Replica Bias schemes
You always get better matching by direct copies, or
instance-ratios, than by geometric dimension scaling.
By using one "copy" as feedback to an op amp based
bias distribution, you force the setpoint to be "what
works" at the business end, rather than "what you
would like to get" at the head end.
Sometimes you want to separate bias racks while
keeping them identical in setpoint, trying to segregate
noise or cross supply domains.
Not sure I understand what you are after (what is the
replica referring to?).