CPU temp vs. CPU core temps

[SquadBumrush]

Hey gang!

Just wondering if someone can offer some clarification between these two measurement types. I'm overclocking an i5-750 using various monitoring/benchmarking/stressing tools (OCCT, Sandra, Everest, etc.) and it's unclear to me exactly how these measurements factor into overclocking stability.

For example, Everest reports CPU temperatures using a singular value and per-core values.

The singular value is typically much lower (~5-10 degrees) that the highest individual core temperature. So low, in general, that it is unlikely that this value is the average/median/mean of the 4 individual cores. Perhaps this is actually the ambient temp?

Intel's whitepaper on the i5-750 states that maximum "normal" operating temperature is 72.7 degrees Celsius. Does this actually mean the maximum of the highest individual core temperature?

Thanks!

 

[Deleted member 184142]

CPU temp is for the socket, not the chip it self, so don't even worry about it, all you care about is core temps. The CPU temp can also be way off, mine sometimes reads as below ambient and other times as 200C+. Also TjMax is 100C, not 72.7C.

 

[SquadBumrush]

Found more info on Everest's site that supports the reply above:

Q#34:Why is the "CPU" temperature considerably different from the core temperatures?
A#34:While the "CPU" and "Motherboard" temperatures are measured by the sensor chip mounted on the motherboard PCB; the core temperatures are measured by the CPU itself, using on-die temperature diodes. Due to the fundamental differences in the measurement solution, it's not recommended to compare those values against each other at all.

 

[King Icewind]

Also TjMax is 100C, not 72.7C.

No, 100C is when the CPU starts throttling itself lowering performance and will possibly cause damage if used all the time. 72.7C is the safe zone where none of that will happen.

You can manually change the TjMax. Coretemp's default is 100C.

 

[Zero82z]

No, 100C is when the CPU starts throttling itself lowering performance and will possibly cause damage if used all the time. 72.7C is the safe zone where none of that will happen.

You can manually change the TjMax. Coretemp's default is 100C.

Incorrect. The CPU can be safely run up to 100C. The thermal shutdown point is at some unspecified value above 100C. 72.7C is the thermal specification which is based on the TCase temperature, so it is therefore completely meaningless to us since we have no way of actually measuring TCase. The number we really have to worry about is distance to Tjmax, and not because if the CPU gets too hot it will become damaged (it's impossible to damage an Intel CPU as a result of temperature alone, since it will shut down before it actually gets hot enough to incur damage), but because we want to keep the CPU far enough away from Tjmax so that it won't end up hitting the thermal throttling point and slow down, which will result in reduced performance. A distance of 10C or more is generally enough breathing room.

Tjmax is a value that is set for each CPU in the factory and cannot be changed. What can be changed is the Tjmax that each temperature-monitoring program assumes. Those programs use the assumed Tjmax along with the distance to Tjmax (the actual value reported by the temperature sensors built into the CPU) to approximate the actual core temperature. Because each CPU has a different actual Tjmax, we can only guess which number we should use based on Intel's target Tjmax for each CPU model.

 

[SquadBumrush]

Because each CPU has a different actual Tjmax, we can only guess which number we should use based on Intel's target Tjmax for each CPU model.

Interesting. Do we have any reliable data about the average variance from assumed Tjmax (100C) for a given CPU? Or is it a matter of gradually heating the CPU to the point of self-induced shutdown (which would theoretically shed some light on the CPU's actual Tjmax)?

Sorry for the newbie questions. I've been conducting my first CPU overclocking experiment based on the (apparently misinformed) idea that the magic number 72.7C is the definitive "do not exceed" value. You suggest that Tjmax-10C is a reasonably acceptable limit, but since Tjmax is obfuscated, it seems like tripping the PROCHOT# and subtracting 10C is the only indisputable way of arriving at that magic limit.

For all the attention that gets paid to temperatures when overclocking, it's a wonder that this critical piece of information is basically assumed by the various tools that we use to monitor performance.

 

[Zero82z]

Interesting. Do we have any reliable data about the average variance from assumed Tjmax (100C) for a given CPU? Or is it a matter of gradually heating the CPU to the point of self-induced shutdown (which would theoretically shed some light on the CPU's actual Tjmax)?

There is no way to determine the actual Tjmax. The only thing we can do is use the target value and assume it's close enough to be accurate. Since the important thing is distance to Tjmax anyway, it doesn't matter if we assume that it is 100C as long as we base all our comparisons off that number.

Sorry for the newbie questions. I've been conducting my first CPU overclocking experiment based on the (apparently misinformed) idea that the magic number 72.7C is the definitive "do not exceed" value. You suggest that Tjmax-10C is a reasonably acceptable limit, but since Tjmax is obfuscated, it seems like tripping the PROCHOT# and subtracting 10C is the only indisputable way of arriving at that magic limit.

Like I said above, you can still use 100C. The thermal throttling point is at Tjmax, so if you set it to 100C in Real Temp for example, you know that "100C" is the number you have to stay away from. If you were to set it at 105C or 95C, then those numbers would be the values to stay away from. So the actual number itself isn't important. What matters is the distance to Tjmax, which is always reliable (okay, not always, but that's a different story, and it's always good near Tjmax anyway).

 

[SquadBumrush]

Thanks for taking the time to share this information. I think I'm starting to better understand that, for all the tools/metrics that are available, at the end of the day finding that limit is still just a matter of taking baby steps toward system instability and then taking one last step back.