This is true, but what the parent comment is getting at is we really just want to be able to address graphics memory the same way it's exposed in CUDA for example. Where you can just have pointers to GPU memory in structures visible to the CPU, without this song and dance with descriptor set bindings.