Games always have too much stuff to render, but are often set in occluded environments, where much of what is in the view frustum is not visible to the camera. Occlusion culling tries to avoid rendering objects which the player can’t see.
Discussion of occlusion for games has tended to focus on the use of GPU pixel counters to determine whether or not an object is visible. This places additional stress on a limited resource, and has latency which has to be worked around, requiring a more complex pipeline.
For Killzone 3 we took a different approach - we use the SPUs to render a conservative depth buffer and perform queries against it. This allows us to cull objects very early in the frame, avoiding any pipeline costs for invisible objects.
This presentation talks about the ideas (and dead ends) we explored along the way, as well as explaining in detail what we ended up with.