Note that in the pipelined case, it is bottlenecked by the operation that takes most cycles.

Throughout lecture, pipelining has been called a "free lunch" because it is a simple way to reduce the runtime of multi-stage algorithms that does not even require parallelizing within each stage. The pipelined program takes the exact same resources and utilizes them better than the sequential program, whereas other forms of parallelization may eliminate some bottlenecks but hit others based off of the limited compute/memory bandwidth available. The only downside of pipelining is the additional memory required for buffering, but typically this is a small cost to pay for the increase in efficiency observed.

The figure just reminds me of the compare between single-thread and multi-thread before.