I will try and find it but also if you are still interested, but a while ago (2 months) I saw a very complete tutorial on terrain generation, and frustrum culling in the GPU (including also LOD in there as well). It was written for C#/Slimdx, but I cant find it at the moment and I have too many bookmarks :)

But it is out there, just have to find it... :)

If you find it, please share it here !!

GPU frustum culling is great, and only recently has it been reasonable to do with new api features. You could do it pre Vk/dx12 but it was kind of a headache for a lot of scenarios.

Since running the culling in either a cs or gs is going to almost certainly mix up your pipeline state a little it probably doesn't matter a ton which you choose -- I'd assume compute would be faster but vendors might have made gs filter operations really fast? Probably not though.

Not sending a ton of unused data to the GPU is another part of it. It's less of an issue than it was a few years ago, but it's still pretty relevant. I guess I could see doing a sloppy CPU cull and then doing cleanup and clipping on the GPU, but I can't imagine that it would be more effective to upload render instructions for every object in the scene and then leave the GPU to sort it all out.

Reducing latency can also be a big part of it. Imagine your command list builds a frustum from the most recent user input in gpu readable memory rather than being given a frustum by the host. Normally input is going to be at least as stale as however many frames were queued up in front of it. Instead the input will be only as stale as the time between when it was polled on the gpu and when its presented, less than a frame. But the only way this can work is if culling is also done on the gpu since it won't know what it's looking at until the command list is actually being executed.