JohnZS, you're just the guy I like seeing data from. My goal is to turn your random number generator into some meaningful temperatures and I think I'm very close to that goal.
This new Cool Down Test is really helping me see what's going on. I took the Intel IDF presentations at face value but all they initially did was get me off track. Intel's graph of slope error was more of a generic line on a piece of paper than the engineering type data I was hoping to see.
With my test, when checking for slope error, I like to compare the 87.5% level to the 12.5% level. I shortened the heat up phase of my test so users wouldn't have to wait too long for the results. For that reason, the 100% level is not your maximum Prime95 temperature but it's usually pretty close to it. The 100% level is also a maximum temperature where 87.5% and lower are averaged temperatures. That's why I like to start by looking at the 87.5% level first.
Anyhow, when comparing 87.5% to 12.5% on your CPU I get the following temperature deltas:
Core 0 = 16.5°C
Core 1 = 15.3°C
Core 2 = 16.1°C
Core 3 = 16.4°C
This shows that slope error is only playing a very minor part in your wildly looking random temperatures. The majority of error in your sensors is at TJMax. The biggest problem is that each core has a slightly different TJMax.
Intel publicly stated at IDF that there is error at TJMax but has only been willing to define that number as plus or minus X for the 45nm Dual and Quad Core CPUs.
Here's the published spec for the 45nm
Intel Atom N270 sensors as listed in the datasheet:
5.3 Digital Thermal Sensor
The digital thermal sensor (DTS) accuracy is in the order of -5°C ~ +10°C around 90°C; it deteriorates to ±10°C at 50°C.
If this spec has nothing to do with the 45nm Dual and Quad core CPUs then in their IDF presentation they should have defined exactly what this spec is. They admitted that there is some error at TJMax but have only been willing to refer to the amount of error as 'X'.
Based on the Atom spec and the data coming from sensors like yours, I think assuming that X is somewhere around 5°C is reasonable. That means the Intel calibration point that TJMax is based off of might only be accurate to plus or minus 5°C.
The values released at the IDF conference are referred to as TJ Targets. TJMax is defined as always being equal to or above that Target value. In this case, TJMax would be more accurately defined as 105°C ± 5°C or somewhere between 100°C and 110°C. That range is a lot closer to the truth than the single TJMax=100C number.
That's the problem. Users including myself have always taken TJMax as a very fixed value with virtually no error in that number. The endless argument about what is the correct TJMax has always been flawed. There is no single TJMax number for every QX9650 processor or for any processor. The 65nm sensors may have had a little tighter thermal sensor specs or they might have been able to do a better job of matching similar sensors with each other on the same CPU but there is always going to be some error. Each processor within a given line is always going to have a range of correct TJMax values. Early testing shows the Core i7 is continuing this tradition.
Knowing this, some users are simply going to give up and say, "Accurate temperatures from these sensors simply aren't possible. There are too many unknown variables." That's what Intel has been hinting at since day 1 but the user community has never been willing to accept that.
My opinion is that 100% accurate reported temperatures may not be possible due to the limitations of these sensors but with a simple calibration, and a few tweaks, you're going to get a lot closer to an accurate temperature than what some of these sensors give you out of the box.
rge's recent testing gives us a point of reference at the low end and reading between the lines of what Intel has recently said gives us more flexibility at the top end.
The present RealTemp calibration formula needs to be reworked and users are going to have to be willing to adjust TJMax on a core by core basis in some situations. I'll come up with a new formula and some calibration numbers you can try later this week. I'm hoping for 4 temperature curves that look more like mirror images than random numbers.
My E8400 is listed as 1.225 volts on the box so there's nothing wrong with 1.25 volts. Didn't my previous post show you that these CPUs can take a licking!
Originally Posted by burebista
Reply With Quote
Bookmarks