Less cells to store means memory optimisation; less cells to calculate - runtime optimisation! On the contrary, in usual WFA the current layer depends ONLY on previous layers, so you can compute all cells in the layer at the same time (in parallel). Furthermore, in the diamond optimisation we can only extend one diagonal at a time (or possibly two — one at each side), while in WFA we can extend all diagonals in the same layer at the same time since they do not depend on each other and can be calculated simultaneously.