]

Lastly, next generation machines will use high order primitives such as NURBS patches, so all this polys per second figures have little relevance.

Still, polygons are used even in high end modeling for offline film renders. I don't know what is the reason for that, I've heard you can't model everything with NURBS, and perhaps they render too slow so replacing them with polygon object tesselated enough is always faster.

Absolutely, there are several reasons to use polys. Flat surfaces are more efficiently modeled with polys, polygonal meshes don't need stitching, cloth simulation only works with polys etc. But high order surfaces are better suited to a gaming platform because they are a compact representation that saves memory (which is a scarce resource on consoles), are faster to animate and have intrinsic LOD. Moreover, next gen will use per pixel lighting, so I expect a shift from hardware rasterizers to a line rendering approach similar to the one used by off-line renderers; in that case poly performances won't make much sense anyway.

 

[randycat99]

Saying that you only need 1 poly per pixel is just like saying you only need one color per pixel: why do we need 16mil colors when a 640x480 image only has 307,200 pixels?

Easy there- I wasn't trying to say that you only need 1 poly per pixel at all. I was simply trying to give some perspective as to how potentially small the polys could get with so many at your disposal for a given screen size. If you can literally throw more polys on the screen than you have pixels, then it seems to me that there is great potential for very finely meshed models to appear on a game screen. Rounded and curved things would actually appear as so instead of a "reduced" hexagon look. Organic shapes could actually be organically shaped instead of faintly faceted and angular, simply veiled by clever textures. I could be wrong, but that's how I picture it given my limited understanding of the matter.

 

[V3]

Even with LOD, you can't get rid of sub-pixel polys or overdraw.

You can minimise it though.

Saying that you only need 1 poly per pixel is just like saying you only need one color per pixel: why do we need 16mil colors when a 640x480 image only has 307,200 pixels?

Well, each of those 307,200 pixels need a color, that's why you need 16mill colors.