I wrote a blog addressing a forum post by caveman-jim a while ago… It seems that he recently edited the post to include the following:
edit – if you’re reading this from a link on ‘scalis blog‘ then I think he’s the one that doesn’t understand how modern GPU’s work on quads of pixels in groups of four (quad quad processing) and tessellation below a certain factor (making triangles smaller than inside a single quad) stalls the pipeline as you have to rerun the quad over and over again, for no discernible increase in visual quality. Crysis 2 is the perfect example of this, despite having 3-pixel triangles, tessellation is used to make a smooth sided shape… smoother, with more complex internal geometry.
Now it’s quite obvious once again what is wrong with this… At first I wanted to just respond to this on the forum. However, my post appears to have gotten deleted within minutes of me writing it (edit: hours later, the post magically reappeared again). So I will just have to post the response on my blog instead then.
Caveman-jim still does not get it. Yes, obviously I know about quads (these people have no idea, do they? Dunning-Kruger much?). But having triangles smaller than a quad does not really change the story. As I said, triangles have to be more than 2 pixels wide in order to see space between wireframes. Which implies they have to be larger than a quad at all times. It seems Caveman-jim also forgot to factor in MSAA. Namely, the quads work at the multisampled resolution, and games are generally tested/benchmarked/played with 4xAA or more, on the more high-end cards. However, the tessellation factors are determined on the triangle size on screen, not on the multisampled resolution.
I can also refer to an earlier blog once again, which covered Beyond3D’s analysis of triangle throughput vs triangle area, showing that Radeons are actually more efficient at handling very small triangles. So if anyone would have an issue with sub-quad triangles, it would be nVidia, not AMD. All the more reason to conclude that triangle size is not the issue here.
Lastly, I can refer once again to my blog on Crysis 2, you can argue that Crysis 2’s artwork does not make the most efficient use possible of the geometry detail. However, arguing that it it uses too much geometry just seems wrong. There clearly is hardware on the market that is capable of handling these levels of geometry. So that is exactly what we want. We want our graphics to be as detailed and realistic as possible, and geometry detail is a very important factor in this. It still surprises me that people are so scared of polycount.
Update: Caveman-jim replied on his forum
Attacking me in your blog is a very bad way to get yourself to be relevant. As such, I’m not interested in your opinion, which seems to self-supporting circular logic (‘I believe X, so I’ll find Y to support it, thus supporting my premise that you’re wrong’) rather than based on discussion with industry figures (i.e. GPU designers, game developers) and analysis of hardware performance results with respect to perceived differences in quality.
+1 to the ignore list.
Well, that’s funny, isn’t it? Boasting that you discuss with GPU designers and game developers… to try and discredit an actual developer? Someone here doesn’t know his position in the foodchain… Not to mention that ‘opinion’ doesn’t quite do justice to the coverage I’ve given on tessellation with various articles on this blog.