RTX 5090 FE Liquid Metal Swap: Thermal Paste Performs Just Fine

[btarunr]

Did you catch our launch review of the NVIDIA GeForce RTX 5090 Founders Edition graphics card? If you did, you'd have noticed the hoops we had to jump through to disassemble the card without breaking anything. As we carefully peeled the card's aesthetic finery and worked out way down to the GPU, we found that NVIDIA is using a liquid-metal thermal interface material (TIM) between the "GB202" GPU and the unified vapor chamber plate of the Double Flow Through cooling solution. After our photography the problem arose of putting the card back together, and we wondered if using a mainstream thermal paste would be a suitable replacement since we didn't have the patience to toy with liquid metal given our review volume. So we pulled out our trusty tube of Arctic MX6, gave it a suitable application, skipped the gasket, and put the card back together.

Here's what we found—the GPU temperatures rise by about 2°C on average over the stock liquid metal TIM. In the graph below, you will see both TIM applications compete with each other over a 370-second stress from a benchmark run. In the chart the start temperatures are slightly different, this does not mean that idle temperatures are higher with liquid metal. At the end of test the maximum temperature reached with the stock liquid metal TIM is around 77.6°C. The Arctic MX6, on the other hand settles to 79.4°C. This +1.8°C temperature increase really isn't significant at all—room temperature changes between summer and winter will cause bigger swings. We also tested performance, and it was spot on, the same as pre-disassembly—not a hint of thermal throttling. Both values are safely below the 90°C thermal threshold for the RTX 5090—that's right, NVIDIA raised the thermal throttle point, it's not 83°C anymore like on the RTX 40-series Ada. Taking the RTX 5090 apart and putting it back together was a challenging experience, but we're glad we didn't have to do a liquid metal application to ensure trouble-free operation. This is good news for all DIYers—don't bother with replacing the liquid metal—a thin thermal paste application works fine, too.



View at TechPowerUp Main Site

 

[usiname]

The conclusion:
Nvidia is using the cheapest $hitty available liquid metal

 

[Vya Domus]

The hotspot temperatures which are not exposed are probably a lot worse.

 

[Lewzke]

5090 FE using stock liquid metal interface? Why not the industry standard PTM? That 1C is not much a difference and PTM is safe and stable in a long run.

 

[FiRe]

"just fine", then proceeds to show ~2 degree temperature difference. In that case I might as well use cotton wool just as well if 2.3% is classed as "fine"

 

[W1zzard]

"just fine", then proceeds to show ~2 degree temperature difference. In that case I might as well use cotton wool just as well if 2.3% is classed as "fine"

What's your room temp?

 

[TheDeeGee]

Then i'd replace it with a Thermal Grizzly Kryosheet and forget about it.

5090 FE using stock liquid metal interface? Why not the industry standard PTM? That 1C is not much a difference and PTM is safe and stable in a long run.

PTM is still paste, and will eventually degrade.

 

[lepudruk]

The hotspot temperatures which are not exposed are probably a lot worse.

And what about VRAM? I saw temps reaching around 90+ C degrees on some tests.

 

[JustBenching]

The conclusion:
Nvidia is using the cheapest $hitty available liquid metal

Or, the bottleneck is the heat sink and not the die. Liquid metal is useless by itself if it's the heat sink that is getting overwhelmed.

 

[usiname]

Or, the bottleneck is the heat sink and not the die. Liquid metal is useless by itself if it's the heat sink that is getting overwhelmed.

For sure the cooler is bottleneck, but better heat transfer will improve the temperatures with more than 2 degrees, especially at ~80C.
The hotspot is the main issue here, probably before the repaste it was around 100C, now maybe over 110C

 

[CarvedInside]

But what about the fan noise? Where the fans spinning at similar RPM?

 

[W1zzard]

The hotspot temperatures which are not exposed are probably a lot worse.

The hotspot is the main issue here, probably before the repaste it was around 100C, now maybe over 110C

Probably .. maybe .. maybe not? cmon people ..

You can manually test for hotspot, because it will still cause thermal throttle.

I stopped the fans with my fingers during Furmark and the card didn't throttle until GPU temp showed 90°C.

If hotspot "it's over a million" like you two claim, then this shouldn't be possible

But what about the fan noise? Where the fans spinning at similar RPM?

Yup, I checked that too of course

 

[remekra]

It may run fine now, but the question is about longevity. There are many cases on the forum or internet about 7900XTX hotspot temps rising after months of usage. I also had that problem with my Gigabyte Gaming OC model. After 6-7 months temps on hotspot rose to 110 and fans were running at 100%. RMAed it and switched to Nitro+ which supposedly uses PTM and didn't had an issue for all the time I used it until I sold it a week ago.

Also edge temps stayed the same the whole time with cool 65-70 degrees, just hotspot rose. If Nvidia doesn't expose that then it will be hard to say what is going on. If the fan speed is related to hotspot as in Radeons rather then edge then fans will go crazy, but if they are connected to edge temp then GPU can either cook itself or just shutdown.

EDIT

Yeah thermal throttle will probably happen before it will cook itself. But still would be nice to see how it performs after couple of months

 

[RH92]

"just fine", then proceeds to show ~2 degree temperature difference. In that case I might as well use cotton wool just as well if 2.3% is classed as "fine"

usually there is about 20C difference between LM and paste so yes 2C difference especially on such high TGP is ... JUST FINE ...

 

[Denver]

Probably .. maybe .. maybe not? cmon people ..

You can manually test for hotspot, because it will still cause thermal throttle.

I stopped the fans with my fingers during Furmark and the card didn't throttle until GPU temp showed 90°C.

If hotspot "it's over a million" like you two claim, then this shouldn't be possible


Yup, I checked that too of course

Is there any rational reason to defend the absence of the sensor? It's an expensive product, I think it's fair to be critical about it.

If the temperature in the engineering marvel is high, imagine in more modest designs. Why hide data?

 

[RH92]

What's your room temp?

He probably lives on Venus . Jokes aside i think it would be nice to test with some well established LM like Conductonaut for good measure. Surprising to see such a small gap so either that vapor chamber design is crazy efficient or maybe stock LM is not of the best quality .

 

[TSiAhmat]

Is there any rational reason to defend the absence of the sensor? It's an expensive product, I think it's fair to be critical about it.

If the temperature in the engineering marvel is high, imagine in more modest designs. Why hide data?

Because people would probably freak out if they saw a high number, even if it is safe. And I assume nvidia tested the gpu intensely so it won't break.
(we already saw that with amp cpu and people complaining that they get to hot/the number is to high)

 

[Vya Domus]

You can manually test for hotspot, because it will still cause thermal throttle.

I stopped the fans with my fingers during Furmark and the card didn't throttle until GPU temp showed 90°C.

If hotspot "it's over a million" like you two claim, then this shouldn't be possible

Is Furmark even a good test for this ? I suspect the driver automatically throttles the card to some extent as soon as it's detected because it's a known power virus, at what frequencies does it run when you have it running ? Because for instance on my 7900XT the clocks don't even go over 2Ghz and it barely touches the power limit, I have come across far better benchmarks for thermals like the AIDA64 GPGPU test where my temperatures shoot up to 80C+ hotspot even though I am on water and using PTM, temperatures are far greater vs Furmark. I wish reviewers would stop using it because it is just not a good test anymore.

 

[Legacy-ZA]

The hotspot temperatures which are not exposed are probably a lot worse.

Yes, I hate that we can no longer see this, it helped so much to notice quickly if you re-pasted wrong, have a bad alignment or had a defective chip from the start. I don't like this one tiny bit.

This is a fun little experiment, I never really was sold on liquid metal, I will stick with my MX-4

 

[_roman_]

I wonder what happens when someone uses a proper thermal paste. MAybe amazon sold me in the past a bad batch of arctic mx-6. That arctic paste from amazon performed very much poorly than my existing noctua one.

That Nvidia graphic card will make a great test field for future thermal interface medium testing.

 

[blufeb95]

Could you test what it's like using PTM7950 or Thermalright helios which would be closer to what they'd be using if they weren't using liquid metal.

 

[AVATARAT]

"Just fine" but for how long?
A month or two?
Hot Spot will bake the paste and without a sensor, the user won't mention it, so - RMA = no warranty.

 

[Ruru]

The conclusion:
Nvidia is using the cheapest $hitty available liquid metal

Are there really differences between different liquid metals?

 

[W1zzard]

Is Furmark even a good test for this ? I suspect the driver automatically throttles the card to some extent as soon as it's detected because it's a known power virus, at what frequencies does it run when you have it running ?

On NVIDIA it is throttled just enough to sit right at the power limit at all times. The temp chart in the news post is using a game-like load of course. But when I want to just heat up a card real fast real hard I'll always use Furmark

 

[Bwaze]

He probably lives on Venus . Jokes aside i think it would be nice to test with some well established LM like Conductonaut for good measure. Surprising to see such a small gap so either that vapor chamber design is crazy efficient or maybe stock LM is not of the best quality .

Maybe the huge area also plays a role - there is relatively small advantage in using LM compared to good paste between CPU heat spreader and cooler, but a large difference in using LM to interface small delidded CPU die to heat spreader.

The (NVIDIA GeForce RTX 5090) GB202 graphics processor is a large chip with a die area of 750 mm²

AMD 9800X3D die size: 70.6 mm²

Intel Core Ultra 9 285K die size: 243 mm²