Depends on what you mean by 'reading'. Figure 1 looks at readings in ADC counts, and while that IS what you get from the basic hardware, my process immediately dispenses with ADC counts by direct conversion to mV, which of course DOES require inclusion of the 'reference voltage', WHATEVER the value, as a necessary means to find mV/LSB. So I use 3000/65536 from my 3V reference and 16-bit AtoD mode, others use Vref=3300mV, and certainly you could use 3600mV, which on 16 bits would simply indicate .055mV/LSB. A sample 738mV would come in as 13,435 counts, and that times 3600 is some 48million, thence divided by 65536 is back to 738. There are probably 'minor effects' of Vdd on the measured voltage to the extent THAT possibly varies the 'current source' used to drive the Vbe measurement of this on-chip temperature sense, but only as a VERY MINOR secondary effect. In ANY case, without further calibration efforts this process will net you +/-10C at best, including datasheet tolerance on Vtemp25 and other sources.
But once the equations reduce to mV, then the final (integer) "25000-(Vtemp-Vtemp25)*inverse_slope" is the single equation to return temperature in mC. The on-chip temperature voltage is an absolute voltage varying with temperature only. Hence, the datasheet spec that Vtemp25 is an 'absolute' (nominal) 716mV and an (inverse) slope of 617mC/mV (recent K60 data). So using the previous sample-example, 25000-(738-716)*617 yields 11,426mC or 11C.
For a reference on the absolute-voltage-nature of on-chip-temp-sense, see also:
Silicon bandgap temperature sensor - Wikipedia, the free encyclopedia
