Anandtech Socket A heatsink review

[Tillmann]

Hi,

it's true that it's not possible to know the exact power dissipation of the CPU when running "just any" app, but that doesn't matter - as long as the same app is running during all heatsink tests (and that's what I did), the power dissipation is constant during the entire test.

Concerning compound: You're right that compound is a major source of inaccuracies. Actually a much bigger source of inaccuracies than the kind of temperature measurement. I took great care to appy the compound in the same thickness for all coolers (I did that with a razor blade), and as said before, I tested each cooler three times to ensure the results are consistent. Between tests, I uninstalled the cooler and re-applied thermal compound.

Most of the coolers came without compound (Global WIN FOP38, FOP32-1, Hedgehog, Alpha, Super Orb), so I chose to test all coolers with thermal compound (I used the same thermal compound for all of them). If I had tested the coolers that ship with a pad with their pad, this would have skewed the results, e.g. the Taisol would have performed worse than the FOP32-1, which isn't true.

bye,
Tillmann

 

[Mikewarrior2]

If you aren't doing a full load on the cpu of worst case scenario's, then you aren't truly testing the heatsinks. Each heatsink has the possibility of being a different cpu usage when being tested. Try something like Prime95 or cpu burn.

also, you never state how long you waited for full load temps. Due to the thermistor reading setup, it takes quite a while for the air underneath the cpu to actually heat up.

third, you still haven't answered the fact that reading from a secondary heat source and using measurements based from those is wrong. The results are not what actual core temp changes would be.

As with the majority of socket A tests show, bad heatsinks can be made to look very, very good(ie chrome orb).

ALso, you claim we are complaining. well, yes, because the comparison is pretty damn friggin' fishy and wrong. on a p3 with a proper internal diode reading mb, the difference from a c-orb and a pal6035 with a Sunon fan is around 5-6C. Do you honestly mean to say that a c-orb mounted on a cpu pumping out 25W more heat performs within 2C of an alpha with a better fan?

And don't say anything about diode inaccuracies, because on a mb that supports true diode reading, a true heatsink comparison can be done because it is measuring temps at the PRIMARY heat source.

Mike

 

[Mikewarrior2]

Oh, depending on whether you heated up the pad to get it to "phase-change/melt" and what kind of grease you used, the taisol typically performs better with the PCTC pad. Care just has to be made in order to heat up the pad.

Also, if the kt7 thermistor was so damn accurate, and the algorithms so damn perfect, how do you explain people with 1.2Ghz t-birds @ 1.85 volts running at 34C load?


Mike

 

[Mikewarrior2]

Let is also be said, that based on how you loaded your cpu, your cpu was putting out only 30W of heat. This is derived from the measurements you took and claim to be very close to accurate.


Mike

 

[Mikewarrior2]

Update

I grabbed a chrome orb earlier tonight to test out on my own home system. Why? Because i wanted to get a theoretical C/W rating on the c-orb on a reliable temp comparison platform.

For this comparison, i used a taisol cek733... versus the chrome orb. THe taisol clip has been modified to place less pressure on the cpu core, but I did not do this with the c-orb. I removed the pads on both, and lapped both heatsinks.

Room temp was a cool 22C, and ambient case temp was a nice 24C. The highest I could get the c-orb to stably run my computer was 950 @1.85V(note, the taisol runs my cpu at 982 @1.85V). To "load" the cpu, i ran a simultaneous combination of cpu-overheat + prime95 for 30 minutes. This is more than enough time, as internal diode's do not take long to go from idle to full load.

Taisol Cek733: "full load" temp was 38C, or 14C over ambient. Observed heatsink value of ~.46c/w*

Chrome Orb: "full load" temp of 43C, or 19C over ambient. Observed heatsink value of ~.63c/w*

* Assumes cpu thermal output of 30W(actually, if this is lower, the c-orb only gets worse as far as observed c/w).

Now to compare these to the socket A readings. Assuming your "full load" was 45W, your chrome orb should have read somewhere in the vicinity of 57C, give or take a few degrees for the thermistor inaccuracy. Now, since you claim that the thermistor/algorithsm are so damn accurate, it is time to actually prove this to us.

Now, assuming a .38C/W(alpha pal6035 w sanyo denki fan) for an alpha, you would get a temp reading in the range of 46C, again give or take a few degrees. NOw notice that the socket A method does not accurate reflect actual cpu core temp changes, due to the Measuring of temp from a secondary pathway.

All my tests were done 4 times for each heatsink, with a complete reapply of arctic silver each time. Ambient Case temp was verified with an external probe(accurate enough for these circumstances).

Note: If we added a taisol CEK733 to your tests, it would, with your method of measurement, place somewhere in the 45-47C range. in actuality, cpu core temp would be roughly 49C. Again, the temps are so compressed due to secondary heat pathway measurement that full/actual changes aren't shown.

Thanks to the many people who have helped me increase my knowledge of thermodynamics. I am no expert, and am certainly a kindergardner compared to people like JohnCar, but I do know a little bit about this stuff .


Mike

 

[Tillmann]

Hi,

you can't compare the values obtained with an internal probe to those obtained with an external probe (may it be located in direct contact to the CPU core, or may it be the onboard thermistor). With Athlon CPUs, there is no internal probe - so what do you want me to do, contact AMD so that they custom-build me an Athlon with internal probe? There's no other solution than using an external probe, and according to my findings, it does not make a major difference (as far as relative temperatures are concerned) whether the onboard sensor is used or an additional sensor mounted right next to the core.

I'll try to make my point clear once again: I want the results to be reproducible, by as many people as possible. If I use the onboard thermistor, an AnandTech reader who has a similar setup can look at the temperature readings of his own board, and compare them to my results. Then, he can judge whether it would make sense to replace the cooler he's currently using by one of the models tested - he can get a idea how much cooling will be improved then. For readers, the results from the onboard probe are the most useful, and this is why I have published those. But I realized that some of you aren't happy with that. I try please as many people as possible, so in future tests, I'll also include temperature readings obtained with other methods.

bye,
Tillmann

 

[Mikewarrior2]

No Duh!!!

You can't compare the internal diode results to an external probe. But it was YOU!!! that claimed that the thermistor solution was just as accurate as an external probe, and it clearly is not. I've seen external probe tests with chrome orbs run well into the 50Cs, and your "under socket thermistor" does not accurately represent that.

The reason I am upset is that, and i've said this before, even though socket a temps tests are "repeatable" and "reproducable", that doesn't mean they are remotely accurate. YOu again fail to acknowledge that a disclaimer should be put up saying that the any socket A temp comparison should be taken with a grain of salt.

Might I also add that due to the nature of socket A temp measurement, only you can reproduce these results. NO other person can reproduce them.


Mike

 

[Tillmann]

Hi,

I said the 'under socket thermistor' would give similar (relative) results as an external probe, I never said anything about internal sensors (except that they're inaccurate).

47°C isn't that far away from "the 50s". If I had used another CPU, e.g. oc to 1GHz with 1.95V, then I'm sure the Chrome Orb would have run well into the 50s even with my measurement method.

Of course average users will not be able to 100% reproduce my results, but if they have a similar configuration, and bend up the thermistor and apply thermal compound, then they will get very similar results, and they can (to a certain extent) compare the heatsink they currently own to the heatsinks tested here. If I had used an external probe, this would not have been possible.

We're getting a little redundand, so I guess I'll stop posting here

bye,
Tillmann

 

[Mikewarrior2]

Okay, so I can get this right.

from your words, a difference of 47 read/interpreted temp is close enough to a 57C actual core temp to count as accurate?

And, also from your words, a difference of core temp that is actually 11C, but is read/shown by the mb as 2C is also acceptable as accurate?

And, not every socket A mb can bend up the thermistor. Guess you didn't think about that, huh? And I guess a 550% error in the comparison numbers is considered accurate in your opinion.

I dont' know, but as far as science goes, a 550% reporting error is pretty damn bad. And that would throw out any results from your "comparison".


Mike

 

[Maverick2002]

wow mike.......you should start a heatsink business of some sort.......

 

[Mikewarrior2]

Well, the whole problem is any old person can start a website. And a whole lot of cheap website reviews are done poorly/incorrectly.

The problem is when a reputable site like Anandtech, where I USED to recommend people to, starts putting out crap reviews. Like I have said repeatedly, a new user comes here, looks at this review, and goes, wow!, the C-Orb performs so good compared to an Alpha. And according to the thermsitor readings, it does look thsi way. I've explained why the readings are wrong, all I am looking at doing is seeing that new users are not confused by this review.


Mike

 

[Modus]

I guess to settle this we need to get a Socket 370 Intel CPU and compare temperatures between heatsinks using three different methods:

1) The on-die thermosistor.

2) A motherboard thermosistor making good, greased contact with the idle substrate directly beneath the core, sending readings to a stable, updated BIOS.

3) An external thermosistor mounted between the core and the heatsink in such a way as to be unobtrusive enough not to affect cooling.

I am no cooling expert, but as I understand it, Tillman is arguing that the relative differences between heatsinks as measured by methods 1, 2, and 3 would be quite similar, whereas Mikewarrior is saying that method 1 is more accurate than method 3, which is in turn more accurate than method 2.

Am I right in my assesment of the argument? Hasn't somebody done a test like the one I described to put the issue to rest?

Modus

 

[Haervii]

Uh, I guess you could say I'm a dumbass newbie, but I thought the TaiSol looked the best, but it's not around, so I thought the FOP 32 looked the best. That is what the review recommended.

 

[Mikewarrior2]

modus,

Unfortunately, method 2 isn't really testable on a p3. There isn't a via p3 board that has a long enough thermistor.

However, even assuming Tillman's claim that that method 2 can give just as good results as method 3, thermistor's touching the side of cpu core's are anywhere from 5-12%(with the 12% off being at load). With my taisol heatsink, and an external thermistor contacting the edge of my cpu core, i get a peak temp of 36C(versus 40C internal diode reading), and an idle of 28C(versus 30C internal diode). Internal diode inaccuracy issues aside, it is very safe to assume that the thermistor readings are a bit lower, since they aren't isolated from outside air contact, etc.

Now, the argument that an external probe is equal to that of an in-socket thermistor is possible, with a greased thermistor, etc, but it still will not show cpu temp changes correctly(ie like with my taisol versus c-orb test).

Not to mention that, if methods 1,2 &3 are so close to each other that it doesn't matter, why would the theoretical measurement on each heatsink be completely off.

For example:

Thermistor Reading / Theoretical Actual Reading

41C / 44C Kanie Hedgehog(.35C/w)
45C / 46C Alpha PAL6035 w/YS TEch Fan(.38C/w)
47C / 57C Chrome-ORb(.63C/W, derived above)

*Assumed 45W of heat from T-bird

Notice the highly inaccurate differences between thermistor readings and theoretical diode readings. Also note, that the theoretical readings would be absolute best conditions possible, requiring no heatsink grease. In all actuality, cpu core temps would be warmer than the theoretical numbers.

The reason for this inaccuracy? Measuring from a secondary heat pathway that is not isolated from outside air contact. Even if they have a set thermal resistance of the cpu pcb, each mb is different, and air still touches the thermistor.

Your assessment of the argument is correct, Modus, and I think i have shown significant and factual data to prove my point that no matter what is done, an in-socket thermistor will still be a very hit or miss solution when trying to "approximate" cpu core temp.

*See My 2:49 AM post for info on how I derived the C/W rating of the c-orb. I did use a s370 p3 and a compatible diode reading mb. The external thermistor is a asus external thermistor(same as the one supplied with A7V mbs).


Mike

 

[Mikewarrior2]

Also, if you used the bios to read the cpu temp, then you by default cannot be running seti@home.

And if you aren't using the bios(it sounds like, since you keep saying the bios implements "algorithms" to correct the temp measurement) to read the temp, what program are you using?


Mike

 

[johncar]

Tillman, MikeWarrior et al,
Just want to detail the secondary path measurement issue that Mike Warrior brought up and which Tillman has never responded to.

When 2 heatsinks of diferent performance are tested at some constant power dissipation/heat flow, two things happen...the better hs results in a lower internal chip temperature...AND...the heat flow thru the hs path "increases" because of the lower thermal res of ONLY that path. This means that the heat flow thru secondary paths, such as the substrate in contact with the mobo's thermistor, "decreases". This should be obvious and is easily proven. Now let's see what must happen qualitatively to various temperatures as a result of this change in heat flow split.

Assume the better hs lowered chip's temp by 10C, then we'd like to see the same difference at the secondary path thermistor...right???
But that would mean that the heat flow thru the secondary path had NOT changed, since heat flow is a function of "temp drop" from internal to thermistor, and a constant/unchanged thermal resistance. So this means that while the temp at the substrate does drop. it can NOT drop the "same amount" as the chip's internal temp.

Finally lets look at some numbers to try to quantify this effect.
We are told by chip mfrs that we can ignore the amount of heat that flows thru secondary paths. So lets assume a 90%, 10% flow split. then lets assume only a 3% higher heat flow thru the better hs path. But that's a 30% change/less heat flow thru the secondary heat flow path, which should result in a 30% CHANGE in temp drop from internal to thermistor, the thermistor temp change would be only 70% of the internal temp change...we are not talking small numbers, and believe we've used "very conservative" stats in the example.

The point is that testing components in the hs heat flow path by measuring temps in secondary heat flow paths is seriously flawed. The main reason users do it is because they are not aware of this flaw and are attracted by its appealling convenience. Serious testers/reviewers should not fall into this trap.
John C.

 

[Nevin]

Thank you John.

Here is a crude diagram illustrating what John is talking about.

Secondary heat path measurement errors diagram

Nevin House
Arctic Silver, LLC

 

[Mikewarrior2]

Thanks JohnCar for taking the time out today to explain that further, something I don't really know how to do.


Mike

 

[johncar]

Nevin, Mikewarrior2, et al,
Thanks, we all need to help each other present the facts and weed out the voodoo thermo. And here's an some interesting addenda to the post.

It obviously holds true for evaluating Arctic Silver as well as heatsinks....can't be done accurately with the cpu bottom thermistor.
Wish Arctic Silver reviewers would get the message.

Also, we install a tiny thermocouple with .005" wires flush with hs surface opposite case top dead center. Guess what should happen if we installed one in a better hs?? You got it...it should read HIGHER for the same app because of the higher heat flow resulting from the lowered res of the hs path. So anyone installing a better hs, and who has a good sense of temp in their fingers should not be surprised to find that the better hs runs slightly "warmer"...although the chip is cooler.

And same holds true if you replace zinc oxide with Arctic Silver, the hs should get slightly WARMER to pass the add'l heat flow induced in that path.

Finally, if we should worry that our interface material should degrade toward a higher thermal res, we would look for the thermocouple in the hs suface to begin reading LOWER, NOT higher, since the increase in res of the hs path would cause less heat flow thru the hs with attendant lower temp at its surface. You see, the TC is on the hs side of the interface...the cpu side/case top gets WARMER, same as internal temp.

However, if we get a lot of dust or the hs/fan slows down, then the TC should begin to read HIGHER. So we have to be alert to ANY change in hs surface TC temp.

Interesting how temps at different locations react to various changes.
But it's simple once users become aware of the heat flow shift principle.
John C.

 

[thermite88]

Mike, Nevin and John,

Secondary heat path measurements were done all the times in real world engineering and can be very accurate if conducted properly. I am qualified to address this because I do that for a living. My advanced degree and over 20 years of experience are in rocket and space system thermal design. (I can't tell you what I actually worked on.)

In most situation, the engineer cannot access the spot where temperature should be made. The same problem facing Tillmann. Measurement at the primary heat path is often unacceptable because it distrubs the heat flux passes throught. The probe, most likely a thermocouple, is position close to the measurement spot with low thermal resistance between them. The probe location is then insulated to the environment so that the secondary heat path where the probe is has a neligible heat loss compared to the primary heat path. Usually, a fraction of one percent is considered acceptable. If this condition is met, the measurement is very close to the actual temperature.

The heat loss can be minimized by super low conductivity insulation. For the most critical application, the heat loss is reduced to zero by actively controlled guard heaters. The thermocouple leads are arranged in such a way that they are routed along a isothermal path to minimize heat conduction throught the leads.

It is not difficult to measure the backside heat loss in motherboard socket. Commercial heat flux sensor can be bought for less than a hundred dollar. It will need radical surgery to the botherboard to install, such as drilling access holes, but it should be well within the resources of a commercial review site. If Tillmann can find an acceptable insulation to the backside of the CPU socket and prove it, it will substantiate his claim that "using the backside thermister is more accurate than the internal thermal diode". You can calibrate the thermister, but not the thermal diode.

It the CPU backside heat loss is as much as 10% of total, then the methodology in Tillmann's review is faulted. If it is only 1-2%, it should be acceptable for a consumer oriented review. Mike can win the argument hands down if he quantifies the CPU socket backside heat losses.

For now, the burden of proof is on Tillmann because he conducted the review.

 

[johncar]

thermite88 et al,
We agree with all that thermite88 has written, but want to restate the critical points he made and detail/clarify a related temp measurement issue.

Thermite has basically called for an isothermal environment between the internal cpu temp and the "sensing element" of the temp sensor, as well as reducing the secondary heat flux to negligible proportions. Both of these requirements can only be met with appropriate insulation and backside heat flux measurement as he outlined.

We stress "sensing element" not for his benefit but for all of those who have been drawing conclusions from thermistors making uninsulated point contact with cpu backside, "assuming" that thermistor is measuring temp of cpu backside, and some even suggesting internal cpu temp. Not at all, due to the "necessary isothermal nature" of all temp sensors' calibrations, all contact temp sensors respond to the temp of their respective sensing elements, whether a thermocouple bead or the semiconductor material "inside the body of a thermistor" making perhaps 0.1% contact with the cpu backside and 99.9% contact with air in the socket cavity. So the temp of socket cavity thermistors' sensing elements is somewhere between temp of cpu contact point and the socket cavity air temp, and highly variable from system to system. Of course, "carefully" insulating the thermistor and socket as thermite88 recommends addresses this issue as well as the minimum secondary heat flux issue.

Thanks thermite88 for adding the weight of your experience to this issue. Will be interesting to see what kind of response we see from various reviewers/testers.
John C.

 

[OneEng]

thermite88,
Although I no longer work in this area, I was in the Nuclear Navy for 6 years and have a very complete background in thermodynamics (for obveous reasons).

I believe there are two issues here. The first issue is about the accuracy and precision of the readings taken. I suspect that the majority of the people here have correctly stated that it is not all that accurate.

The second and more important issue is if the measurements taken can fairly represent the relative performance of each CPU cooler. I believe they can.

Reguardless of the percenage of heat transfer through the secondary path, a better cooler on the primary path would result in a lower temperature in the secondary path.

The repeatability of the test performed should also be quite good. The inaccuracies of the thermocouple should also be very repeatable and thus be irrelivant in determining which cooler is best. Keep in mind, this review was not attempting to determine temperature measurements accurately, it was simply trying to determine which cooler performed better.

The only point MikeWarrior made that I agree with is that although an Alpha cooler only shows a 2deg difference in this test, it could be quite signifigant in the actual core temp difference depending on the amount of heat transfer going on in the secondary heat transfer path.

The methods used do not make any difference IMHO in the relative outcome of the comparison. If the relative results do not match up with others posted on other sites, perhaps one should look to the other sites for sources of error?

 

[thermite88]

OneEng wrote:

<< The only point MikeWarrior made that I agree with is that although an Alpha cooler only shows a 2deg difference in this test, it could be quite signifigant in the actual core temp difference depending on the amount of heat transfer going on in the secondary heat transfer path. >>

The only point that you agreed with is the only point Mike tried to make.

The Taisol CGK742092 costs almost $40 shipped. The Alpha $36. But the Chrome Orb can be had for only $15. Mike felt that the temperature presented in the review is misleading because the $15 Chrome Orb would seem to be the adequate solution and probably the best buy. He argued that the Chrome Orb is a much poorer performer than even the cheaper Taisol.

I have the Chrome Orb. I cannot install it on the Asus CUSL2 now because it interfers with the capacitors next to the CPU socket. When I used the Asus P3B-F with slotket before, I found the Chrome Orb to be a very capable performer, comparable to the Alpha PALFC-35t with Sunon fan. Both kept the CPU core temperature of a overclocked P3-945 at 1.7 volts vcore to 10 degrees C above ambient at Prime95 loops. At 1.9 volts vcore, the temperatuer went up another 2C.

 

[Mikewarrior2]

Thanks Thermite88,

OneEng,

That was my only point. The rest of my posts were &quot;my proof&quot; of this point. Tillman has decided not to respond. I still believe irresponsible to leave review posted as is, because while the heatsinks are probably in correct order, they are very misleading.

Also, i feel this is a very interesting site.

Heatsink Tests

Mike

 

[johncar]

To clarify a minor issue....thermite88 stated that a thermal diode cannot be calibrated, but Intel says they can. They recommend a specific procedure utilizing special electronic kits from Maxim or Analog Devices?? which are interfaced to the thermal diode/motherboard and run with special software using another PC...not a very simple process.

Basically corrects for variable ideality factors of thermal diodes by using a forward and reverse current technique.

Don't have the specific url but specific ref was found at Intel's site with search engine set to &quot;thermal diode&quot; which gets a half dozen or so of related app notes.
John C.

PS:-One app note details that the variable thermal diodes of Xeon cpus, used on servers where operational reliability needs to be very high, are each calibrated during a max load operation that results in a &quot;thermal byte&quot; representing an equivalent &quot;max temp&quot; being written to an info rom in each chip. Then the output of the thermal diode, transformed into an operational &quot;thermal byte&quot; can be compared in real time to the stored value. This apparently does not get one a temp in deg C per se, but rather a means of being able to operate below some predetermined safe limit. Noted here to demo the variability of thermal diodes and Intel's cognizance and actions re same.