The Official Thermal Interface Material thread

[Borc]

I wish Igor would put these all on the same graph rather than having separate graphs for pastes and phase change materials. PTM7950's results are so good it seems anomalous. Thermal conductivity is one thing while real world performance is different. 75% of pastes in Tom's Hardware's mega test perform well. Helios didn't perform as well as I expected in Igor's test but Igor is testing what's possible not real world performance.

As long as you aren't trying to cool your PC with Kroger Gun Protection toothpaste you are fine. As enthusiasts we want the best cooling possible. We spend extra money to get 1% better results. Meanwhile the difference between most pastes and phase change materials is within margin of error. In my own testing the Thermal Grizzly kryosheet gave my GPU 2.7% more performance than MX-4. Was that money I spent on the Kryosheet worthwhile for a mere 2.7%? Yes. I would do it again. It was fun for me to test.

The gaps are much bigger on a small laptop CPU or bare die chip in general. Some pastes in this THG review are so much worse on my laptop CPU. For some reason most people think thermal paste is just for desktop CPUs with heatspreader and think the gaps are always so small between pastes on every chip.

 

[Nordic]

I finally applied the kryosheet to my CPU. I have minor temperature gains compared to a year old application of Kryonaught Extreme. Not really enough of a difference to matter.

 

[freeagent]

TFX is so thick I do not understand how it pumps out

So far so good with my application of Heilos to my GPU.

SYY-157 on my CPU and things are just dandy I suppose.

Pump out.. really?? I don't even know what to say. Its 2024 lol.

 

[Nordic]

Strange how no major reviewer has tested Thermalright TFX, while TF7 and TF8 seemed common. It appears to have mostly fallen off the radar, despite claimed superiority over the TF8.

I've been happy with it for repasting laptop and NUC and a desktop build, with no issues except for the horrifying thickness. Wonder what they put into it.

Cant find anything about TFX apart from multiple complaints about pump out. I think PTM pads are honestly the way forward.

Igor's just did a test showing TFX had thermal conductivity similar to PTM7950.

 

[FreedomEclipse]

So far so good with my application of Heilos to my GPU.

I was considering pasting up my 7800XT with heilos before putting it in. After seeing igors review I wasnt too sure about it. How are your before and after temps?

 

[Shrek]

Here is something I never quite understood; look at the temperature drop for a typical power, area, thickness and thermal conductivity

P = 100W
A = 3 cm^2
k = 20 W/(m.K)
Dx = 1/20 mm

How then can thermal pastes differ by say 2 to 3°C degrees?

 

[Borc]

TFX is so thick I do not understand how it pumps out

Too much filler maybe? Here is a good explanation of what can happen: https://koolingmonster.com/insights...ivity-and-is-it-always-good-for-thermal-paste

There are many complaints about the fast pump out of Thermalright TFX. I also can confirm TFX degrades really fast, on my laptop CPU it did.

Metal alloys that are most often used in thermal paste include: aluminum oxide, boron nitride, zinc oxide, and aluminum nitride

So what difference does this make for thermal paste? Recall that thermal paste is not just these alloys. There is also an oily (most of the time silicone) base to host them. Therefore, the more of the alloy you add to the base, the higher your thermal conductivity and the higher the ability of the thermal paste to transfer heat out to the heatsink.

This is where we crossover from theory into reality. Because most of the time our thermal paste doesn’t simply sit in a vacuum – it needs to be applied correctly and to function for long periods of time.

Achieving a “high” thermal conductivity (in this case, let’s say 10W/mK+) is actually very easy – all you have to do is keep adding filler (metal alloy) until you achieve that level of conductivity. But thermal paste’s job is not solely to transfer heat – it is also a gap filler for the micro-imperfections between the surface of CPU and heatsink where oxygen might be trapped.

If you add too much filler, you risk making the thermal paste too hard to spread, resulting in thick, uneven layers which result in poorer real-life performance. And because of the lack of base, high-filler thermal pastes will also degrade significantly faster when put through real use which will quickly result in poorer performance.

 

[FreedomEclipse]

Here is something I never quite understood; look at the temperature drop for a typical power, area, thickness and thermal conductivity

View attachment 360503
P = 100W
A = 3 cm^2
k = 20 W/(m.K)
Dx = 1/20 mm

How then can thermal pastes differ by say 2 to 3°C degrees?

higher aluminium/zinc oxide content?

 

[freeagent]

I was considering pasting up my 7800XT with heilos before putting it in. After seeing igors review I wasnt too sure about it. How are your before and after temps?

Marginal at best, about 3c better than SYY. Stock was pretty decent, I should have left it.

 

[ShrimpBrime]

Here is something I never quite understood; look at the temperature drop for a typical power, area, thickness and thermal conductivity

View attachment 360503
P = 100W
A = 3 cm^2
k = 20 W/(m.K)
Dx = 1/20 mm

How then can thermal pastes differ by say 2 to 3°C degrees?

Margin of error.

 

[Shrek]

Maybe so, but then most any reasonable paste will do (regardless of thermal conductivity).

I am wondering if surface effects play a role that is not always recognized, namely, the analysis I did is not valid.

 

[ShrimpBrime]

Maybe so, but then most any reasonable paste will do (regardless of thermal conductivity).

I am wondering if surface effects play a role that is not always recognized.

Yes, this could be variations in surface mating and surface area. Still would be minor.

I would think ambient temp sway would be a culprit too. My furnace has a 4° temp swing from on to off for example.

 

[Nordic]

Here is something I never quite understood; look at the temperature drop for a typical power, area, thickness and thermal conductivity

View attachment 360503
P = 100W
A = 3 cm^2
k = 20 W/(m.K)
Dx = 1/20 mm

How then can thermal pastes differ by say 2 to 3°C degrees?

I always assume large differences come down to user error in application of paste. One benefit of these phase change pads and the kryosheet is you remove user error. You can't apply it too thick or unevenly.

 

[Shrek]

There is still ShrimpBrime's 'My furnace has a 4° temp swing from on to off' which indeed would explain the 2 to 3°C degrees, although since his location is 'Nowhere' I'm not sure if he meant F or C

 

[Borc]

I always assume large differences come down to user error in application of paste. One benefit of these phase change pads and the kryosheet is you remove user error. You can't apply it too thick or unevenly.

On a small laptop chip I can say the difference can be huge between a top paste or PTM and mediocre average paste. Runny pastes like MX-4 doesn't really work and would probably even worse for longevity. Kryosheet didn't work either by the way, phase change pad is better for a laptop chip. Most people doesn't really know that, I mean who is really testing it out with 10+ pastes. Mostly it's based from heatspreader chip testing where the margin will be much much smaller.

 

[ShrimpBrime]

There is still ShrimpBrime's 'My furnace has a 4° temp swing from on to off' which indeed would explain the 2 to 3°C degrees, although since his location is 'Nowhere' I'm not sure if he meant F or C

4F would be 2C temp swing. I am in US.

Is he now here, or no where? Some prefer me not here maybe.

 

[freeagent]

We lurves ya

At least I do..

No eye contact makes it ok somehow

 

[lexluthermiester]

TFX is so thick I do not understand how it pumps out

It doesn't.

 

[Athlonite]

TFX is so thick I do not understand how it pumps out

It's probably not pump out but far to much used in the first place

 

[Acfeiln]

Here is something I never quite understood; look at the temperature drop for a typical power, area, thickness and thermal conductivity

View attachment 360503
P = 100W
A = 3 cm^2
k = 20 W/(m.K)
Dx = 1/20 mm

How then can thermal pastes differ by say 2 to 3°C degrees?

I'm not sure I fully parse your question, so apologies in advance for any misunderstanding or if I say too much you already know. But that equation (a model for thermal conduction) is not one for ΔT between CPU and ambient, nor is it one for comparing two TIMs. Instead it tells you about conduction through a single TIM application. Your ΔT is the temperature difference between the CPU-end and cooler-end of the TIM itself. (Are you asking: if that number is so tiny, how can there be so much variation among different ones leading to single-digit CPU temp changes? If so, completely valid, I think I just didn't read you write the first time. Regardless...)

Seperately, how is your Dx so small? 5% of one millimeter? Where is that from? (This will show up later as well.)

In any case, I got curious, so here's an attempt to model the CPU temperature difference between two TIM applications. Lots of simplifications and bla bla bla I don't expect extreme accuracy or anything, but I burned an evening with this in my head so here we are.

Modeling CPU Temp. Difference Between TIMs​

Here's a visual sketch of our two imaginary setups.

I'll call the TIM applicaiton thicknesses by d rather than Δx to save some space. The two cross-sectional areas are denoted by A. Temperatures are all denoted T, and a note about each one:

• T1 is the temperature at the CPU/TIM 1 interface. T2 is the temperature at the CPU/TIM 2 interface. The quantity we're after is T1−T2, the difference in CPU temps given two different TIM applications and all else being equal.
• Tb is the temperature at the TIM/cooler base plate interface. For reasons I'll outline in the assumptions, this will be the same for both TIM applications.
• Tambient is the ambient air temperature. We assume this is the same for both TIM applications.

Addtl. Assumptions I'm Making​

1. The CPUs are identical, operating in steady state, and dissipating power P in both cases. This means the power through each TIM and through each cooler must be P: otherwise the TIM (or cooler) would itself be accumulating or losing internal energy over time, neither of which are stable scenarios.
2. Perfect (read: I will completely ignore) thermal contact conductance between the CPU/TIM and TIM/cooler interfaces. You're welcome to follow the process here and relax this assumption on your own to see how it turns out!
3. I'm ignoring any inhmogeneity in temperature across the surface of the chip, so I'm not considering "hot spots."
4. Identical coolers. This makes a big difference, because it allows me to get away with not modeling the cooler at all. The point is that the cooler thermally connects the TIM contact with the ambient temperature—perhaps with heatpipes and fins and a fan; perhaps through liquid cooling and a radiator and fans; perhaps something more exotic. But as long as they conduct the same power P, and as long as that's linear in ΔT between b and ambient (as it would be for a convective model), that forces the two Tbs to equality, regardless of what other dependencies there are in the cooler's operation (as long as those conditions are equal for both TIM applications). This lets us ignore the cooler.

Working it out​

Start with TIM 1. We have a relationship between the heat conducted and temperature change along the thickness d1 (the heat conduction equation you cited):

,

where k1 is the thermal conductivity of TIM 1. We have the same equation for TIM 2, with the subscripts changed. These are both "true" equations for conduction along each of the TIMs independently. But the fact that they must be conducting the same power and the equality of their Tbs connects them.

Now we'll take the difference:

,

and voilà! I'm leaving this here to highlight the fact that not only the conductances but also the thickness and cross-sectional areas of the applied TIMs will make a difference here. But that's difficult to control and/or measure, so I'll make one further simplificaiton before plugging in numbers: suppose the two TIMs are applied with the exact same thickness over the exact same area. (This is obviously not how it goes irl: different properties of the TIM and variance in user application will lead to different numbers there.) Then we have:

.

We can quickly see this passes a sanity check: if k1 > k2, then T1 < T2, so a higher conductance TIM indeed gives a lower CPU temperature. Plugging in your numbers P = 100 W, A = 9 cm^2 (assuming you meant 3 cm × 3 cm area), d = 0.05 mm (though this seems insanely thin, where'd you get this from?) and comparing, say, Conductonaut at 73 W/m K vs TL-A40 at 15 W/m K (both from here) we get a temperature difference of about 0.3 K. But I'll draw attention again to that d: the temp. difference is linear in that, so adjusting to half a millimeter of thickness brings us to ~3 K temperature difference. We can also see this scales linearly with conducted power: so the difference will get bigger with more power draw. It does not scale linearly with the difference in conductances, though, which is why liquid metal isn't like an order of magnitude better than the competition.

 

[KLiKzg]

Here is something I never quite understood; look at the temperature drop for a typical power, area, thickness and thermal conductivity

View attachment 360503
P = 100W
A = 3 cm^2
k = 20 W/(m.K)
Dx = 1/20 mm

How then can thermal pastes differ by say 2 to 3°C degrees?

They do not...as there is only the application impurities, that are in question...that is why I said: all these tests with CPU are not OK.
Example: if I do not like one company, I am could sabotage the spread & get worse results.

Now, if you look at Igor's lab (which is probably ar some university), then you can see what professional measurement is all about:

• always same equipment
• always the same pressure between source & radiator unit
• 3 sets of data measurement in order for a quite the good correction

 

[Ferrum Master]

tests with CPU are not OK.
Example: if I do not like one company, I am could sabotage the spread & get worse results.

• always same equipment
• always the same pressure between source & radiator unit
• 3 sets of data measurement in order for a quite the good correction

I agree, also that most of these tests are bullshit same as audio reviews, bullshit presentation without structural data, subjective, prone to failure and moon phase and morning horoscope. I better watch comedians, at least they make me laugh not because of sadness. The best bullshit comes when reviewer makes a complaint that is impossible to solve with the current technologies.

There is one caveat to this test. It only has one energy source, hotspot, basically you measure only one axis, some materials are worse or better transferring energy in different axis, or totally suck at it like graphene.

Also PTM in paste form needs a break in period, paste needs drying also, then it gets usable, people use it the wrong way. I also only use the PTM7958 pads only... but that's Igor... he often manages to mess up also.

 

[Shrek]

Seperately, how is your Dx so small? 5% of one millimeter? Where is that from?

...
d = 0.05 mm (though this seems insanely thin, where'd you get this from?)

Take a ream of photocopy paper

• 500 sheets = 5 cm

so, 0.1 mm a sheet, and that paper is thick; so, 1/20 mm may actually be too large an estimate.

so adjusting to half a millimeter of thickness brings us to ~3 K temperature difference.

Half a millimeter seems unreasonably large, even 1/20 mm may be too large.

I burned an evening with this in my head

The start equation is all that is needed, no need to complicate it.

 

[ShrimpBrime]

They do not...as there is only the application impurities, that are in question...that is why I said: all these tests with CPU are not OK.
Example: if I do not like one company, I am could sabotage the spread & get worse results.

Now, if you look at Igor's lab (which is probably ar some university), then you can see what professional measurement is all about:

• always same equipment
• always the same pressure between source & radiator unit
• 3 sets of data measurement in order for a quite the good correction

But no processor. The end resulted data is great for the measurement of those two blocks of copper. Neither is a common cpu cooler, and neither is a cpu. If it where a cpu, it would need to be nickel plated copper.

Is this testing actually accurate for the intended uses? At least what we do have is constant pressure. So that's a good thing.

So this thermal paste does X wm/k an hour between those 2 specific pieces of copper. That's what we get from that lab testing.

The question is, would this study be close enough to realistic usage in a PC??

 

[Acfeiln]

Take a ream of photocopy paper

• 500 sheets = 5 cm

so, 0.1 mm a sheet, and that paper is thick; so, 1/20 mm may actually be too large an estimate.



Half a millimeter seems unreasonably large, even 1/20 mm may be too large.



The start equation is all that is needed, no need to complicate it.
View attachment 360568

Yeah I definitely was an order of magnitude off in my head as to what's reasonable, that was my bad. But I'd be surprised if we expect much thinner than say 0.1 mm consistently in an application (there will presumably be thinner and thicker sections), because that amounts to a full half-size compression of something like a Kryosheet, and if it's reasonable to expect more than that consistently I'd imagine they'd make it thinner to start with. But I don't know anything really about these materials so lots of room for me to be wrong on that!

That said, I'll push back on "start equation is all that is needed" with an asterisk: it's true that's all you need to estimate a temperature drop through a particular TIM from CPU-end to cooler-end (which may well have been what you meant from the start). But if the goal is explicitly to compare two TIMs you definitely need both and a way to combine them, even if it turns out the numerics matched pretty well, and my burned evening was plenty of fun (I modeled the cooler explicitly and worked all the way through it before realizing I could completely ignore it...fun!)