With modern fillrates and deferred rendering, is occlussion culling still relevant?

Question

For example, while it's the current top-of-the-line GPU, the GTX 980 has a staggering 72.1 gigapixels/second fillrate, which with back-to-front rendering and/or Z buffer checks, seems almost ridiculously large, possibly even at 4k resolutions. As far as polygon counts go, modern GPUs can do tens to hundreds of millions of textured triangles without a hitch if you batch and/or instantiate them properly.

With forward rendering, the amount of fragments that shaders will run on can quickly become overwhelming, but with deferred rendering, the cost is usually more-or-less constant, depending on the resolution, and we long since passed a point where most shading or post-processing effects can be done in realtime in 1080p.

Either way, the limiting factors nowadays are most commonly draw call counts, and shading costs, both of which are kept relatively low by proper deferred rendering and geometry batching, so with that in mind, is culling more than just backfaces and out-of-frustrum polygons of any substantial benefit? Wouldn't the costs(CPU/GPU time, programmer time) outweight the benefits, a lot of the time?

Answer 1

Yes, occlusion culling is still worth it.

At minimum, a draw call that you skipped due to culling is a draw call that doesn't have to run the vertex shader. Triangle count goes up as quickly as GPUs start supporting more triangles, because why not? With unified architectures, vertex shaders use the exact same hardware that pixel shaders do, so every vertex that you skip because of culling is more compute time for the stuff that you can see. Not to mention all the other stuff you're skipping (CPU draw call processing, and throwing the tris far enough through the pipeline that the rasterizer realizes that it doesn't need to shade them).

There was a great presentation from two Ubisoft studios at SIGGRAPH 2015 about GPU-driven rendering pipelines. On the surface it's about some of the things you mentioned: batching and instancing, reducing draw call counts. But one of the major advantages they get out of a GPU-driven pipeline is incredibly fine-grained occlusion culling: better culling than you would normally see at the draw-call level. It's all in the service of asymptotically approaching the goal: processing only what you can see, which means that what you can see looks better.

(Also: consider console, mobile, VR, and desktop-without-the-latest-and-greatest-GPU-that-money-can-buy. Even if all your tris disappear into the gaping maw of your top-of-the-line GPU, you may not be the primary target.)

Answer 2

It depends on the game style just how much culling is needed. For instance first person shooters benefit from this a lot, having a lot of stuff in the frustrum at any one time, while an overhead view RTS does not since you are effectively looking at a plane with things on that plane. Even in an RTS though, doing a "depth only render" to eliminate overdraw is still useful.