I am a little confused on the light purple bar - Synchronized. Is this the overhead for coarse grained locks or locking without any TM? It seems unfortunate that the increase in overhead for STM even with all of the optimizations is still 2-3x worse than simple synchronization. But I guess this is where the advantages of TM come into play (atomicity, isolation, serializability).

@tennis_player I think you're right that this is the overhead for coarse grained locking. An important note is that this overhead is measured for a single thread, because even though STM looks 2-3x worse that simple synchronization, one could imagine that you could potentially get much faster overall execution when using many threads with STM over coarse grained locking, because the additional overhead would be amortized by gains from being able to concurrently work on TM operations.