IIRC i have used only one iteration and got equal performance only with that. With 2 iterations AND a large branching factor i assume the basic approach wins also on GPU.

But one iteration can be enough, depending on detail, kind of game etc. (one 4-branch interation == two binary tree iterations)

Downside of Iterative is that you need to keep lodcuts in memory for all nodes. Basic approach needs memory only for nodes you select for an update (important if you want a reasonable detailed large world where updating all nodes each frame would be to slow).

It remains a difficult decission, probably you'll end up implementing multiple variants... as always... ;)