View Full Version : CPU temperature results

Terry Porritt
17-12-2004, 05:49 PM
Thought I'd report on some cpu temperature testing, it isn't that scientific but could be of interest.

I bought the following, A DSE case with 3 built in temperature probes and front read out, an el cheapo MSI KM4M-V motherboard with onboard video/sound, An Athlon XP 2400+ 266FSB cpu with stock AMD heatsink and fan, A Thermaltake Volcano 11+ Xaser heatsink and fan with fan speed control options. I also lapped the copper base of the Volcano to improve surface finish and flatness. Flatness is estimated to be within a light band or two.

I opted to use the manual speed control for the Volcano.

First of all I checked the DSE temp. probes against a Fluke thermocouple/digital multimeter, with all in a glass of warm water at about 45deg C. The probes all agreed to about 2 deg C. The Fluke had been checked a long time ago against the ice point and in boiling water.

I mentioned in an earlier posting that the AMD heatsink surfaces were not flat, and that I lapped both sides of the plated copper base plate and the surface of the aluminium heatsink to a reasonably high degree of flatness.

One DSE probe I taped to the cpu next to the core using 'thermal tape', one is dangling in the case and one is positioned at the front air inlet to a 120mm case fan. (There is another 120mm case fan at the rear).

A program called PCAlert came with the motherboard, and this displays cpu temperature as reported by the BIOS from a thermal diode in the cpu core, system temperature from I dont know where, and cpu fan speed.

I have 3 different heatsink thermal compounds: Titan "silver" grease from DSE, which is not silver but probably aluminium particles/flakes in a silicone grease base. This has a very nice soft consistency enabling a very thin film to be applied;

Arctic Silver 5, 'polysynthetic silver thermal compound', this is rather stodgy and a stiff backed razor blade seems the best way of applying a thin film. The greyish colour of the silver also suggests that it is silver oxide rather than pure metallic silver.

The third compound is a white silicone grease that came with the Volcano, and it says Silmore-Taiwan on the packet. The blurb with the Volcano says Dow Corning T340.

I have tested both heatsinks each with all 3 thermal compounds with Windows XP running 'at idling'.

Now the Volcano noise is just not tenable with the fan running flat out at 4800rpm, the quoted noise level says 48dB, and 17dB at 1300rpm.
I found 3000rpm gave just about acceptable noise, so that was the speed for the tests.

OK, so after all that preamble here are the results:

NO significant difference in cpu temperature between either of the heatsinks or the thermal greases.

Typical readings are:
CPU temperature (from the thermal diode),54 deg C; System temperature,25 deg C; Ambient air temperature 20 deg C.

With Volcano fan running at max speed, cpu temperature 52deg typical, 25 system and 20 ambient as above. That is only 2 degrees lower than at 3000rpm

The DSE probe attached to the cpu next to the core reads typically 15deg C less than the core thermal diode.

Now Thermaltake publish a cpu temperature rise against watts dissipated curve for the Volcano 11+, which says at 70 watts the temperature rise is 24deg C.
So, if the fan is at max speed, the cpu temperature should not be above 44-49 deg, and really somewhat less as an idling 2400+ will be at about 60/65watts?

So what does all this mean?????

Well, if the heatsink surfaces are really flat, and have the finest surface finish possible, then the real area of metal to metal contact is maximised, and only the faintest trace of heatsink compound is required.
Under these conditions the thermal grease film is essentially isothermal, ie no temperature difference across its thickness, and so the results confirm what I had suspected: viz. that you can use any old thermal grease.

The stock standard AMD heatsink and fan is also more than adequate, and of course costs little when you get a 'retail' boxed cpu.
The proviso is that it needs working on to provide much better flatness and finish, something not everyone can necessarily easily do.

The other significant aspect is the difference between built-in thermal diode temperature, and that obtained from a probe against the cpu case.
Some boards have a thermistor inside the socket, and it is meant to be bent up to contact the underside of the cpu.
I have this on my Abit board, again I suspect that the 'real' core temperature is about 15deg higher than the reported 42deg typical for Athlon Thunderbird 1.33Mhz.

Now before anyone starts claiming their cpu runs at 35deg C when the ambient air is 20deg or more, just stop for a moment and think about the physics.

Graham L
17-12-2004, 05:57 PM
What are you playing at, Terry? Bringing facts in to spoil a subject which is full of good "theories", even better opinions, and faith in advertising.

Disgusting. :eek:

Murray P
17-12-2004, 06:47 PM
What are you playing at, Terry? Bringing facts in to spoil a subject which is full of good "theories", even better opinions, and faith in advertising.

Disgusting. :eek:

Too b****y right Graham.

For a start this thread should have been posted in the Chat forum Mr Porritt, more appropriate to what the typical post will be, ah hum, including this one :D

The other thing we can say for sure from this evidence is, the Thermaltake HSF is pile of overblown, noisy junk.

One more thing, we'll have to find a way to drop your CPU in that glass of warm water along with the other measuring devices while it's at idle. I think I've already solved the dropping in the glass bit, someone else can work on the, while it's idling bit, ok.

PS. The rest looks fine, excellent work :@@:

17-12-2004, 07:00 PM
well....as long as each setup was given three days use in order for it to show what its longterm running temp will be(they can drop by 15 degrees in the first 3 days after a new install),and that the probes were tested for reliable readings up and down the temperture spectrum and not just in 1 glass of water,that room temperture and case temperture were contriolled so that each test had an equal base,and that the system was tested at idele,and at work,half load,full load and a few hours sustained load would do the trick.

Would also be a good idea to see how each setup operates in poor conditions,such as a hot room,or a poorly ventilated case.

Besides which,rather then machinging the bottom of the heatsink its far easier to rub a small out of paste into the surface to fill up the irrglarites and then wipe off the excess,resulting in getting a perfect contact with the cpu.....which of course only gets a razor thin covering......

Terry Porritt
22-12-2004, 03:45 PM
I was going to post further on this topic when I had some thermocouple readings from the contact surface of heatsinks with the cpu die, but it looks as though the thermocouples on order wont now arrive before Xmas.

So here are some further simple physics sums to explain some of the strange statements surrounding cpu temperatures and thermal compounds.

Let us take the very basic steady state heat conduction equation, and we will take this as read, you can all look it up in school certificate physics, or any Engineers Diary, Machinerys Handbook, or even the internet, or whathaveyou.


Q is the steady rate of heat conducted, Watts; K=thermal conductivity, Watts/metre.deg C; A =area through which heat is flowing, metres^2; ΔT temperature difference deg C over a length ΔL metres.

Let us look at two thermal compounds, white generic Dow Corning T340 silicone thermal paste, and Arctic Silver5 silver compound with "polysynthetic oil" base and 88% thermal conductive filler.

The thermal conductivity of Dow Corning T340 (from interner search) is quoted as 0.73 Watts/Metre.degC.

The thermal conductivity of arctic silver translated from the crazy American system of mixed units on the Arctic Silver website comes out at greater than 8.9 Watts//Metre.degC

{When I say crazy units they are mixing metric system watts and metres with inches. That doesnt give much confidence in their figures or credibility does it? But that is all we have to work with, and the 'greater than' comes from their site too.}

The particle size for the conductive filler in Arctic Silver is said to be 0.49 micron, a micron is 10^-6 of a metre,ie a millionth of a metre.

So the absolute minimum thickness of Arctic Silver that you could have between two perfectly smooth surfaces is say 0.5 μm

Now real surfaces have hills and valleys, ie a surface finish that can be measured with a surface texture measuring instrument like a "Talysurf".

The smoothest most common engineering surfaces belong to gauge blocks, or slip gauges, as any toolmaker can tell you. These are precise length gauges that can be 'wrung' together, and any residual film between the surfaces will be less than 0.02μm.

So to consider thermal compound film thicknesses of 0.5-1.0 μm between the polished surfaces of a cpu die and a heatsink is quite reasonable.

The other characteristic of metal surfaces is the small real area of contact as opposed to the apparent macro area of contact. It can be as low as one hundreth of the apparent area of contact. It is this factor which goes to explain why the classical Amonton law of friction gives a friction force dependent only on loading and not on area.
But of course they didnt have gauge blocks in Amontons day.

{Anyone who wants to do more research into the physics of surfaces in contact, and the real area of contact could start with reading the classic work of Bowden and Tabor, The Friction and Lubrication of Solids, around 1950, with later reprints and additions}


Consider a film of thermal compound separating the cpu die and the heat sink surface.
Im going to assume very smooth lapped surfaces so that the amount of compound in the surface texture is small.

For the sake of the calculation let the film thickness = 1.0μm
Let the rate of heat transfer=70 watts.
Let the die size = 11mmx7mm, the size of an athlon XP 2400+ this equals 7.7x10^-5 square metres.

For Arctic Silver the temperature rise across the thermal compound is

ΔT= 1x10^-6 x 70/7.7 x10^-5 x 8.9 = 0.1deg C

For Dow Corning white thermal paste the corresponding temperaure rise would be:

ΔT = 1 x 10^-6 x 70/7.7 x10^-5 x 0.73 = 1.2deg C

So there is about a 10:1 ratio between the two compounds, but since we are only talking of ~1deg C for white thermal paste this would hardly be measurable given the variabilities between operating setups even in the same computer.
This confirms my measurement of no significant difference between compounds.

NOW, we come onto the the Arctic Silver statements and that of Metla, that after a period of time the cpu temperature drops.
Arctic Silver say 2-5deg C after about 200 hours, Metla says upto 15 degC after 3 days.

How can this be explained??????

One thing is sure, the composition of the thermal compound is not transmogrified by alchemical means into a different wonder substance. :D

No, any cpu temperature drop can only occur by increased heat transfer through the compound and into the heatsink, and this can only happen by the thickness of thermal compound decreasing with time.

So what thickness change will bring about the quoted temperature drops??

For Arctic Silver5
For a Δ(ΔT) of 5deg C, the corresponding change in compound thickness works out at 0.05μm, (thats 0.002 inches or 2 thou for those that like inches).

Whereas if we take the change of 15degC quoted by Metla, then the compound thickness would have to decrease by 3 times that or 0.15mm (an incredible 6 thou) :eek: That would suggest the compound has been trowelled on.

Obviously if the surfaces are lapped quite smooth and flat (and the cpu die already is, you cant lap that, you would just make it non-flat), then only the thinnest film has to be initially applied, a film ideally that is 0.5μm thick (it cant be thinner), or say 1μm, and NO cpu temperature drop over time will be observed because the film is already as thin as it can ever get. QED.

On the other hand if the heatsink surface is very rough, and we only have 1/100 of the apparent area in contact, then most of the compound will be in the valleys of the surface. The peak to valley depth will be many times the thickness of the compound at the asperities, and the temperature drop across the interface will increase as the surface roughness increases.

To carry more rigorous sums we need to be able to estimate the real area of asperity contact between clean surfaces under Hertzian contact loading, and to estimate the mean surface texture depth that will be filled with compound, as far anyway as a compound particle diameter of 0.5μm would allow.
This could be done with extensive effort with a Talysurf and statistical analysis, and I have done similar in the past, but it aint worth the effort.

Now the another point Metla raised was carrying out better temperature probe calibration over a range of temperature.

It just isnt worth it. Two points is enough, room temperature which I didnt mention but assumed people would realise that, and an elevated temperature around the temperaure of interest.
Else the ice point, and the boiling point of water could be used for accuracy. These probes are linear enough for our purposes, if they werent they wouldnt be any good would they?

So in summary, the better the surfaces in flatness and texture the better the heat transfer because the thermal compound can be very thin, ~1μm or less,
and differences between compounds are not really that significant.

If any significant drop in cpu temperature is observed over time, then it means the compound was applied too thickly to start with.

What I plan to do next is to drill through the AMD heatsink at the centre of where the die comes, normal to the surface and bond in a thermocouple with the end dressed flush with the surface, and the surface lapped 'flat'.
This will measure temperature directly at the die surface.

The heat transfer will be reduced slightly because of the hole diameter. Hopefull report on this in the new year.

22-12-2004, 03:53 PM
Wow Terry! I was lost half way... but you could probably pass a PhD with this! :D

Terry Porritt
22-12-2004, 04:30 PM
Oh gosh a bad typo:

>>>For Arctic Silver5
For a Δ(ΔT) of 5deg C, the corresponding change in compound thickness works out at 0.05μm, (thats 0.002 inches or 2 thou for those that like inches).

That should read 0.05mm NOT 0.05μm

Sorry about that it makes a world of difference.

22-12-2004, 04:35 PM

And stop quoting numbers i pull out of thin air...

Anyhow, It will be interesting to see what results you come up with from the next stage,You should seriously consider having this work published on a site somewhere.

Wonder what Whetu makes of it.

For what its worth,I don't use Artic Silver,I use their ceramic paste, And near every comp I build sits at about 30 degrees as reported by the inbuilt hoodacky whatsit.

22-12-2004, 05:04 PM
That is some great scientific work you have done there Terry, the second stage of your investigations should be interesting as well.

This article is just in time for my annual summer paranoia about my CPU temperature .. :rolleyes:

22-12-2004, 05:28 PM
This article is just in time for my annual summer paranoia about my CPU temperature .. :rolleyes:[/QUOTE]

This year we may not get the summer thing, or may allready have had it, judging buy the last week or two.


Terry Porritt
22-12-2004, 05:36 PM
It's all fairly basic, but the relationship between temperature rise, flatness, surface finish and thermal compound thickness should be explained more fully by the manufacturers of heatsinks and thermal compounds.

Here is an overclockers site that looks at thermal greases and does the same calc. as I've done more or less:

But notice that he derives correctly an expression for thermal resisitance and the units are deg C/Watt if you check the formula. Then he quotes the manufacturers crazy mixed units for thermal resistance as deg C. inch squared per watt for a 0.001 inch layer.

That is just plain confusing.

Looking through my school exercise notebook for 1st year sixth form, I think that is year 12? There is a nice experiment for measuring thermal conductivity using the LEES' disc apparatus, I wonder if our Graham knows about that

The main purpose for fitting a thermocouple against the cpu face is to calibrate the internal cpu thermal diode. I cant think of any other way of finding out how accurately that temperature is reported by the BIOS.

22-12-2004, 08:00 PM
According to AMD they don't recommend grease on "lidless" ie no heat spreader CPU because it tends to dry out after a time leaving voids especially if it was thinly applied. Phase change pads are reasonably foolproof and a few degrees don't matter so long as the case get decent airflow.The grease they do approve does change with heat and becomes denser.

As far as thermal probes go, even CPU with on-chip diodes have the problem that the diodes aren't in the hottest part of the core so some correction has to be done to compensate.

Terry Porritt
22-12-2004, 09:05 PM
Following on from what Pauls says, here is what AMD says on the the topic:

here (http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/26951.pdf)

I think drying out is somewhat of a red herring, especially when you think of power transistors, where conventional white thermal paste is most often used.

Maybe after about 5 years? one should think about renewing the paste.

The other thing is that I notice is that AMD thermal pads cost $10.80 each from Ascent. So if you remove a stock AMD HS&F it's going to start getting expensive to replace the phase change pad.

I have been using Titan TTG S101 silver grease on an AMD 1.33 Thunderbird for a full 2 years now, with no observable temperature rise due to any drying out or leaking out, so I'm reasonably happy about that aspect.

22-12-2004, 10:35 PM
Without wanting to state the obvious, but doesn't the real value of thermal grease come in once you start overclocking?

I mean, the fact of the matter is that you are measuring temperatures of CPU's which aren't even under load or overclocked.

Once you start overclocking and measuring temperatures, along with ambient temperatures, etc then you get some real value. The way i read it, you are forgetting about the benefits of stock cooling vs aftermarket cooling in overclocking or high stress applications.


23-12-2004, 12:55 AM
what I'm starting to wonder is what your trying to achieve,The conclusion that your heading to is that stock cooling is fine on a cpu running at stock speeds,and that thermal paste does its job as designed.

Not exactly new knowledge, Or anything that had to be proven.

Murray P
23-12-2004, 01:29 AM
It's very interesting all the same. There is also a lesson that we shouldn't take things at face value, especially what the marketing droids tell us. Of course we all know that anyway don't we, and why, and how to spot it every time.

Just digging for diggings sake has it's rewards too. Observing a good navey at his work, well a bit of that ethic can rub off.

Good stuff TP :thumbs:

From what I got the gist of, overclocking will still not be adversley affected by the thermal compound you use, the smear and the surfaces (and how close you can get them) will. So, there's a very good reason for the fearsome clips and springs that clamp our shiny bottomed, expensive, heatsinks down on the CPU other than the fact that every now and again the machine is going to get a boot across the room :badpc:

Terry Porritt
23-12-2004, 08:03 AM
Dead right Murray. Since the temperature drop across a THIN film of thermal grease is about 0.1deg C for high conductivity greases and say 1 deg C for white paste at 70 Watts, then no matter how much you overclock the thermal grease is not going to make any significant difference.

The contact conditions are important, and so is the heat transfer coefficient or thermal resistance of the heatink.
Many of the exotic HSFs quote very low thermal resistances like 0.2 C/W as against 0.5C/W for more stock units, but when you look at the noise, those figures are achieved at round about 50dbA, ie very noisy, too noisy for a normal person to endure.

If you are trying to get rid of 70 watts at 0.5 C/W then the temperature rise will be 35C, and with an exotic unit going full bore it will be 14C. The basic physics cant be denied by wishful thinking or inaccurate temperature read out.

Actually I'm not at all interested or concerned with overclocking. I mean if you wait a year you will get a cpu that is twice as fast as last years souped up overclocked cpu with shortened life, and which people then try to off-load on TradeMe :D

Also Im only concerned at this time with AMD Athlon XP, so like has to be compared with like.

Im also concerned with the accuracy of motherboard temperature readouts.

23-12-2004, 09:22 AM
Actually I'm not at all interested or concerned with overclocking. I mean if you wait a year you will get a cpu that is twice as fast as last years souped up overclocked cpu with shortened life, and which people then try to off-load on TradeMe :D

Well... I dunno about that... My Northwood P4 3.0 was bought at least a year and a half ago... basically when it was the brightest, shiniest thing I could lay my hands on... it can clock quite nicely to 3.5 with a little tweaking, and a year and a half later, clock speeds haven't sent this little beaut into the lower end of CPU's yet!

As for your points, I guess what I'm saying is that I don't think your results were that surprising to me. Thermal paste isn't going to magically lower the standard temp at stock speeds, or while the CPU is at idle.

I'd also question using just one type of CPU too! To prove the test, I would've loved to see how the Prescott chips fared. Oh, and also under stress would've been nice too... like a CPU burn in test through Sysoft Sandra?

At the moment, the test looks like comparing the performance of a car running on LPG vs a car on 98 octane - but instead of measuring performance, etc, you are measuring them while their engines are idling and aren't moving anywhere!

Good work on the calculations side of things though... lots of numbers which didn't mean a lot to me! ;)


Murray P
23-12-2004, 07:12 PM
At the moment, the test looks like comparing the performance of a car running on LPG vs a car on 98 octane - but instead of measuring performance, etc, you are measuring them while their engines are idling and aren't moving anywhere!

Good work on the calculations side of things though... lots of numbers which didn't mean a lot to me! ;)


You have to establish a basis for your tests before you can start looking at variables, of course the test has to be designed to exclude as many variables as possible, or be altered and re-measured as it progresses to exclude them.

The idle speed that Terry chose is arbitary, he could have chosen full blast, overclocked or whatever, as long as all were exactly the same so that results could be compared. However, I suspect that it is easier to get consistency, without the influence of of benchmarking software, by testing at idle, less effort required and less danger of damaging the CPU. It's still a measure of the "performance" of the object of the study.

The beauty of all good studies/theories/evidence, is that counter arguements can/(must) be forewarded, try to disprove it rather than prove it and you can't go wrong. Even if you do disprove it you have gained something. Prove it (in your own mind) and your study is likely to be invalided, sooner or later.

BTW, Intel are not going for increased clock speeds to increse the "power" of their CPU's so much as they're increasing cache and cores.

Terry Porritt
23-12-2004, 07:46 PM
I see PF1 is back on line good oh.

Murray is right again :thumbs:

Neither the thermal grease, nor the heatsink know if you have overclocked :lol:

All that we are concerned with in these tests and considerations is the amount of energy/second in watts that is coming out of the cpu and through the thermal grease into the heatsink. The more constant that is the better for purposes of comparison and calculation.

We hear a lot of figures bandied about without the benefit of background detail and conditions of measurement. So there is a lot of cross talk.
All I'm trying to do is to bring in a bit of science and rational thinking, rather than 'wild' statements.

The thermoucouples arrived today!

I had forgotten that drilling through the centre of the heatsink would hit the spring clip.

So what I've had to do is make up a small drill jig to drill a 1.2mm hole at 45deg to the surface of the heat sink, so that the thermocouple wires can be brought out through the fins to one side of the clip.

The AMD heatsink has a 3mm removable copper base and I've driiled that 0.6mm diameter in the centre to take the thermocouple.

The thermocouple has been calibrated in crushed ice/water for the ice point and in steam corrected for atmospheric pressure for the steam point.
Purified Water BP was used, all very scientific :D

Stay tuned for more results.

Graham L
24-12-2004, 02:20 PM
One thing which most people seem to have missed is the need for good contact between the surfaces. Extrusions are not famous for flatness. Even "powerful" springs aren't going to get good contact if you have .020" of bow.

I wonder what results you could get using Wood's metal as the "gap filler". :D

Graham L
24-12-2004, 03:29 PM
This Application Note (http://www.intel.com/design/network/applnots/30114801.pdf) from Intel might interest you.

The Pentium figures for Junction-Case resistance of 0.21 C/W and 33 C/W for Junction-Board are pretty good.

(A power MOSFet I use has 1.19 C/W for Junction to Case, and 0.5 C/W for Case to Sink. If it's just mounted on a PCB, the Junction to Ambient of 62.5 C/W would cause problems a bit before the dissipation of 120W is reached, even though the Max Tj is 175C.)

"°" doesn't work any more to produce the degree symbol here; perhaps I should use "K/W" rather than degrees Celsius/W. That would confuse everyone.:D

Terry Porritt
28-12-2004, 08:09 PM
First, I appologise to those whose attention span has been conditioned by 10 second sound/video bites, this is long-winded :eek:

In a previous posting, I described drilling the AMD stock heatink to take a small thermocouple. Here is a photo of the copper base plate that is screwed to the aluminium heatsink. If you look carefully there is a small spot in the centre. That is the thermocouple bead lapped flush with the copper.

The photo quality is not that good as I only have a very very old video camera or a webcam for pictures.

Perhaps a word about laps and lapping.
In general engineering cast iron laps are most commonly used. These are flat plates with grooves cut into the surface into which surplus lapping paste can go. Ideally one would have a range of plates for different grit sizes as it is difficult to remove coarser grit from the pores if fine grit is to be subsequently used.
I have 2 plates, but use only 400 grit or finer. Wet & Dry silicon carbide paper on the lap or a surface plate with kerosene or turps is ok for either coarse work, or to give a polished finish after first lapping flat on the lapping plate direct.
Then there are brass laps and pitch laps for optical work.

Be aware that using wet & dry paper will cause the edges of the object being worked on to round off, so it is difficult to make small objects flat all over that way. Trying to lap a cpu die would be madness.

There are techniques used to keep lapping plates in a flat condition. The traditional method used 3 lapping plates in correct sequence against one another to produce 3 flat surfaces.
I use spot lapping and then check the lap against another known flat using engineers blue.

The AMD copper base was originally grossly out of flat by up to 0.05mm. It had all the appearance of being hand finished on a belt linisher before plating, not the sort of tool for this job.
The Volcano HSF had a good surface finish, and only required a light lap to bring the copper surface to a high state of flatness, and was then finished on 800 grit paper. The copper of the Volcano also seemed a bit softer than that from the AMD.

Here is another picture showing the AMD HSF with thermocouple lead, sitting on the lap I used, and there is a set of gauge blocks behind, one of which I used to cross check a small straight edge come thermal grease spreader I made from lapping a Stanley knife blade straight to within a wavelength of light.

I would have liked to have fitted a thermocouple in the surface of the Volcano 11+ Xaser HSF, but it is in the too hard basket at this time.

The thermocouple hole was drilled with a number 72 drill,ie 0.025in. or 0.635mm diameter. This is less than 0.5% of the area of the cpu die, so is assumed to have insignificant affect on heat transfer into the heatsink.

It would be really nice to be able to measure the power going into the cpu from the power supply, and I expect this could be done possibly by using a modified power supply to enable current flows to be monitored.

Failing that, the best that can be done as far as I can see, is to take the AMD quoted power values, which for an Athlon XP 2400+ is 68 watts maximum, 62 watts typical (whatever that means).

I looked around for a program that would load up the cpu, and found a package, cpuburn4.zip from here:

There are programs included for testing Pentiums and AMD K6 and K7 cpus, and floating point arithmetic is used to stress them.

I thought the CPU Usage meter in Taskmanager would give a reasonable measure, but it seems to shoot up to 100% quite easily. So I'm not too sure that 100% usage really means what it says. For example the burnK6.exe program for K6 cpus gave 100% usage even though the cpu temperature was lower than when using the burnK7.exe program which also gave 100% usage but ran the cpu 5 degrees hotter.

When running burnK7.exe, opening other programs or even just opening My Computer was exceedingly and painfully slow. So I think it reasonable to assume the cpu usage was 100%, and that this would correspond approximately to 68 watts.

So before giving results of measurements, lets recap. on the objects of this whole exercise,

1.Compare some different thermal greases to see if there are significant heat transfer differences.

2. Compare an AMD heatsink and 60mm fan supplied with a retail boxed Athlon XP 2400+ with a Volcano 11+ Xaser heatsink and 80mm fan.

3.Compare the temperature readout from the inbuilt cpu thermal diode with the temperature at the cpu die/heatsink interface, and also with the temperature readout from a probe taped to the cpu adjacent to the die. This latter measurement should be comparable to the use of an onboard thermistor contacting the cpu from under the socket, as on somewhat older motherboards.

Here is what I found:

1.Thermal greases.
I couldn't measure any difference in cpu temperatures between white Dow Corning T340 paste, Titan TTG-S101 Siver Grease, or Arctic Siver 5 silver grease. This was mentioned earlier.

The most likely reason is the thinness of the applied grease film.
The flatter the surfaces and the finer the finish of the surfaces then the thinner the heatsink compound can be applied.

It is possible that with very flat fine surfaces that white compound would give a thinner film than something like either Arctic Silver or Arctic Ceramique, both of which have particle sizes of about 0.5micron.

Certainly I did see when I removed the Volcano, was that where the die had been, using white compound, was almost plain copper surrounded by white compound, and a perfect imprint of a small rougher texture rectangle that AMD cpus have in one corner of the die. This was more pronounced than with using Titan or Arctic Silver

Also with fine finishes, the 0.5 micron particle size of Arctic compounds could be too big to get into the surface surface texture to fill the valleys.

Flatness of the surfaces, in the final analysis, is more important than surface finish. There is no point in polishing a heatsink with 2000 grit paper to give a mirror finish if the surface is out of flat by say 0.5micron or worse.

The two measures should be commensurate, say aim to get 0.05micron for flatness by lapping, and 0.05micron Ra surface finish from 600/800 grit paper.

2 & 3. Heatsinks and CPU Temperatures.

It is easier to lump these two topics together.
Temperature readings from a thermocouple at the fan intake, room temperature, and from the AMD heatsink have an estimated uncertainty of not more than +/- 2 deg C, with the readout being on a digital multimeter reading to 1deg C. The probe next to the die I'd also guess at +/- 2deg C

The accuracy of the built in thermal diode temperature is unknown.

The following readings were taken over the Christmas period 24/26 December when the ambient room temperature was remarkably constant at 21 Deg.

The AMD 60mm fan speed is 3924rpm given by MSI PCAlert monitor program.
I set the Volcano 80mm fan speed to 3125rpm, a speed that had a similar noise level to the AMD and just a bit faster peripheral speed.

For brevity I will give the results for the heatsinks both having Titan Silver Grease as the thermal compound, but the readings for the other compounds taken a week or so earlier were virtually identicle as mentioned previously.

Firstly with the cpu idling, just sitting at the Windows desktop, 0% CPU Usage from Windows Task Manager:

Ambient 21 deg C.
Fan intake 25 deg C
Heatsink surface 46-47 deg C
Thermal diode 57 deg C
Probe next to die 40 deg C

Ambient 21 deg C
Fan Intake 24 deg C
Thermal diode 56-57 Deg C
Probe next to die 39 deg C

Next I ran each heatsink in turn with the cpu stressing program burnK7.exe, and with the CPU USAGE meter reading 100%. Temperatures seemed stable within 15 minutes, and readings were taken after 30 minutes running.

Ambient 21 deg C
Fan Intake 26 deg C
Heatsink surface 64 deg C
Thermal diode 69 deg C
Probe next to die 52 deg C

Ambient 21 deg C
Fan Intake 25 deg C
Thermal diode 67-68 deg C
Probe next to die 51 deg C

Then with burnK7.exe running, 100%cpu usage, the Volcano fan speed was wound up to its maximum reported 4600rpm, these were the temperatures.

Ambient 21 deg C
Fan Intake 27 deg C
Thermal diode 63-64 deg C
Probe next to die 47 deg C


1.The first thing to notice is that the difference between heatsinks is minimal.

2.The cpu thermal diode is reported by the BIOS as hotter than the heatsink/cpu interface by about 10 Deg C when the cpu is idling, and about 5 deg C hotter when the cpu is running burnK7.exe and cpu usage is said to be 100%.
This may not be unreasonable considering that this diode is buried somewhere inside the cpu.

3.If it is assumed that the cpu is consuming 68 watts with burnK7.exe running, then we can estimate the thermal resistance of the AMD heatsink, as (64-26)C/68W = 0.56 C/W.
This value seems to about what is quoted for similar looking non-exotic HSF units.

We can also estimate the thermal resistance of the Volcano when it is running at top speed if we assume that the cpu interface temperature is 5 deg C less than the thermal diode, and we get about 0.46C/W.
This can be compared with the manufacturers temperature rise curve versus watts dissipated which has a slope of about 0.34 C/W, not that good a match, but then increasing the speed and the noise a hell of a lot didn't do that much to the temperature.

5. As this work involved many installs and removals of heatsinks, I placed some wood pieces between the motherboard and its mounting plate so that no bending strain would be placed on the board. The Volcano was not that easy to fit, and I had to replace a small screw that fouled against a capacitor, with a countersunk one. The PSU also had to be removed each time to fit the Volcano, but the AMD HSF was small enough not to have to do that.

4.As a disclaimer, these results apply only to my set up, and I am open to criticism, just hope I havent made too many typos or transcribing errors.

5. Now to finish, an appropriate tune from 1929, Turn On The Heat, the Jack Hylton Orchestra:


28-12-2004, 09:20 PM
Thanks Terry, for that very interesting info.

It would be interesting to repeat these tests with some other branded coolers to see if your calculated thermal resistance is close the the claims, and if it makes a difference over the stock equipment.

I wonder if it would be acceptable to speculate then, that a better method of bringing cpu temperatures down is case ventalation, to bring the ambient temperature of the cpu cooler as low as possible.