Hardware errors are gonna vary from device to device chip to chip, even wiht same mod applied. for starters theres an error in resistor values even at 1%, for the resistor (r52 mod) then there is plenty of other resistors on the board as well - the main one thats gonna effect voltage is r50, connected to DGND from the buck convertor, plus the resistors for the DPLL voltage. All these will give a variance although small, but when measuring voltage in.00 volts it can make a difference.
Secondly these chips are Intel or AMD chips, manufactured to the same constiant and quality control as thoose so each chip itself will have a different optium performance. Not to mention the whole assembly of the grids isnt to the high standards and quality control we would like.

CPUminer i believe uses the hardware errors to adjust the hash rate, no hardware errors over x period hash rate goes up. x hardware errors over x period hasrate goes down. until it finds a balancing point.
When running a batch of grids from the same command file - with setting each one indivdually - a varying amount of hardware errors will be produced, as each device will start of the same and autclocked until it gets the hardware errors. After running CPUminer for say 24hr, if you take a note of the hash rate for each grid, set a batch file wiht autotune off and set each one indivdually youll probably find over the next 24 hours that the hardware errors are a lot lower if none at all.