Consider this scenario: You are running Availability Group in Synchronous-commit mode between two independent datacenters. You cannot have any data loss and the application is very latency critical.
I would like to have an option so that the transaction is committed, when it is in memory on both replicas. The hardening of the log to disk can be delayed.
If one of the servers would crash, so that you are running on only one server, then the hardening of the log should no longer be delayed.
This can be considered safe if the datacenters are fully independent (including power). It is then very unlikely that you lose both replicas at the exact same time.