OpenGL still needs to at the very least run the vertexes through the modelview and projection matrixes before it can determine whether or not they're outside the view frustum. I don't know the internal details fully, but I would expect that there are then 3 cases: fast accept (triangle is fully inside the view frustum), fast reject (triangle is fully outside the view frustum) or a third case where the triangle intersects the view frustum and it needs to go down to per-fragment testing. Older drivers might not even be able to fast accept or fast reject entire triangles. There is also vertex submission overhead on top of this; if your scene has 30,000 triangles in it, every one of them must be submitted, processed and tested by OpenGL before it can decide to accept or reject.

Running your own frustum culling in software in advance of this has traditionally been the solution. The idea is that you check on groups of triangles rather than individual ones. So if you have 300 triangles in a model, you run your own test on the entire model and just don't add it to your draw list if it fails.

There are plenty of articles on frustum culling on the web, and "bounding box" or "bounding sphere" are also good search keywords. Here's a reasonable place to start: http://www.lighthouse3d.com/opengl/viewfrustum/

Frustum culling an entire object is a good idea. I compute each object's bounding sphere offline. You might also need to optimize your glDrawElements call by optimizing for the pre and post transform cache. There is a discussion about this in the OpenGL.org forum, in the advanced section.

Warning : it is 4 pages long!
http://www.opengl.org/discussion_boards/ubbthreads.php?ubb=showflat&Number=257252#Post257252

Hi,

Consider that in most cases when people talk about frustrum culling they're talking about culling on a per-object basis. What you want is to cull triangles
from being sent to the rendering pipeline that are part of the same object.

Now, there are methods to do that but I guess not every engine needs those. The
easiest is to subdivide the triangles to groups, partitioning them.

In many cases its done with a bounding volume hierarchy but that's not the only option available to you. Even a small grid or AABBs or spheres may do the job
nicely if you can get away with something simple.

Just using a lower detailed mesh and using it beyond a certain range is probably the simplest way to start. Memory is cheap and plentiful, and is a resource to be used, so go ahead and use it.