Lex wrote on Oct 14th, 2010, 5:41am:1. They reduce mismatch of current due to otherwise different drain voltages.
Just to clarify/expand a bit.
M12-13 could have very different drain voltages from M1 unless the cascode is used. Also, the power supply rejection of the M1/M12/M13 mirror will be poor if the channel lengths are short. You could get similar performance without the cascode by making the the channels longer at a cost of more cap on the drain and slower response of the M1 mirror.
For verification, I like to make sure Vds > Vdsat+100mV over all corners. Preferably 150mV greater.
The "bottom" device of a mirror should have gm as small as possible (usually limited by bandwidth)to minimize noise and mismatch caused by Vt mismatches. On the other hard, the cascode device should have as large of gm as possible to increase gm*ro and minimize the pole due to the impedance looking into that node.
rg