Anandtech Socket A heatsink review

[Mikewarrior2]

Anand's Heatsink Review

Hmmm... Quite honestly, I am disappointed in this review. The reviewer fails to mention that socket A temp reading is highly inaccurate, as this review clearly shows.

First off, there is only a 2C difference between a Chrome ORB and a an alpha pal6035. Only 2c difference between a heatsink that rated at .82c/w versus a heatsink rated at .38c/w.

Hmm, math aside, the difference between the two heatsinks(core temp wise), would be at least 15C, if not more.

This serves are more reinforcement for the fact that these readings, by virtue of reading from a secondary pathway, do not show the cpu load temp, nor does it show differences between heatsinks very well. not only are all the heatsink results highly compressed, they are not good reference for the unknowing Anandtech reader.

Any particular reader may come to the Anandtech website, and say, ooh, the Chrome orb is good enough, because it performs within 2C of an Alpha. The truth of the matter is that the chrome orb performs significantly worse than an alpha, only that the temp measurement method is so unreliable that is hows almost no difference. To the unknowing reader, he doesn't see this, and assumes the c-orb as a great heatsink.

This is truly the first time I have been utterly disappointed by a review here. While almost every single review provides an excellent reference for anyone from a beginner to intermediate to expert user, this review clearly does not. THis review does not provide neither an accurate, nor a good reference for those in the socket A heatsink market.

Lastly, for those who say the review is fine because they used the same mb, it is above-mentioned that the results are only a fraction of what cpu core temp would change in the test situations. A 2C difference in the review may be a 10C difference in core temp.


Mike

 

[NaughtyusMaximus]

I was waiting for you to jump in and say that about this review. In all honesty, it looks like the review was well thought out, and well implemented. The problem being that it wasn't thought out enough. Surely there must be better ways (PIII anyone) of measuring temperature.

 

[Mikewarrior2]

THe whole problem is, it does look well thought of to the average user. And it serves as potential confusion to the average/beginning user.

The problem with p3s is that they simply do not produce the amount of heat that duron/t-birds do. A heavily overclocked p3 at 1ghz at 1.9volts will push at most 35 or so watts. an overclocked t-bird(at 950, 1.8v) puts out roughly 55.

 

[NaughtyusMaximus]

Which poses the question:

How can you make an accurate heatsink benchmark?

I would probably lean toward the PIII, even though it has lower heat output, since it does have the thermal diode. Perhaps the results wouldn't reflect the heatsink's operation in (un)optimal conditions, but they would at least be accurate.

 

[Mikewarrior2]

I'm sure a big website like Anandtech would be able to "borrow" a thermocouple for testing purposes. As little as I do know about thermocouples, i do know that they are less effected by outside temps(idle substrate and air) than are thermistors, and are extremely accurate(way more accurate than an intel internal diode).


Mike

 

[etech]

Accuracy is relative. If you are just looking for a delta t, the absolute accuracy of the measuring device is not that important. The repeatability is more important.

I think 2Cool has the right idea in using a TEC for the heat source. I would add a temperature controller and a 1/4"? copper plate. The plate could be grooved on each side to accommodate a thermocouple. The smallest thermocouples I normally see are in 1/16 tubing. Use one to control the temp of the TEC and the other to measure the effectiveness of the heat sink.

That would give you a controlled and repeatable experiment.


If you want absolute accuracy look into RTD's. Thermocouples have a stated accuracy of +- 2 degrees when purchased.

 

[HellRaven]

I agree, I too felt it was not a very good review. The findings in it conflict with tests from overclockers.com, hardocp.com, and 2cooltek.com. It also goes against common sense.

Like Mike said, a newbie coming here that doesn't know much about cooling could be lead to believe that a chrome orb, while a good cooler for 800mhz or less processors, is a good choice for their blue core tbird or pick of the litter duron when they try for 1ghz. Why spend extra cash on an Alpha or Globalwin? I know I myself wondered that until I read alot of cooling reviews before I got my duron.

 

[Technonut]

I agree with Mike also. I have personally owned a few of the HSF's that were reviewed for my own tests with an external probe. (I know... not 100% accurate, but close as one could get)

I posted the same findings about the ArctiCooler awhile back regarding the HSF not making contact with the pads on the Duron/T-Bird, and the clip being very difficult to install. I do not agree that it cools as well at my tested voltage (1.85) as the Alpha, or FOP series coolers. At least they did not recommend it.

I have personally found the Alpha with a 38CFM Delta to be the best cooling solution for me. I know a lot of people do not like the noise, but I put up with it due to the great cooling it offers. I have found that even with an Y.S Tech fan, the Alpha does better than the FOP with the same fan. I understand the Taisol's are good also, but I have not picked up one of those to test yet, but probably will soon.

EDIT: The reason that I have not considered a Hedgehog is due to the fact that the places that I have seen offering it will not allow it to be returned, even if damaged in shipping because of the copper construction. I have also heard that it is not a good choice and damage prone if you install/uninstall it often, (which I am known to do) and needs the bottom lapped. It has less cooling pins than the Alpha, and the Alpha is of forged construction. I also wonder if it will discolor in time. Anyway, I'll just stick to my Alpha until the Kendon CPU Radiator is available in the States.

 

[Mikewarrior2]

Another Clear problem with the Kanie Hedgehog is that I've seen sites where, upon purchase from a store, the bottom was un-finished/un-polished. The workmanship/quality of the hedgehog varies so much, that a nicely lapped/handpicked(especially because of the soldered on pins), could easily outperform another heatsink. But is this indicative of the average purchaseable hedgehog? no.


Mike

 

[Caitiff]

I understand the points that the rest of you have made regarding this issue, but there were some points that I saw here that I haven't seen elsewhere. Case in point, the inability to mount either the Chrome Orb or SupaOrb on the Epox 8KTA series of boards. This information was vital to me, and I hadn't seen it anywhere else, and I have been looking. I also think that although the method of testing the heat was not optimal, it DID give a 'level' reporting tool across all the fans represented. I too would have like to have seen a temp spread done, no load vs heavy load etc, as they do on some sites, but the observations made regarding the various choices I thought were spot on. Just another perspective.

 

[DaddyG]

I agree with Mike's comments and add a couple of my own.

I think that for purposes of equalizing the test, that the PCTC interface material should be removed BUT a strong disclaimer stating that PCTC is the ONLY material recommended and approved by AMD should be added. The average John Q Public reading this review assumes that the PCTC pad is crap, which is most certainly is not.

Testing methods which use a 'program' to heat up a cpu are simply flawed. Just how many Floating Point Double Precision Multiply operation does it take to create 1 watt ??.

I think that Anandtech has started testing heatsinks with little real knowledge and even less testing equipment.

Update: Since the AMD approved heatsinks use PCTC which fills imperfections in the heatsink base, many manufacturers don't take the time to put a super smooth finish on the bottom of the heatsink. Since regular thermal grease does not fill the imperfections, this is another reason why the tests are flawed.

 

[Danlz]

Tillman is more than capable of burying most of us in thermodynamics. I think he did an excellent review for the general readers of this forum. His CPU temperature reading are not intended to be absolute, but relative; challanges to his readings and methodolgy are as certain as howling at the moon. Besides, temperature is only one small piece of the equasion when selecting a HSF, and many of us believe that bang-for-the-buck, noise, availability, and ease of mounting are of equal priority. Thank you for the excellent round up Mr.Steinbrecher!

 

[Mikewarrior2]

Danlz,

I am not disputing his personal knowledge. I am disputing his heatsink test. The fact of the matter is, when socket A temps are concerned, they are "relative", and are "repeatable", so they seem like good tests. However, when the readings are this far off(ie chrome orb on near par with alpha and taisol and others), they should not be reported. the fact remains that the test scores are a mere fraction of what cpu core temp change would be.

I am almost certian he knows more about thermodynamics than I do, but at this point, I believe i know more about why socket A temp readings are wrong than he does. For you to claim that us challenging his methodology and measurements is the same as howling at the moon, I implore you to read through the thread. The problems with this method are very, very clear.

Of course he isn't lying about his temp measurements. But his methodology is not good. He could be showing 1/5th of the change in core temp, or 1/3rd, or 1/2. who knows? His temps are right, but he clearly fails to mention the errors of relying on that thermistor to compare temps.

And to that, I will say, thank you for a very confusing review to the general and new computer user. Thanks to you, i'm sure Thermaltake will have to produce many many many more crappy chrome orbs.


Mike

Update: Ditto on the PCTC comment by DaddyG. Also, I do not subscribe to the school of thought that Anandtech "rulez" or Anandtech is always number 1. They are capable of bad reviews, as is any website on the internet. AND IMO, this is one of those bad reviews. The fan part was good, but it seems like a secondary thing considering the main thing people look at are temps.

 

[Tillmann]

Hi!

Let me comment about the points mentioned:

- Choice of the onboard sensor
I did do quite a lot of experimenting before chosing to use the onboard sensor for this test. E.g. I used a thin NTC installed in direct contact to the CPU core (just like in the Slot A Athlon comparison many months ago). It is true that the _absolute_ results obtained with this configuration were a little different from the results obtained with the onboard sensor (but only a little), but the _relative_ results (meaning: the CPU temperature difference between heatsink A and heatsink B) were identical in the cases I tested. Also, I've looked at more accurate alternatives to NTCs: Thermocouples are way too thick. Thin Film Platinium sensors are very accurate. but too thick, too (the thinnest I could find were Heraeus Sensor-Nite sensors, they are 0.9mm +-0.2mm thick, and that is too much). If anybody knows a source for thinner platinium sensors, let me know (tillmann.steinbrecher@anandtech.com).

For comparing cooler performance, only the relative differences between various coolers are relevant, and the onboard sensor does a good job measuring these - so I chose to use the onboard sensor: This way, anybody who has the same motherboard and CPU can verify my results.

- Accuracy of the onboard sensor
The type of NTC that is used for the KT7's onboard sensor has an accuracy of 1%, which is very good. However, what limits accuracy are a) the contact to the CPU (this depends on if and where the sensor touches the CPU, whether thermal compound has been applied, etc), b) the quantization errors from the chipset's DAC, c) inaccuracies in the comparison voltage and/or resistance used for the DAC, and d) inaccuracies due to imperfect algorithms for calculating the CPU temperature displayed in the BIOS. The inaccuracies caused by a), c), and d) only affect _absolute_ temperature measurements, those caused by b) are very small. In order to limit the inaccuracies caused by a), I bent the sensor a little up and applied thermal compound.

It is true that the accuracy of the onboard sensor isn't very good for absolute temperature readings (e.g. if the temperature displayed is 45°C, the actual temperature might be 44°C or 46°C), but as said above, I only care about the accuracy of the relative temperature differences - and for this, the thermal sensor is good, and the results are quite accurate, even if you don't believe them (Chrome Orb vs Alpha)

Some people tend to think that the temperature displayed by the BIOS is exactly the temperature measured at the thermal sensor, and therefore actual CPU temperature is higher than the temperature displayed. This is not the case. The BIOS measures the temperature underneath the CPU (which is lower than the displayed temperature), and then calculates the temperature of the CPU core from that value. Of course that's not as good as measuring inside the CPU, but as I have verified with a thermal sensor installed in direct contact with the CPU core, the BIOS does a good job calculating the temperature.

Before doing the test, I have discussed the testing method and the advantages and disadvantages of using the onboard sensor with other techies in great detail in a German tech newsgroup. If you're interested and if you speak German, I can find the thread for you on Deja. One of the guys had similar arguments as you, but later on agreed that using the thermal sensor is fine for this particular purpose.

- Intel on-CPU thermal sensor
The integrated thermal diode found on Intel CPUs is inaccurate, much less accurate than the onboard sensor on the KT7. You can see this very easily by looking at the temperature displayed: Sometimes the temperature will jump by more than one degree within less than a second. Intel doesn't even specify any data on accuracy. It's ok for checking whether the CPU is about to die because it's overheating, but it's not perfect for accurate measurements. I must admit I did use the internal diode of an Intel CPU for the last Socket cooler roundup, but I'm not going to do this anymore in the future. The test results of this roundup are more accurate than in the last one (and still, nobody complained back then).

- Chrome Orb vs Alpha
The thermal resistance specs come from different manufacturers, who use different measurement methods (and this _does_ make a difference - for example Global WIN specified the thermal resistance of the old VEK12, which was introduced in 1998, to be 0.16°C/W - LOL!). It is really true that the Chrome Orb is very close to the Alpha performance-wise. If you don't believe this, buy both and test yourself, with whatever test method you like. You'll never be able to measure 15°C difference, rather 2°C, or maybe 3 or 4 if you use a CPU with a higher power dissipation.

- Comparison of my results to the results published by 2CoolTek and HardOCP

a) 2CoolTek:

They use a Peltier element for testing. This is totally unsuitable:

- a Peltier element doesn't provide an exactly defined thermal load. Its power dissipation will depend on many factors, in particular the delta T between cold an hot side (if this value is high, power usage will decrease, because of the Seebeck effect). Also, the power dissipation depends on the thermal load on the cold side.

- The Peltier element doesn't have exactly the same size as the CPU die. Since the contact area is different, the results obtained with a Peltier element don't represent performance with an actual CPU.

- The heatsink clip (and the pressure it puts on the CPU) has a MAJOR effect on cooling performance. The higher the pressure, the better the thermal transfer. 2CoolTek's test doesn't take this into account at all.

Don't misunderstand me, I'm not bashing 2CoolTek. They are a great retailer, certainly among the best, and I highly recommend them for purchasing heatsinks. But the test results they post are often incorrect and misleading.

b) HardOCP

I've just looked at Kyle's tests again (e.g. http://www.hardocp.com/reviews/cooling/supaorb/index4.html), and they pretty much confirm my results (e.g. the Super Orb performing on par with the Alpha PAL6035/YS Tech, the FOP38 outperforming these...). So I really don't see your point. In HardOCP's test, the Agilent ArctiCooler performed a bit better than the Alpha, which I could not confirm. Maybe HardOCP was using a CPU with a higher power dissipation, where little differences become more obvious. But I must admit that I tested the ArctiCooler only once, all other coolers were tested three times to ensure the results are accurate and reproducible. After I broke my first KT7 with the ArctiCooler, my motivation to test it on the replacement board wasn't very high. Therefore I didn't include the ArctiCooler in the table with the temperature bar graphs.

Duh! That sure was a long posting, but I hope it clears some of the points

bye,
Tillmann

 

[Mikewarrior2]

Actually, that clears up very, very, very few points.

You must first understand that you aren't measuring cpu core temp on a kt7 with the onboard thermistor. With flip-chip construction, the core has been rotated to the top(heatsink) half of the cpu pcb. What the thermistor on the kt7 is touching is IDLE substrate. It is a secondary pathway compared to the cpu core, and therefore does not show reflect the full amount of change that the cpu CORE would exhibit between these various heatsinks.

The bios does not do a good job "calculating the temperature". It does a good job of approximating the temperature. The vast evidence for this is the mere fact that the change from idle to full load on kt7 and other socket A platforms is often less than 5C. When reading internal diodes(internal diode inaccuracies aside, it does show change effective), the change is 8-12c on a much cooler running cpu. This is also evidence of the fact that it is measuring a vastly inaccurate and unpredictable secondary heat pathway.

The hardocp review does the same thing wrong as your review. It is measuring from a secondary heat pathway that does not reflect the full level fo temp change of the cpu core. Your measurements aren't wrong, neither is the thermistor, but that doesn't mean that the on-board thermistor is accurately portraying cpu temp change.

It is the fact that it is a secondary pathway that makes it unacceptable to use as a comparison platform. Search hte BBS for JohnCar, he knows much more about this. It is accurate for where it is measuring(idle backside substrate). But it is not accurate for comparing cpu core temp changes between the various heatsinks.

Your relative temperature differences between the different heatsinks does not necessarily reflect the "relative" changes in cpu core temp

Mike

 

[Mikewarrior2]

also, might I add that no, someone else with a kt7 probably wouldn't be able to replicate your temps.

the reasons for this:

He may be using an older bios

He may not be bending the thermistor up

He may not be putting grease on the thermistor

He may have a higher ambient case temp. With thermistor readings, ambient air temp plays a larger factor in temp measurement since the thermistor is not isolated from the ambient air.

Also, What was the ambient case temp? I can believe 19c as ambient air room temp, but how about the case? THose components run pretty damn warm

Mike

 

[Howard]

Why not use the Raytek MiniTemp Non-Contact Thermometer? It's the handheld one that tells the temp of whatever it's pointing within an accuracy of +/- 2%.

Here

 

[Tillmann]

Hi,

I am aware about the characteristics of a Flip Chip CPU. Still, I have compared the temperature readings of the KT7's BIOS to temperature measurements taken from a Thermistor in direct contact to the CPU die, and I can only repeat that the relative temperature differences measured are the same.

Of course it's true that not everybody who has a KT7 will get exactly the same temperature readings, but everybody who owns one can easily reproduce them, by flashing the UL BIOS, bending up the thermistor, and applying good thermal compound. If I had used an extra thermistor, it would be necessary to obtain exactly the same thermistor in order to reproduce the results, and that's hard.

About difference between idle and full load, when measured with the onboard sensor vs the thermal diode - I really don't see how you have measured that, no AMD CPU has an internal diode for temperature measurement.

Intel CPUs do. And yes, with Intel CPUs the temperature difference between idle and load is much bigger than with AMD CPUs, due to a more effective implementation of the suspend-on-hlt feature. But I don't see what this has to do with the accuracy of the KT7's thermistor.

The ceramic/metal part between the CPU die and the thermal sensor does have a clearly defined thermal resistance, I don't see what's "unpredictable" about it.

Motherboard temperature, as specified in the roundup, is 31°C. This is the reading from the BIOS. I also measured with a standard thermometer about two inches above the memory slots, and there the temperature was 29°C.

bye,
Tillmann

 

[ModemMix]

they should have definatly made use of a more accurate temprature measuring tool, i am happy that my purchass of a g-dub fop 38 wasnt in vain its an excelent preformer if not a bit loud, but im half deaf anyways so its ok.

ModemMix

 

[Tillmann]

Hi,

@Yucky:
one of the manufacturers of these infrared thermometers (I don't remember if it was RayTek or another mfr) has contacted me about these IR thermometers a while ago, and I discussed with them whether these would be suitable for measuring CPU temperature. We came to the conclusion that they aren't, since the part of which the temperature should be measured must be "visible". Since the CPU is covered by the heatsink, its temperature can't be measured by an infrared thermometer.

bye,
Tillmann

 

[Mikewarrior2]

Hmmm... Okay, now i feel bad about destroying your review.

You say your ambient CASE TEMP is 29C.

Your best heatsink the Kanie Hedgehog, is running at 41C Full Load.

at 12C difference, and roughly 55W, the kanie hedgehog would have to be .22C/w in order for that to be even remotely possible. How can you possibly claim that your measurements are accurate? Your temps are at least 7C off, and that is if you had a perfect connection between the kanie hedgehog and cpu(hence not requiring grease).

You are measuring a secondary pathway. It does not matter what algorithm's the bios uses to calculate the temp, because the temp comes from a secondary pathyway that is not isolated, does not solely contact cpu, and contacts idle cpu substrate. Changes are not measured in full by this method.


Mike

 

[Tillmann]

Hi,

I have another comment for DaddyG:

We do have the test equipment, but we use it only where it really makes sense. For Socket CPU coolers, there's no real alternative to using a real CPU and motherboard for testing. It is true that with a CPU "simulator" more accurate measurements are possible, but it is very hard (IMHO nearly impossible) to accurately simulate the contact area (size, location, material) / clip system for a Socket A CPU.

For Slot A CPU coolers, things were different (and more easy). Since the Slot A Athlon had a thermal transfer plate, on which the heatsink was installed, it was very easy to build a "simulator" that perfectly simulates the contact area between CPU and heatsink, and provides a clearly defined thermal load, which allows super-accurate measurements (We built such a simulator many months ago, for the Slot A cooler roundup - see http://www.anandtech.com/showdoc.html?i=1136 for info on the testing methodology).

But for Socket A CPUs, this isn't possible. So I don't see what "equipment" we could use for testing. Ok, we could use external sensors to measure the temperature, but this doesn't have any major advantages, as I explained in the other replies.

But, in future Socket A cooler reviews, I'll include two measurements: one obtained with the onboard thermistor, and one obtained with an extra sensor that is installed in direct contact with the CPU. Not because I think this is strictly necessary (I still don't think it is), but to keep you guys from complaining , and to show that I take reader suggestions very serious. Ok?

Concering the interface materials: It is true that phase change thermal pads are superiour to classic (e.g. graphite) pads. But good thermal compound is still better, and yes, it does fill imperfections on the heatsink base very well.

bye,
Tillmann

 

[Tillmann]

Hi,

@Mike:
The power specification of 55W is only reached when "worst-case" instructions are being executed. "Full load" (e.g. when running seti@home) doesn't mean the CPU is dissipating the maximum amount of power. I used seti@home because it's widely available to anybody. There are programs that will get the CPU to run hotter then seti@home - e.g. during DivX encoding it'll run hotter (I've been told from various people, haven't measured myself). But since the DivX codec is pirated software, I didn't use it of course. If you don't care about that, give it a try and measure temperatures yourself.

Also, you're talking about the absolute temperatures again. I never claimed that these were perfectly exact (in fact I even posted earlier that there are little differences between the absolute temperatures measured at the core and from the onboard sensor), I just claimed this for relative temperatures, which allow direct comparison of heatsinks.

But as said above, in future reviews I'll also post temperature readings obtained with a sensor that is in direct contact to the CPU, ok?

bye,
Tillmann

 

[Maverick2002]

HAHA mike i KNEW you were gonna post something about this LOL
yea i agree that after our (intelligent) conversation the review seemed a bit off target, but you gotta give em props for stripping the thermosistor and using grease......
yea the whole chrome orb thing is bad as balls.......but overall i think it wasnt TOO bad, although i would like to have seen something like what 2cooltek did with his tests

 

[DaddyG]

Tillman,

Look around the net, tests have been done in a very precise thermal chamber. You are NOT controlling the heat output of the CPU. Full load tests are subjective. As I posted, just how many Floating Point ops does it take for the cpu to consume 1 watt.

When you remove the pre-applied thermal compound you completely invalidate the manufacturers specs. A simple mis-application of the compound can produce quite dramatic results.

Case in point, Taisol uses the 09 in their part # to indicate that Chomerics T-725 PCTC is pre-applied in a heat fused format. Part #'s without the 9 indicate no thermal compound.