I thought that time would be much lower.. Any hack you know to improve it?
20us seems to be about average for "minimal" arduino functions.
There have been "improved" versions of HardwareSerial.cpp that (in particular) reduce the size of the FIFO pointer from 16bits to 8bits, which would help some. You could contemplate writing "Serial.write(buffer, size)" that writes more than one byte at a time, which should theoretically save some time, if you do it right. Other that re-writing HardwareSerial::write, I don't know of any way to make it go faster.