This solution loses 1 rate on firework w for geometry reasons, other than that inputs are pulled at pseudo-period 4 with min latency on the final products assuming the atoms in the last input are both used for the final product. It's likely that a better layout can save the 6 cycles lost on firework w. Pulling inputs at period 2 or 3 is theoretically possible but I'll be surprised and impressed if anyone pulls it off.