GPU-driven rendering in video games still requires some data transfer from CPU - at least for transformations of game objects. In some cases it can create a bottleneck. That's why it should be considered to implement game logic on GPU and limit data transfer only by player input.

I once experimented with such approach. It generally works, but it's hard to do so - one need to use only a shading language for general game logic, debugging is practically non-existing.