Curvature compensation CMOS bandgap reference

[mcsquare]

I am designing a bandgap reference using a typical temperature-independant voltage architecture. When i sweep the bandgap reference (1.2V) across temperature from -40 celcius to 100 celcius, i get 1.19V at -40 celcius and 1.205 at 100 celcius. How do i compensate this. I thought the curve is supposed to be a parabola curve, but mine is not a complete parabola. What i can do with it?

 

[blue_orangee]

you could adjust the ratio of resistances.

 

[electronrancher]

Trust me, they never come out where you expect them to, although it sounds like you are damn close. I have a process that i feel the models are exceptional - made by the simulator company itself for the low price of $30k USD. These models still simulate bandgaps 2 LSB lower than they come out - ha! Silicon is a tricky beast.

Anyway - your cell has a rising tempco, so you have too much PTAT. If this is a brokaw cell, you'll need to lower the "tail" resistor a little, or raise the PTAT resistor a little. This will lower the amount of (+) tempco and with a little tweaking you should be able to get the parabola you want.

Another way to think of it is this... A bandgap basically adds a Vbe (-2mV/C) to a PTAT voltage to make a flat voltage about two diodes high. If your diode is 650mV, and your ptat voltage is 0.259*ln(8) = 53.8mV. You will probably want a 12:1 ratio in order to gain that 54mV up to 650mV. For a 2k ptat resistor this gives a 24k gain (tail) resistor in the brokaw cell. But you always need tweaking... If Vout rises with temp, reduce tail resistor. If it falls, increase tail resistor.

I highly suggest that you build many, many simple bandgaps before you try to add the curvature compensation circuits. Adding in a curvature-correction circuit can mask a real problem, and give a good result in sim that will be horrible in silicon. I have seen this mistake. Both the reference and the curve-correction were slightly off, and the total sum could not be flattened. The guy would have been better off just using a simple reference.

 

[wchenx]

you should also consider the nonlinear case, there is always some difference of voltage vs T.