Or you know, just enable backface culling and let the hardware do that for you! Performing your own culling algorithms with that granularity isn't worth it. Throwing stuff at the GPU in big batches is....

Yup, it sounds to me as though the CPU-side culling approach here is going to have more overhead than just throwing it all at the hardware; a cube is a very lightweight geometric construct that the GPU can cull very easily and quickly, compared to traversing a scene, doing all of those dotproduct tests, building up dynamic buffers, etc. The cost of doing this for even a single face may well outweigh the cost of just submitting all 6 and letting the GPU do it.

CPU-side culling works best when it's done on a quite coarse scale than can reject 10s, 100s or 1000s of objects with a single test, but when you get down to individual faces or triangles, you're talking serious diminishing returns.

For a minecraft type game, would it be better to use hardware instancing or create faces of cubes based on what's visible? I'm not sure on how many cubes can be rendered using instancing, but i want to know if it's worth my time to try it out.

HW instancing will win by a mile, but that doesn't mean you have to draw all of the blocks! I'd be tempted to split your world up into batches of say 20x20x20 cubes (possibly even larger than that), and then only draw those groups that are visible (with a single instanced draw call per group).

Once you have a list of drawable cubes you can traverse that list and extract visible faces using the normals. Using the faces, populate another array that you copy to your instance buffer. The geometry buffer would then only be a quad.

Or you know, just enable backface culling and let the hardware do that for you! Performing your own culling algorithms with that granularity isn't worth it. Throwing stuff at the GPU in big batches is....

Yes! Yikes. Don't do this on the CPU. Let the graphics card do its job.

Casadilla, have you researched all the functionality of the default graphics pipeline? I keep seeing questions like this from you. The pipeline is designed to take this kind of work off the hands of the CPU. Any time you're looking at running some operation on all your graphical entities you should immediately be looking to offload that to the GPU. The common operations are just a matter of throwing a switch, and more complex operations are available through shaders. Don't optimize until you start to approach reasonable performance limits, and once you do reach that point, optimize at the large scale, like menyo and Rob are saying rather than at the per-instance scale. The GPU is designed to do batch geometry operations. It can likely cull the faces on thousands of cubes in parallel.