For HBM, which stacks many DRAM layers, the precision of this step is effectively the yield. Each added layer makes joining harder, turning back-end processing into the bottleneck.