RobG wrote on May 6th, 2013, 8:56am:Not sure what the process problems are - if you remove R1 and R2 the output voltage will be pretty much process independent PTAT. Opamp offset and mismatches between the resistors will dominate the errors.
Also be aware that you have to compare this voltage to something - generally you will do an A/D conversion so the reference that you use for it may be a worse source of error. Or it may give you either a trimmed Vbe or a PTAT votlage you can use since the reference voltage is Vbe+PTAT.
The process variance of just the diodes vs. the bias current makes it a little more complicated.
The implementation I did on this many moons back was in the era of the Pentium I, when thermal monitors first got introduced to microprocessors.
The methodolgy used to get rid of all the calibration nonsense involved a pair of PN junctions with a geometry ration (8:1? 12:1? I forget.) and also a pair of high-low current (that was 10:1) into both of the junctions. With all that, you end up doing a set of difference measurements where most of the process dependent variables fall out of the equation, and you are left with purely a temperature dependent function. The geometry ratios and the current ratios become the control of the output differences, with all the absolutes falling out of the math.
Sorry I don't have more specifics here but I did that design way back around 1992, so the details are a little muddy.
Do a IEEEE JSSC search on this, there was a lot of temp sensor stuff done back there.