Rockhammer, Thanks for reply, and we are going to wind up in complete agreement, as we solve "communication is man's biggest problem".
Yes, we are in total agreement on items 1 and 2 of your last post. Then you wrote.....quote.....
One thing I don?t agree with this statement:
?But most testers claim that the temperature "differences" reported by socket thermistors are the "same" as what the cpu sees, and that is "not accurate", it's wrong.?
Yes it would be wrong to contend that, but from what I?ve seen, most reputable testers have at least made some effort, if not gone out of their way, to point out the limitations of the socket thermistors. end quote.....
So here again we agree on the technical issue, but perhaps have sampled different ratios of sites reporting test results, and the key words are "most" and "reputable". Our experience is that "most testers" are not even aware of the technical problem, and they are the ones spreading the erroneous interpretations of test results. We have no problem with the "reputable" ones who clarify the socket thermistor's limitations. Problem is that "most people" don't know the difference between the two types of testers.
Quote again....
Also, you said: ?The mechanism which causes the problem is that the measurement is being made in a secondary heat flow path instead of inside the cpu. The reported "compression" of cpu temp differences is the nature of the heat flow circuits, it's not a measurement error. It's an error of interpretation by people who do not understand the physics nor the issue.?
I partially agree with that. Measuring a secondary pathway is often not the ideal, but many times it is all we have and occasionally it can yield superior results. Check out this quote from the Pentium III spec sheet:
?Note: The reading of the thermal sensor connected to the thermal diode will not necessarily reflect the temperature of the hottest location on the die. This is due to inaccuracies in the thermal sensor, on-die temperature gradients between the location of the thermal diode and the hottest location on the die at a given point in time, and time based variations in the die temperature measurement. Time based variations can occur when the sampling rate of the thermal diode (by the thermal sensor) is slower than the rate at which the T
junction temperature can change.?
That is from Page 48 of this document:
ftp://download.intel.com/design/PentiumIII/datashts/24526407.pdf. Given that statement, it might be argued that it is actually be better to measure a secondary pathway since it will tend to average out the temporal and spatial variations that Intel?s internal diode is unable
to quantify. End quote.....
We are familiar with the many Intel app notes re their thermal diodes and their quoted shortcomings. However, we believe those comments are addressed to the issue of monitoring internal cpu temp for the purpose of preventing failure by exceeding some temp limit, which by the way, Intel has never published. To confirm that we refer to an app note about Xeon chips for servers. Intel describes how each Xeon is tested at some arbitrary high cpu loading and the varying accuracy thermal diodes are calibrated by implanting a unique "thermal byte" into each chips' info roms. Then in normal operation, appropriate readout equipment converts the status of the thermal diode into an equivalent thermal byte which is compared in real time to the stored limiting value in the info rom...for whatever appropriate action etc.
But you are correct in that no one knows for sure that the thermal diode measures the "hottest temperature". However in the context of using the thermal diode in a "comparative" testing program of cooling hardware where it's obligatory to use exactly the same loading app, then temporal and spatial temp variations should not be an issue. And in this context we are concerned not so much with knowing the temp of the chip, but in knowing as best we can the "change in temp" of the chip when subjected to different cooling hardware.
There is no way a sensor partway along a secondary heat flow path encumbered with its inherent temp compression could be superior in that regard.
There is a valid place for the socket thermistor...as an indicator of change in those chips without thermal diodes. If a system is stable at some max loading, note the thermistor temp...and don't worry until it increases significantly at the same or lesser loading. When it does it should signify that chip temp has increased also...time to worry.
Reviewing all comments in context of testing cooling hardware's effect on internal chip temp, think we are pretty good agreement all around. Thanks for helping to explore and expose this issue to the overclocking community.
John C.