Cull using which technique? BSP yes, other trees yes, but take a look at something like ROAM. If you go for the all out culling version you will be getting structures where you are rendering out few polygons at a time. Forcing it in this cas with today''s hardware is a lot faster.
Then take a look at some types of LOD. Reducing a mesh from 5k to 2k polygons using a real time face remover then subsequently reorganising all the various buffers will cost you more time than sticking the whole model through the pipeline.

quote:Original post by Captian Goatse
With Geforce 4 hardware it is more efficient to cull than render 2000 triangles. I don''t know what type of hardware you are running with, but triangle reduction and culling is extremely important.

Depends on the size of the triangles. With 2000 very large triangles, I could see that being an issue because of fillrate limitations. But at that point you might as well tesselate those triangles because modern hardware (CPUs and GPUs, not just cutting edge stuff, GF3+) can transform a lot of vertices very quickly. With 2000 small triangles culling would be unnecessary.

You won''t see a huge advantage using LOD on 2000 vertices unless you have some kind of crazy vertex shader that is extremely slow. I mean, Nvidia, ATI and Microsoft recommend that you render with batches of around 1000-2000 polygons (with less than 500 batches total), so what''s the point in reducing polygon count significantly lower than the recommended batch size?

Of course as a whole LOD and culling are important, but with 2000 polygons in the scene LOD is really not necessary, especially when the average game is drawing 25,000-200,000 polygons per frame.

I wouldn''t worry about elaborate LOD schemes. The things to avoid are overdraw and too frequent state changes (switching textures). Sort by texture, sort objects from front to back, render as many polygons at once as you can. Even old and crummy video cards can draw 10k+ poly scenes at playable framerates when set up properly.