I've been kind of hoping a post like this would show up, though I was reluctant to start one.

The short answer for me is that I find it easier to make alarms in real world values (thus floating points) and for certain calibration curves for temperature measurements it's very hard to avoid using them. For the long winded answer, see below.

Most of my applications involve (among other things) temperature measurements at cryogenic temperatures for superconducting coils and protection of such devices. The temperature sensors (Ruthenium Oxide to be precise) we have evolved to are 2 wire resistive sensors with a resistance range from ~130Kohms to ~100K (~20kelvins to 290 Kelvin’s respectively).

We calibrate the devices in-house and the R vs. T curves generated for each sensor different enough to require a calibration curve for each sensor. The possible temperature change in our superconductors is slow enough (seconds to minutes for a significant temperature change) that the overhead for processor intensive floating point conversions are not very costly in performence.

These measurements are mainly for alarm purposes, not scientific (though it's nice to be as accurate as possible), so a 12 bit A to D in an 8051 based SOC is deemed OK for our purposes (uses SILABS C8051F040). As I mentioned above, it's easier to use real world values for "alarming" purposes.

We put a precision current source through each sensor, measure the signal back though a diff amp -> the on board AtoD, get the sensors resistance, and then run resistance through a conversion routine to get temperature.

Thats all I can put out today, time to go home.
have a good one

Andy