This is your problem. There is only one row in the pgbench_branch table, and every transaction has to update that one row. This is inherently a seriaized event.
Indeed it was!
One solution is to just use a large scale on the benchmark so that they upate random pgbench_branch rows, rather than all updating the same row:
pgbench -i -s50
With a scale of 1000, everything except the END took roughly the latency time. Interestingly, the END still seems to take more, when threads/clients are really ramped up (100 vs 8). Why would that be?
Alternatively, you could write a custom file so that all 7 commands are sent down in one packet.
How would you restructure the sql so as the make that happen?
