Comment by lonk11

Running one layer 4 times should fetch the weights of that layer once. Running 4 layers makes you fetch 4x parameters.

The recurrent approach is more efficient when memory bandwidth is the bottleneck. They talk about it in the paper.