The red numbers represent measurement points for 3 monitoring techniques: as we all know: what you need for proper monitoring are temperatures both for the CPU-package & "system" [air-intake to the fansink from inside the system-case]. 

It is the relationship between these two which is important, rather than the absolute accuracy of either. This relationship defines the efficiency of the fansink - its °C/W s-a

What you're doing is measuring heat in & heat out, then - to find out a fansink's efficiency - dividing the difference by how much power the CPU is using.

4 > 5 = "bcase1" - the default monitoring setup on a socket370/A motherboard: the in-socket sensor will either be mobo-mounted approximately in the centre of the socket, or on a wire spring - which may be bent gently up to press the sensor against the centre-base of the CPU. "System" is sometimes taken from the monitoring IC itself, or off a mobo-mounted sensor, or a "roving probe" 

2 > 1 = "tcase" - the conventional technique followed [more or less] by fansink & CPU manufacturers: a small thermocouple is inserted through a hole in the fansink base & secured - together with its connections, by some form of epoxy; the base & thermocouple/epoxy are then flushed-off. Air-intake temperature is taken by a shielded sensor adjacent the fansink's cooling-fan.

3 > 1 = "bcase2" - burningissues' technique: a cheap unshielded bead-thermistor is pressed up against centre-base of the underside of the CPU-package by foam insulation filling the socket370/A; a second [encapsulated] bead-thermistor adjacent the fansink's cooling-fan takes the air-intake temperature.

All these techniques have specific faults, both theoretical & [what we're interested in] practical.

"bcase1" 

The ordinary mobo sensor > mobo IC > monitoring utility package gives completely unacceptable results for CPU & generally poor results for "system" [air-intake] temperatures. 

Individual components may be of fair quality; but we have found the performance of the package [in every modern motherboard examined] to be deeply flawed in linearity. By this we mean, that when you set the CPU temperature to read accurately when the system is at HLT idle, it will read inaccurately - generally a great deal too low - at 100% CPU-load. The appalling error-level we found in a webs'-worth of GOrb tests - an average error of 57% & all on the low side - reflects this non-linearity as well as testers' basic failure to calibrate motherboard monitoring utilities.

Linearity errors come partly from measurement error from the under-socket sensor - especially when the sensor is mobo-mounted - but also apply to wire-mounted sensors touching centre-base of the CPU. 

On our present test-system [MSI 694D-ProA] each wire-mounted sensor becomes ~6C out over a ~19C range. Neither baring the tip of the sensor nor applying h/s grease significantly alter this for the better.

These errors are clearly due to the data supplied by the manufacturer of the monitoring IC; this is more-or-less tweaked for the specific model by the mobo-maker & the resultant algorithms included in the BIOS code. All & any motherboard monitoring utilities call on this & faithfully report what they are given.

Imperfect though it may be, this package is what we all have & must make the best use of; Burningissues has kept focused throughout on any outcome of superior accuracy in fansink testing being of direct & clear benefit to all us users with this default package. 

This page has accurate data from which many users will be able to calibrate their motherboard monitoring utility - below we look at two ways of getting this data.

"tcase"

CPU-packages are carefully designed to offer minimum resistance to heat passing through the case-top - the current FC-PGA/TBird/Duron packages have an extremely short path from their minute heat-producing cores upwards. Any technique attempting to measure from anywhere but adjacent the case-top relies upon an assumption that their alternate path is proportionate, over the measured temperature-range, to this most direct heat-path. This diagram is an impression of the heatflows being measured, & from which points:

As you see, a tcase sensor directly measures change in the major heat-flow from within it; importantly this is adjacent the source & along the direction of change. 

By contrast, any "bcase" technique is indirect: it is like applying a stethoscope to the patient's chest, rather than dangling a microphone down their throat. A wit might argue bcase is like applying a stethoscope to the feet while asking for a loud cough.

To the naive, this might appear to make the tcase technique so unarguably superior that it's not worth considering an alternative. However, if we examine more closely the practicalities when using tcase this superiority is less clear:

The tcase technique, in its present form, is to insert a very small thermocouple into a hole perhaps as small as ~0.8mm diameter drilled in the fansink base: the quality [& expense] of the components is high, as is the specified resolution of temperature differences.

 - here is a diagram - equally simplified - of differences between tcase & bcase2:

Whenever you compare two ways of measuring the same thing; one interventionist, the other diagnostic, the benefits in an approach can be considered in relation to its aim & ideal. 

The tcase technique aims for directness, & its ideal is a "point" measure - an infinitely small thermocouple in an infinitely small hole: this would [obviously] measure without affecting what is being measured. The reality is different:

a) measurement intrusion: however small the thermocouple seems, it is relatively large in relation to the tiny core of modern .18 process [soon .13] CPU's. 

Even if the hole drilled for it is as small as possible & the securing epoxy minimal, a substantial part of the direct heat-path between core & fansink is obscured &/or altered in thermal resistance. 

This means - obviously - that the fansink's performance is being altered [probably degraded] by the measurement process in a non-quantifiable way.

b) placement: temperature across the tiny core of a CPU varies - we have been quoted the figure of 8C across the core of a PIII. 

If the thermocouple is claimed to be sufficiently small not to create serious intrusion-errors, then it is small enough to resolve such differences across the core. 

To get repeatable measurements unaffected by this across-core difference the thermocouple [& the fansink it is within] must therefore be placed in precisely the same relation to the exact same spot over the core. 

This means the base of any fansink tested must be drilled - say by means of a jig located in relation to the fansink's clamp - with precision. 

The fansink when in use must then be located on the lugs of the socket370/A with equal or greater precision . . . . .  

 . . . . . . A reasonable observer looking at the clamping-system of fansinks knows this to be an unattainable goal.

c) "absolute" error [plural]: These expensive little thermocouples are carefully produced in batches & warranted to measure absolute temperature within a certain margin of error.

However, each fansink requires a new thermocouple; this maximizes the allowance which must be made for error between comparisons.

d) feedback: thermal conduction to & from the fansink body, the epoxy, &/or the connecting wires to the thermocouple is exceptionally difficult to quantify. 

If this is minimized [by using, say, larger volumes of epoxy of known insulation properties] then measurement intrusion - as in: a) above - is further worsened.

e) joint quality: The thermocouple/epoxy combo must be flushed-off to the fansink's base: in practice this means that no fansink "tcase-tested" can be in "outta-the-box" condition, since a flushed-off base will have a flatness & finish different [in most cases better] to that common to the manufactured item. 

Refinishing noticeably affects the possible thickness & thus component °C/W of the heatsink grease [what you in fact measure, then resolve by calculation to °C/W s-a, is °C/W j-a; this is the sum of the °C/W's of the fansink & heatsink grease/compound].

f) expense: disproportionate [several hundred dollars minimum]: overall accuracy of this technique depends on matched accuracies between air-intake & CPU-package temperature sensors & their instruments: such accuracies are priced in direct proportion to outlay. 

These practical issues, added together, make the tcase technique flawed. It is, however, the conventional "approved" technique - this factor could be an attraction to some.

We considered it is a last resort if no other technique could give useful results; & took a thorough look at those alternatives which involved a real CPU in a real motherboard.

[there are obvious techniques using artificial heat-sources, useful for private comparisons, if done well: see: 2cooltek & overclocking.telefragged for crude implementations]:

bcase2

The technique we decided to try, & the ordinary mobo-sensor-package, share the theoretical disadvantage & practical advantage of measuring CPU temperature through-package. 

As you see; there are many possible heat-paths for waste-heat to travel through. The proportion of this waste heat to that going upward to the fansink base varies between package designs; but is obviously minimized by design in all, & is unlikely to exceed 10% of the total. 

More importantly, bcase is a measurement of inference; not only is the heat-flow being measured in an unknown [& possibly changing] proportion to core temperature; but it is in the reverse direction to the heat-flow of interest - that from the core through the fansink to the air.

These are obvious & profound theoretical disadvantages to any bcase technique; to balance these, its aim is simple: to be non-invasive, cheap & practical in use.

a) any non-invasive [diagnostic] measuring technique has the categoric advantage of not affecting the process being measured.

b) measures the fansinks as we buy them: "outta-the-box" without performance-changing modifications.

c) you can change fansinks & CPU's just as easily [or not] as usual, without disturbing the sensors: this also means all sensors are shared between comparisons, increasing basic accuracy.

d) since it measures "through-case" it's analogous to the default sensor/IC setup on motherboards: °C/W s-a figures measured throughcase translate well to real-world use on real motherboards with real CPU's

e) it costs $20 & one hour's work.

f) the sensors of the specific cheap digital thermometer used have a known matched performance & share the same circuitry: though their absolute accuracy may be imperfect, their relative accuracy - what matters - is easily tested [as good].

g) Burningissues' results - mostly with PPGA packages - compared against an exhaustive series of quite refined "tcase" tests carried out by a public-spirited fansink designer, are unexpectedly reliable & accurate. We have reason to believe our technique, carried out carefully, gives results inside 5% of absolute ºC/W 

Yes: it works - & to us this is sufficient reason to encourage other enthusiasts to try this for themselves. 

We would - of course - be glad to receive any criticism or ideas for improvement.