If you have to choose, try to keep the GUI card in the x16, and put the CUDA cards in x8.
Juan – it’s not the total bandwidth which matters, it’s the bandwidth when Resolve needs to use the GPUs. So, running 4 GPUs in a shared x16 slot is not the same as running 4 GPUs in x4 slots each, you will lose some performance. It depends on what resolution images you are working on, and what sort of processing you are doing, so difficult to quantity. For HD images, you’ll see very little loss in performance sharing the bandwidth.