Quote Originally Posted by Dua|ist View Post
But first of all we're talking about temperatures at load around 70C (whichever Tjmax they used, this is much above the 50C, so no "slope error" here, right?).
That's what I used to believe but that's not what Intel said at IDF. Their version is that there is some error at TjMax and slope error goes from one end of the temperature graph to the other which makes sense based on how temperature sensors work.

They called the amount of error at TjMax 'X' which is somewhat useless information but does confirm that there is some error and TjMax is not a fixed, written in stone number. Perhaps X means about plus or minus 5C. You might think that plus or minus 5C is way too large an assumption but that is actually a tighter tolerance than what Intel lists their 45nm Atom sensors at when tested at 90C.

If you put a handful of processors in a 70C oven and ignored the slope error for the moment, they could indicate anywhere from 65C to 75C based on the data coming from the on chip digital thermal sensors. If you take into account slope error then the temperature range might even be a little wider than that.

What I'm trying to say is that we're all putting too much faith in the quality of data coming from these sensors. The 45nm sensors are far worse than I ever imagined. How can you possibly compare temperatures when there is such a huge amount of error in these readings?

And then if we assume their measurements are correct, -9W actually seems not that small difference, and it resulted not in -1C but in +2C according to their testing, which is what confuses me most.
That just confirms that uncalibrated temperature readings are worthless. If power consumption goes down then the core temperature will also go down. If it doesn't, then either the core temperature reading is not accurate or the power consumption number is not accurate or maybe both.

I did some testing today while running Prime95 small FFTs and an 8 watt reduction in power consumption at the plug was resulting in a 2C drop in core temperatures. That's with a Tuniq Tower on high at 4000 MHz and a core voltage of 1.40 volts. I tried to keep everything as equal as possible and used the Clock Modulation feature to switch back and forth between a duty cycle of 87.5% and 100%. This reduced power consumption by about 8 watts while still maintaining the same CPU frequency and voltage while running Prime95 small FFTs. That might not be 100% scientific but it's probably pretty close. The CPU temperature was in the low 50C range while testing.

I've always found that if a Core CPU is running too hot then Prime95 small FFTs will fail or LinX will fail or your computer will re-boot. Too hot depends on how hard you are pushing things. The harder you overclock and over volt, the cooler you'll have to run to maintain stability. If you are overclocking then your temperatures will be nowhere near the thermal throttling or thermal shut down point so there's no need to worry about what temperature you're at. It's just a number and for most people that are unable to calibrate due to sticking sensors or simply don't bother, it's not even a very accurate number. Overclock as high as you want and let stability be your guide.