JohnZS: Your sensors are still near the top of my list of bizarre sensors so thanks for posting that. I originally came up with the idea that some of these sensors move along different slopes. At the IDF, Intel jumped on that band wagon and agreed with that theory. With your sensors, slope error might be a minor issue but it doesn't seem to be the major issue. The slopes for all 4 cores are quite similar and none of your sensors are sticking.

Your data would lead one to believe that during calibration by Intel, TJMax may only be accurate to something like plus or minus 5C. The 45nm Atom sensors have Intel documented error amounts similar to that at TJMax but my contact says that the 45nm Core 2 Duo based sensors are not like that at all. After looking at your data, I guess I'd disagree. It's too bad that Intel has decided not to share the amount of error at TJMax with enthusiasts.

I recommend using Prime95 Small FFTs when testing but I tried using "Maximum Heat and Power Consumption" like you did and the results were very similar to Small FFTs. On Max Heat, the power consumption is very consistent at each level just like Small FFTs is. The maximum power consumption at the wall was a few watts less so the peak temperature was about 2C less when I tested but each step down in power was unique and very consistent. Linpack testing creates more heat but it doesn't have the consistency that Prime95 has.

I'm still looking at data at the moment. When testing, you'll get the best results if you make sure that there isn't a lot of junk running in the background. Vista has a lot of background baggage like SuperFetch you need to watch out for.