Defeating thermal paste pump-out in high power GPUs

[yoadknux]

Hello friends,

I'm wondering if anyone has experience with dealing with pump-out of the thermal paste in high power GPUs.
I've owned a TUF 4090 for about two months, of course I had to push it to 600W and stress test it, and I've observed high GPU hotspot temperatures ~95c, although core was reasonable ~75c. After changing thermal paste (Noctua NH-T2), I got a great improvement: Hotspot 75c, Core 65c. The problem is that two weeks into the application, the temperatures steadliy trend upward: Hotspot ~84c, Core 68c. Those are still great temperatures, but I'm afraid temps will continue increasing, and on top of that the ambient will also increase with the summer coming up. Meaning, if left unnoticed, I could see it hitting 100C+ in August.

I think with cards like the 4090 the thermal paste pumpout is more severe due to two reasons:
A) 600W load <=> 0W idle mean the heat cycles are very severe
B) The heatsink itself is much heavier (my TUF weights 2.2kg!), meaning it gradually distorts the die-cooler contact point due to gravity.

I guess the direction I'll go with next is switch from the NH-T2 to a more dense paste like MasterGel Maker or Gelid GC Extreme. But I'm wondering if there's a way to help the card maintain its mechanical stability? I use a GPU support bracket, but I'm not sure it really helps the heatsink, since the contact point is the GPU shroud. Anyone have any experience with those things? Comments apperciated!

 

[freeagent]

My experience with Noctua paste was not very good, it pumped out on way older cards. You want a nice thick TIM. My newer GPUs have not been opened up yet as there is no need. Maybe look at things like case flow? I know my 4070Ti enjoys the environment that I have placed it in.

 

[EarthDog]

A) 600W load <=> 0W idle mean the heat cycles are very severe

So, what 4090 uses 600W consistently? They are 450W cards. 33% power increase is a ton......

But I'm wondering if there's a way to help the card maintain its mechanical stability? I use a GPU support bracket, but I'm not sure it really helps the heatsink, since the contact point is the GPU shroud. Anyone have any experience with those things? Comments apperciated!

I don't think there is a need at all. AFAIK, the GPU shroud is attached to the heatsink anyway.... so supporting that supports the heatsink and card.

Anyway, I also don't think it's the card or a 'heavy heatsink' so much as it is the TIM used. You'd want a more viscous paste, indeed. I replaced the stock TIM on mine (4090 Suprim Liquid X) with some thermo grizzly and it's been consistent over time.

 

[mackerel]

That might be why the stock pastes tend to be solid when cold. Haven't used the newer Noctua but the older one seemed very runny to me. Think Cryonaut is on the thicker end but maybe even look for something even thicker. Probably matters more how thick it is when hot and I can't say I tested that myself.

 

[67Elco]

PTM 7950

 

[Evilsizer]

not sure if it is still being sold, but i go by PK3. last time i got some PPC's had 3 tubes left.

 

[yoadknux]

When I opened the card today I saw a clear circle, with no paste, in the center. Yikes...
I re-pasted with Thermal Grizzly Kryonaut Extreme (pink, lol). Judging by the paste application, its definitely thicker, we'll see how it holds up. With the fresh application, running the same benchmark I get -15c hotspot temp and -4c core temp, but I want to emphasize that it's the fresh application that makes the difference and not the paste itself (temps are about the same between fresh Noctua and TM paste).
I also placed an order for a CM MasterGel Maker as a backup in case this doesn't work out. As for PTM7950, I heard about it, I'll go with that if I see that the paste approach completely fails.

 

[67Elco]

The paste won't fail...it's just the PTM 7950 is a longer lasting alternative.

 

[EarthDog]

I'm still a bit hung up on a 4090 that pulls 600w in any consistent fashion... lol.... and now "no paste" in the center. The gao paste is filling is well less than 1mm.. perfect applications tend to look the same. Be sure you're not confusing just right with too little.

In other words... for 600w and the paste...




Unrelated, please tell me you aren't using furmark...

 

[yoadknux]

I'm still a bit hung up on a 4090 that pulls 600w in any consistent fashion... lol.... and now "no paste" in the center. The gao paste is filling is well less than 1mm.. perfect applications tend to look the same. Be sure you're not confusing just right with too little.

In other words... for 600w and the paste...




Unrelated, please tell me you aren't using furmark...

I agree that only with FurMark you can see consistent 600W load (Yes, I tried it). But occasional 600W spikes are real even with 450W power limit. Just open HWInfo and look at GPU Rail Power. Here's what I get after running TimeSpy Extreme Benchmark:

As for the center gap, I definitely should have taken a picture. I'll try again next time (I hope not)

 

[EarthDog]

Ahh, that reading. Yeah, an "instantaneous peak". Its also included in the gpu power draw in hwinfo already (the one thats reporting a lot closer to the power limit of the card). It's not frequent, nor consistent and not something to base things on in this case. You couldn't get that consistent in furmark as your cards power limit is somewhere around 450w at stock.

If you want to keep it simple, gpuz and the total board power is what you're looking for. I wouldn't account for those normal blips. They happened forever, now you can just see it in software.

 

[yoadknux]

Ahh, that reading. Yeah, an "instantaneous peak". Its also included in the gpu power draw in hwinfo already (the one thats reporting a lot closer to the power limit of the card). It's not frequent, nor consistent and not something to base things on in this case. You couldn't get that consistent in furmark as your cards power limit is somewhere around 450w at stock.

If you want to keep it simple, gpuz and the total board power is what you're looking for. I wouldn't account for those normal blips. They happened forever, now you can just see it in software.

Yeah, I agree with you.
Anyway I'll leave those readings here so I know what to expect when I repeat the measurement by the end of April. The lower temps are with stock PL (450W), the higher temps are with 600W, no OC, both under TimeSpy Extreme Benchmark.

 

[EarthDog]

So, you're raising the power limit to reach 500W (again, thats a peak/blip - what was the average through the run?)? Did you flash the bios to get that much more out of it?

I'll run some tonight on mine and see how it behaves. I can tell you though, IIRC. I only get a 6% increase on mine.

 

[yoadknux]

So, you're raising the power limit to reach 500W (again, thats a peak/blip - what was the average through the run?)? Did you flash the bios to get that much more out of it?

I'll run some tonight on mine and see how it behaves. I can tell you though, IIRC. I only get a 6% increase on mine.

No, I didn't flash any BIOS. By using MSI AB with default bios, I have access to 133% power limit, which is 600W maximum.


I agree with you that in principle you have to work hard to find an application that uses 600W consistently other than FurMark. The TimeSpy Extreme only manages to pull a maximum of 500W not including the spikes. But I'm guessing this number can increase if I overclock the card (this is at stock clocks).

 

[Woomack]

These cards are not designed to run at more than TDP for an extended time. The 600W is the maximum power design, but more for short-time spikes. The card shouldn't go even close as you are asking for problems.
I also don't think you understand the main purpose of TIM. It is supposed to fill the gaps between the die/IHS and the cooler, and that's all. In short, you don't use too much of it, as every higher-pressure cooler will leave barely any TIM on the die/IHS. This is normal, and if there is not much TIM after the cooler installation, it means it perfectly fits and has an optimal design.
I have no idea how it was tested that ~20°C difference after reapplying the TIM, but it's impossible unless the cooler wasn't properly mounted or the TIM was too thick and too dry. The difference between TIMs is typically up to 5-6°C (more like 1-3°C, as even that factory-applied isn't so bad nowadays). Also, Coolermaster TIM is pretty bad. I was testing it, and I won't ever use it. No matter what you use, you won't get more than ~1°C better temps than on Noctua NT-H2. It shouldn't degrade for about 2-3 years. Thermal Grizzly is typically more liquid. The only better option would be Alphacool Apex, but again, I haven't seen more than ~1°C difference.
Btw. even if you unlock the power limit to 600W, it doesn't mean that the card will actually go that high. It's because the voltage and other limits are locked. This is also why, except for spikes, you see ~500W.

 

[yoadknux]

These cards are not designed to run at more than TDP for an extended time. The 600W is the maximum power design, but more for short-time spikes. The card shouldn't go even close as you are asking for problems.
I also don't think you understand the main purpose of TIM. It is supposed to fill the gaps between the die/IHS and the cooler, and that's all. In short, you don't use too much of it, as every higher-pressure cooler will leave barely any TIM on the die/IHS. This is normal, and if there is not much TIM after the cooler installation, it means it perfectly fits and has an optimal design.
I have no idea how it was tested that ~20°C difference after reapplying the TIM, but it's impossible unless the cooler wasn't properly mounted or the TIM was too thick and too dry. The difference between TIMs is typically up to 5-6°C (more like 1-3°C, as even that factory-applied isn't so bad nowadays). Also, Coolermaster TIM is pretty bad. I was testing it, and I won't ever use it. No matter what you use, you won't get more than ~1°C better temps than on Noctua NT-H2. It shouldn't degrade for about 2-3 years. Thermal Grizzly is typically more liquid. The only better option would be Alphacool Apex, but again, I haven't seen more than ~1°C difference.
Btw. even if you unlock the power limit to 600W, it doesn't mean that the card will actually go that high. It's because the voltage and other limits are locked. This is also why, except for spikes, you see ~500W.

I'm well aware of what thermal paste does and how its supposed to be applied. I've seen with my own eyes a gradual degredation of the hotspot temperature, within weeks/months, even with the NT-H2, and I'm sure I can replicate it.
My thermal paste application is thin layer w/ spatula.
In the picture below you can see the same identical test (FireStrike Ultra 20 loops) with identical test conditions, before and after re-applying paste.


With the Noctua->Thermal Grizzly, I didn't document it, but I saw TimeSpyExtreme temperature drop from 90c to 76c. There was a clear improvmenet. And like I said, I think it's the fresh application that made the difference and not the paste itself.

It's not just me. Other people are reporting those things as well:

https://www.reddit.com/r/overclocking/comments/15yia3i/4090_strix_thermal_paste_replacement

4090 Strix Thermal paste, dint and hotspot

so I was getting 70+c core temps with hotspot up to 100c so decided to tear this down and repaste it and what I was greeted with I can only say I find it totally unacceptable!!! Especially for such a expensive high end card, this particular bright spot appears to have been a hard bit of paste...

rog-forum.asus.com

And guys, please don't fixate on the 600W number. Fine, make it 500W, or 450W. That's still a lot of heat that's extracted from the core to the heatsink through the thermal paste. I know that in past years you could apply paste to a 100W CPU and it would last for years, maybe even for the entire lifetime of the CPU, but I'm starting to be doubtful whether this is true for high power GPUs

 

[EarthDog]

but I'm starting to be doubtful whether this is true for high power GPUs

Does wattage matter to the thermal paste or the heatsink? 90C is 90C to the paste as I understand things... be it from 450W or 5W.

Now, the 'effort' of the heatsink is higher to keep more wattage at 90C, but not sure what that does to thermal paste (my guess is nothing outside of temps get warmer and it becomes less viscous). Other cards with less wattage can easily hit 90C too.

Did you notice in your link the first (only) reply in the thread said it happened on his 4070ti (mentioned 4000 series from asus, specifically) which is considerably less wattage (~285W).

Seems like this may be an issue for Asus on their 4000 series cards. Note that my reapplication wasn't because of rising temps, but because I took the card apart for the review, and it needed it. The thermal paste on the suprim was nearly solid, lol, and perhaps why we don't hear a flood of similar stories from their cards or any other brand.


EDIT: Again, I'm not denying it's happening or that reapplication doesn't temporarily mitigate it, but my money isn't on the wattage but on Asus' thermal paste and how it behaves at temperature. I'd go as far as saying my TG Kryonaut is less viscous than the stock MSI TIM and still keeps the GPU temps fine.

EDIT2: In thinking a bit more critically, My 4090 is water-cooled, doesn't peak anywhere close to 90C, and this doesn't happen with the same wattage (450+).

 

[Evilsizer]

if your worried about thermal paste degrading over time with those high of temps. just get some thermal putty and be done with it. i posted a few of this persons videos in the past. seem they are doing more testing lately, i have not watched any of the newer ones yet. i do suggest watching his videos on testing. https://www.youtube.com/@snarksdomain3453 their channel is also how i choose my thermal pads for the 2060 super that needed to be refreshed.

 

[yoadknux]

Yeah, I agree with you.
Anyway I'll leave those readings here so I know what to expect when I repeat the measurement by the end of April. The lower temps are with stock PL (450W), the higher temps are with 600W, no OC, both under TimeSpy Extreme Benchmark.



View attachment 365568

I'm quoting this post so you can see the differences in just 1 week.

Same benchmark, same settings, same peak power consumption. GPU temp +3.2c, Hotspot temp +9c, Memory Temp -2c, Hotspot-CORE Delta 9.3->15.1c.
Core temp +3.2c is no big deal. Hotspot delta increasing is what bothers me. And this is very consistent with memory temp dropping, because they use the same heatsink: the card tanks the heat instead of transferring it to the heatsink, therefore more cooling for the memory.
I think there are three ways forward
1) say "**** it, its an invisible temperature reading that doesn't affect performance" and just monitor every now and then that it doesn't go above 100c
2) Try to resolve possible mounting pressure issue by changing to softer thermal pads or thermal putty
3) Try PTM7950 instead of thermal paste

I'm leaning towards option 3

 

[yoadknux]

Last night I got 90c hotspot again with the benchmark (+14c change compared to 24/03 repaste - compare it to the 600W picture).

When opening the card, I saw the exact same thing that I saw with the Noctua paste: Most paste out of die, strange ring in the center.







I tried one final thermal paste, the Cooler Master MasterGel Maker, which is the thickest paste of all 3. Fresh temps are nearly identical to the Kryonaut Extreme from 24/03.



PTM7950 expected to arrive around the end of the month, and I intend to use it in case the CM paste also degrades.