There's definitely something funny going on with timers, as the command-line sleep utility seems to sleep much longer than intended, especially if the device is saving power. Verifying with perl sleep:
# perl -e 'printf("%d\n", sleep(10));'
117
i.e. a 10 second sleep took 117 seconds of wall clock time. This after disabling tohd to save power. I began to wonder about this when powertop runs never seemed to finish, or even start (preceded by sleep 120...)

EDIT: when the system is not idle (i.e. interactive and screen is on), the sleep timing seems to be accurate to the second. However, it starts to be off as soon as the screen goes dark.

Started and completed while the screen is interactive:
# perl -e 'printf("%d\n", sleep(10));'
10

Started one second (=reaction time) after the screen goes dark:
# perl -e 'printf("%d\n", sleep(10));'
17