Ah, that makes sense. So, when someone talks about moving stuff to the GPU, it means to reduce the amount of code being run by the CPU, offloading it into the card.
Does CIG have to do a lot of custom stuff, or are current gen libraries good enough for most stuff? Do you think card makers will eventually help with optimization?
The choice of GPU vs CPU work is often very case-specific. GPU work is good if the work can be done in a massively-parallel way, eg image processing. Most of that work will be stuff that we write. CPU rendering costs tend to be about organising things into lists and submitting the scene’s work to the GPU. There are quite a lot of ways to reduce the render thread’s work, from straight optimisation, to moving bulky work to worker threads, to actually producing the GPU work on multiple threads, as Vulkan and D3D12 allow.