Nice to see some comparison shots, but I don’t think they are conclusive. Proper profiling with several averaged samples needed 
Plus, you can’t really compare cascades like that. In CE cascade updates are distributed over several frames. Somebody should be kind enough to provide a unified test scene 