Maximum tolerated variance in the mode's frame-cost signal, in ms².
When the rolling variance rises above this value the provider asks the
indexer to decrease scaler quality. When it falls below a quarter of this
value (equivalent to half the corresponding standard deviation) the
provider asks for an increase. The band in between provides hysteresis so
the system does not oscillate around the target.
The signal the variance is measured on depends on presentation timing.
When frames are presented on a fixed display cadence — vsync, or any
mobile compositor — wall-clock frame time quantizes to that cadence and
masks instability, so the provider measures variance on CPU and GPU work
time instead. Otherwise it measures variance on wall-clock frame time
directly. For the full behavior of this mode, refer to
AdaptivePerformanceStableFrameModeProvider.
As a rule of thumb the value is roughly the square of the tolerated
standard deviation in ms: the default of 16 corresponds to about 4 ms of
jitter.