P870DM3 Liquid Metal Repaste Results

(Maybe wrong forum, if so, please bump this to the root Sager/Clevo forum)

So I've meaning to do a repaste since seeing the incredibly high temps I was getting. For reference, I was hitting 91C on the master GPU and 81C on the slave GPU within about 10-15 mins of max fans on Heaven (4K, max settings, all tesselation sliders turned up to max).

I popped it open and did a heatsink mounting alignment check using some crappy Hydronaut that Newegg sent me by mistake. Not great.

Spoiler

So I added a 0.3mm shim to the master GPU, and that gave me good enough contact on both to attempt a liquid metal repaste. I was about to make the leap, but then I realized that the shim wasn't going to adhere to the CLU without sanding it, then I ran out of CLU. So I had to close it back up with a standard paste. The good news was that I was no longer seeing +10C on the master, the bad news was that I was still hitting 91C on Heaven.

Fast forward to today when a bunch of CLU and Conductonaut that I ordered arrived. I sanded the shim with a fine grit paper and voila, the Conductonaut decided to stick around instead of just beading up. Because I'm absolutely paranoid, I also used the last of my Hydronaut and some far too much Kryonaut to make a gasket around the dies to prevent it from escaping.

Spoiler

Results (idle and 24 minutes of Heaven, both max fans, ~70F ambient):

Spoiler

Overall, quite pleased with the results. Average clocks seem to have held pretty high, which fits with the fact that it never hit the thermal downclock threshold of 91C. Not sure if it ever actually dipped, since this also includes a few times where I tabbed out.

So this is where someone tells me that liquid metal mixes with traditional paste to form a superconducting goop that's gonna fry my cards, right?

 

Actually, I've seen others pack non-conductive paste around fivr to seal out moisture and haven't heard of them mixing. For longer term use, I recommend using liquid electrical tape on local electrical components. But, aside from using that on those components, using a cheap paste to build a barrier doesn't seem like a bad idea...

Sent from my SM-G900P using Tapatalk

 

That's great news, what's the GPU/CPU temps on the same test with auto fans? A fan setting that is a bit quieter and closer to normal running, like when gaming?

It's interesting that the 91c/81c went to 85c/85c, any idea how that might be explained?

The average GPU temp went down 1c from 86c ( (91+81)/2 = 86 ), so the combined heat sink and GPU conductivity improvement balanced out the average temperature, and lowered it by 1C ? That's cool

 

The delta was caused by improper heat sink pressure on the master GPU. You can see this in the first photo where there is significantly more paste than on the slave GPU. Pascal is designed to throttle back down at 91C, so once the master hits that temperature both GPUs get downclocked. When I added a shim onto the master, both GPUs made it up near 91C, but the average clocks were higher since it took longer to throttle. This latest run doesn't seem to have throttled at all since it never hit the threshold temperature.

 

Excellent. Great to see results.

So 0.3mm shims eh? I was curious about the size as I have 0.1mm to 0.5mm locked and loaded to be used if necessary.



::iunlock::

 

Just one shim. The slave GPU had good enough contact with the heatsink to paste directly.

 

why not use electrical tape... that gasket you make is nasty :|

 

Because I would need like 30 layers to make it thick enough to reach from the PCB to the heatsink.

 

If it's applied properly it should not be required but I understand what it feels like when working with such expensive parts

 

Yeah, between the GPUs being $1200 a pop and needing a shim to fix awful heatsink spacing, I figured it would be wise to err on the side of paranoia caution.

 

Also slight differences in the pads can cause the heatsink not to sit flat, I adjusted mine and got it to sit down closer

 

Meaker@Sager has a great point here. Thermal pads are often thicker than need be, made to compress to some degree. If they reach a point, they may not be able to compress and will not allow for the heat sink to sit flat. I've seen some stretch the pads slightly to adjust it and make it thinner. But, if this is not what is wrong, you'll have destroyed the pad and need a new one. Also, you would need to identify which is the culprit first. This also isn't a universal issue, so may not apply.

Remember, when designing these systems, the heatsink and cards have tolerances on how high or low components sit. Because of this, there are certain problems you may find (think logically). This is one of them, but cannot be said to be primarily the problem in this situation...

Sent from my SM-G900P using Tapatalk

 

If they were independent heat sinks, I would agree with you. But since the DM3 uses a single monolithic heatsink, there is zero margin for error, since both cards must be perfectly flush. I also failed to mention that I had to install a set of very small washers underneath the master card before all this so that it wasn't angled, since even a shim wouldn't fix that (without a lot of very awkward sanding).

 

Meaker how did you adjust it? I'm curious. Did you bend the heat sink?

 

Bending a vapor chamber would lead to tears, the pads are what I adjusted.

 

I'll be using pressure paper to make the adjustments. Pretty expensive stuff...

How much pressure do you think the heat sink applies to the die and IHS? Under 60lbs?

::iunlock::

 

Definitely under 60. I recall someone saying around 20 or so.

 

Well noted. Thanks!

 

Laptop heatsinks are use much less pressure than their desktop counterparts.

 

Yes indeed. I'll have some accurate psi measurements soon and will report back.

::iunlock::

 

Did you ever get the psi measurements?

 

I'll have them here shortly, likely this week with the results posted in the 775 thread(s). The pressure paper came in a few days ago and currently I've been running benches on the P775DM3 on stock paste to obtain a solid base.

Cheers