P870KM1 SLI "things".

I recently bought a second 1080 for my KM. My setup is as follows:
- KM1 flashed with stock TM BIOS and stock KM ECs,
- 9900K,
- TM CPU heatsink,
- TM GPU vapour chamber,
- Two 1080,
- AUO B173HAN03.1 screen running at 144MHz.

First thing I noticed after mounting second 1080 was high temps (~50C) and fan noise while doing „non 3D” regular things (web, office etc). After inspection I saw a very high GPUs utilization at idle, even on desktop. Both cards almost never properly sleeps at lowest P state and 139MHz clock. When I open one tab (!) in Chrome (no fancy stuff, just plain text web page or something light like this forum) and scroll it up and down fast, both cards goes to 1511MHz clock (!!), shows 40~50% 3D utilization (!!) and eats 40W each (!!!) and both fans starts to ramp up. It`s insane, 80W just to display a fast scrolling letters!

I`ve never noticed this with single 1080 but it doesn`t mean that it wasn`t happening. So I don`t know if this behavior is SLI specific (and that`s why I noticed it only now) or general 1080/Pascal/nvidia issue (single 1080 was the same but vapor chamber cooled it well enough that fans stayed quiet and I never noticed).

Interwebz lecture says that is normal, and that`s a cost of feeding 144MHz screens but those reports mainly focuses on multiple displays configuration. That I can believe (much more screen pixels to handle) but in one 144MHz screen scenario it looks like some serious bug in nvidia drivers.
It`s driver and Windows OS agnostic, happens here with 38x.xx/39x.xx/4xx.xx and Win7/10 combinations.

Can you confirm it being normal with yours 144MHz+1080 SLI setups?

To avoid that, I found two solutions:
- switch screen to 60MHz refresh (120MHz is the same as 144MHz) and everything is back to normal, cards idles at 139MHz clock, web browsing won`t wake them, fans stays quiet, all is fine but… 60MHz
or
- stay with 144MHz and block both cards in P8 state with nvidia inspector and add 3D programs/games to the exclusion list (threshold setting didn`t work for me) which is only a minor inconvenient for having 144MHz with proper (cool and quiet) idle operating.



Is there anything more, another solution to this issue?


Also I would like to hear your opinions about cooling efficiency, some „audit” from others with TM VC and 1080SLI.
I always aims to stay at stock clocks/turbos but to undervolt GPU as much as it`s stable.
So master card is undervolted to 0,925mV


And slave is undervolted to 0,900mV


Cooler running card (lower voltage=lower wattage=lower temps) is on purpose put in the slave slot which is harder to cool.

Vapor chamber is polished like that:



For both GPU I needed to use 0,5mm shims because of 0,5mm gap (precisely measured – that`s the reason I couldn`t use liquid metal), shims also polished but not so much:


Thermal paste used – Gelid Extreme.

Now on to testing. Stress test with Time Spy Extreme so cards were slammed hard constantly for 20 minutes (max fans, 22C ambient).
Master peaked at 175W and slave at 180W but 90% of time they were in 155-160W range:


As for temps, master peaked at 90C and was flicking between 89-90C constantly. Slave peaked at 91C and stabilized with 90C flat:


Both cards started to throttle at around 10 minutes of test - clock was flipping from 1657,9MHz (lowest) to low 18xxMHz. They throttled still stayed in highest P state and turbo range:



For reference, the same stress test in the same conditions and configuration (shim, gelid, max fans, 22C ambient, etc) BUT with single card (master one, that „hotter” running at 0,925mV) gives me 63C max – no benchmark loop can bump its temperature higher and it was (is) great, with SLI it`s not so great anymore (or is it?).

So this temps are normal and typical for 1080 SLI?

 

Normally yes. But it can be sli cable issue. What is the number from sli cable?

Fastest Customizable Gaming Notebooks & Mobile Workstations | CLEVO Integrator

 

It`s 6-43-P8720-010 and it`s fine I`m sure.

 

Your polishing will have also created more of a gap, with the right thermal pad layout liquid metal use is possible.

 

Measurement was done after polishing, it`s 0,5mm for both GPUs. Gap is the same even without any thermal pads in place so with liquid metal applied contact (and heat transfer) is terrible. I`ve tried.

 

Was that measurement done outside the case?

 

Inside, with both cards in MXM slots and secured with screws and with vapor chamber fully mounted (all screws in place).

So what, polishing/gap/liquid metal aside - temps should be lower?
I'm reading TM1 1080 SLI reviews and they showed even worse temps (with stock everything so it looks correct).

 

I would say those GPU temps are WAY 2 high.

To give you an accurate idea, that card you purchased from myself, wouldn't exceed 65 degrees under load in SLI when sitting on LM that was with both cards sitting on +550 memory and boosting up to core 2088 which they were stable at and passed all tests.

I would agree with @Meaker@Sager that you have a gap there.

From personal experiance a small gap even 0.5mm will create 12 degrees onto temps straight away ( i think this is right - going from @Meaker@Sager info) when he helped me on my watercooling build.

You shouldn't be getting more than 65-70 in SLI in my opinion if using LM and everything is touching correctly.

 

Also it should be hitting 200 watts ! it's thermal throttling here dude, you have an issue with contact i would say.

 

I'd listen to brother cope here, the whole contact issue is fresh in his memory

 

Amen to that

 

I've added additional 0,5 mm shims for both GPU to definitely close any gaps there (so 1 mm in total) and... temps are exactly the same.
I`m out of ideas gentlemen.

 

Could be the vapor chamber is not operating properly.

 

And how is that possible?
There is no visible defect on it anywhere (dent, puncture, etc).

 

When you did the tests, was that with max fans fn+1?

 

Yes, of course - fans were blasting past 4k RPM.
I'm a blind baby in a foggy forest.

 

It sounds like the vapour chamber is somehow bent and isn't making contact with the chips.


A good way to troubleshoot is, do the temps shoot VERY fast to 91 when you hit a stress test? @artpra

 

That would point out a poor contact with die straight away but it`s not happening. If you look at OP there is a graph with temps for both GPU combined.
It takes about 9 minutes until temps stabilizes at 90C and 91C suggesting a proper (?) contact and slowly reaching thermal capacity.

 

This is indeed very trange. I'd go back to basics. Remove both copper shims and only leave the thinnest thermalpads on. Then testfit the gpu onto the vapor chamber and with a flashlight try and see whether the die contact is proper. Do this outside of the chassis.

If it is fine, add the thicker thermalpads and see if they cause poor contact.

 

Thicker pads will reduce contact, the RAM should be 0.5mm and go from there.

 

I personally would remove ALL thermal pads for testing and drop on K50 PRO, this will eliminate any potentially raised thermal pads you are missing.

If you are still in the same position after > Remove Shims > you said 1MM gap - that's a possible 24 degree increase... 90-24 = 66 degrees - this is what both my cards maxed out in SLI and others have reported the same temps, i really do suspect your issues are due to some sort of pad/VC damage.

If it's still the same then I agree with @Meaker@Sager - sounds like the vapour chamber has failed.

You can get a brand new 1 now for the KM1 for around £140 i think.

Your running a CPU TM heatsink, so I am assuming you have had to mod your current KM1 heatsink somehow?

Can you show us some pictures of the heatsink sitting flat on a table - to see if it's warped etc. Also the layout of it installed the laptop.

 

I spent more than half a day for this and here are my findings:
1. VC is perfectly flat, no warping no bents, totally symmetrical. When laid down on thick glass (as flat as you can get) there is no wobbling, it rests firmly on 6 copper plates which contacts with memory modules (via pad).

Gap between glass surface and heatsink (where the GPU cores are) is the same for both master and slave card and it`s about 1~1,5 mm. All is great so far. Next!
2. Cards outside the KM, secured with 8 screws to the VC, no pads on both cards - gap between both GPU cores and heatsink is exactly 0,5 mm. Yeah, nothing new here.
3. Cards inside the KM, secured with 4 screws to the motherboard, 8 screws to the VC and with 4 screws to the frame, no pads on both cards - gap between cores and heatsink is (no surprise here) exactly 0,5 mm for master and slave. Next!
4. Cards outside the KM, secured with 8 screws to the VC, pads being added one by one and inspected each time with flashlight (if there is proper contact with VC, not more than is need) until both cards are fully padded. Full success: pads are in place, all with proper height (fitted exactly for my VC sample). Gap between cores and heatsink is still exactly 0,5 mm for master and slave, so each and every pad is correct - they don't raise the VC at all. All good.
5. Cards inside the KM again, secured with 4 screws to the motherboard, 8 screws to the VC and with 4 screws to the frame, with all pads on both cards in place (as in step 4) - gap between cores and heatsink is exactly 0,5 mm for master and slave. Bingo!
6. It`s time to mount it, so: VC, Gelid, 0,5 mm pure copper shims (to precisely fill both gaps), Gelid, GPU cores. All good to go.
7. Time Spy Extreme stress test, max fans (Fn+1), monitoring log start and...

Red is master, green is slave. They are hitting 90C and throttling like before (that`s why I ended stress test at 13 minutes, there was no point to run it further). But there is some progress - previously it taken 10 minutes to reach 90C now it takes 12. And pads are proper for sure - there was no black screen (board is properly cooled).
Hurrah, whoah, fanfcukingtastic!

 

I would be really careful dude, those cards are hitting 90 degrees, that's way too hot for them in my opinion.

Can you upload some pictures of your thermal pad layout you are using on the cards - with an edited pic of each width of every pad?

I really cannot see it being anything other than thermal pad layout or vapour chamber compromisation.

Have you modded your KM1 heatsink, i am assuming so as your running a TM heatsink ( show us pics )

I personally never ran copper shims with my VC, why do you need the shims bro ?

 

I'd like to chime in with some of my experiences. I've been fighting a very similar issue for a while. Laptop is in my sig.

Some background. After first getting my laptop I flashed modded vbios's which raised the TDP limits of both cards to 250 watts. I repasted with liquid ultra and installed thinner thermal pads (0.5mm for vram, 1.0mm for chokes, 3.0mm for mosfets). GPU's were set in afterburner to 120% TDP with a 155mhz core OC and a 500mhz mem OC. Master was undervolted to 0.975V and slave was undervolted to 0.9625V.
Everything was thermally stable long term in games. I mainly play the Witcher 3 maxed out with mods at 4k 60hz which results in 80-95% load on the cards. Thermals never went above the low 80's for several months.

After a few months temps started to creep up to the point where the cards started bouncing off the 92C thermal limit. I figured maybe the liquid metal dried out so I repasted with new liquid ultra, and cleaned the fans/radiators thoroughly. This resulted in no change in thermal performance. Next I tried lapping the VC which resulted in atrocious performance due to presumably uneven levels between the two contact patches.

I bought a brand new VC and installed it with liquid metal. No change in thermal performance. Still bouncing off the 92C limit. I attempted to solder additional fin stacks to the radiators and popped the VC.

I bought another brand new VC and did the following:
-Instead of lapping this one, I chemically polished the surfaces with a mild acid and NaCl to ensure only the smallest possible amount of material would be removed (probably just a few atoms thick). The polish came out great - similar to OP's pics above.
-I next applied thermal Grizzly Conductonaut to the VC contact patches and let it sit for 2 days in a dust free environment so the gallium could oxidize into the copper as much as possible.
-Then I applied more Conductonaut VC and installed the VC into the laptop with a very thin foam dam barrier to prevent Conductonaut from escaping and to minimize the amount of Oxygen present to react with the gallium/copper.
-At the same time I also did a backplate retention mod on both 1080's to significantly increase contact pressure.
-I also used thermal epoxy (instead of attempting to solder) to attach additional fin stacks to the VC radiators to take advantage of the open gap where the fans blow over the tops of the existing radiators.
-I then removed all of the metal and plastic grates from the bottom and rear of the bottom bottom cover. The fans, VC, and radiators are all 100% open to the outside environment.
-And finally I purchased a 100CFM laptop cooling pad.

What's the final result? Peaks to about 82C under heavy load, with averages in the mid 70's. Not great at all considering the original heatsink with no mods other than liquid Ultra and thinner thermal pads managed the same thermal load at about 80C for months . . . until something changed.

I also attempted to run with no thermal pads, except for 2.0mm pads on the mosfets and Thermal Grizzly Kryonaut on the chokes and vram chips to further increase contact pressure. The difference was negligible. I ultimately went back to the thin thermal pad configuration because I wasn't too comfortable with the contact patches for some of the vram chips. Do vram chips need significant cooling? I know on a lot of cards - desktop and mxm, they are not cooled or barely cooled.

This has been an extremely frustrating, expensive battle which has gone on for almost 6 months. What can be done to definitively get these cards under control?

What would the SLI cable have to do with temperatures?

 

The ram making contact helps as there is no incidental airflow.

 

Pads are 100% proper, there's no doubt. Spent many hours on them, inspected each one, experimented with different thickness until I've found such configuration that touches VC properly and doesn't make the 0,5 mm gap any larger.

How can I check if my VC is empty or not functioning properly (except, well - my temps lol)? Maybe it really is broken.

I'm not using KM VC at all, I'm using TM variant. No mods whatsoever, brand new, straight from the box in to my KM1.

I'm using shim because there is 0,5 mm space between VC and a GPU core - even without any pads in place, "naked" cards and VC aren't making proper contact. Without shim temps are even worse. But let's assume that shim isn't needed and should be removed. Keeping it wouldn't still be the reason of such high temps: it would result in great core temps and sudden black screen due to overheating card components (not touching VC because of unnecessary 0,5 mm lift caused by shim).

Something else is going on, dunno.
Is it possible that KM VC is SO MUCH better than TM one?

@aarpcard I feel your pain mate, stay strong!

 

The TM vapour chamber is better than the KM1, maybe you do have a defective VC.

Your cheapest option is to buy another and test, those cards shouldn't get to 90 for 10 minutes, you will end up burning 1 out and costing you a small fortune bro !

 

@aarpcard already tried this way and ended with zero improvement but another VC is the only solution I can think of.

Anyone here have a spare TM VC? I would like to borrow it for short, non harmful test.

 

I can't see what else it can be guys, sorry but your eyes would struggle to see a 1mm gap under a heatsink when its mounted.

You would need to use pressure tape.

It has to be some contact issue =)

 

I'm going to get a hold of some pressure paper and see what that reveals.

 

Disabled SLI, ran the same stress test (Time Spy Extreme full loop, ~20 minutes) but now only on master card only (with slave still inside), max fans, 23C ambient: card was constantly on full blast, full turbo clock all the time (stable 1911MHz), was drawing ~155W (with peaks to 170W) and... it stabilized at 63C! Freaking 27C lower compared to enabled SLI!
It doesn't look like contact issue, unnecessary shim or wrong sized pads - it looks like great temp with a great cooling solution. Amazing temp, exactly the same as when I had only single 1080. But then with SLI on - it all craps.

So VC is working, maybe. I`m even more confused now.

 

Well it is double the thermal load so it puts the cooler under a lot more work.

 

My SLI setup wouldn't go past 70 ever at full load on Wticher 3 - there is something more to this

 

What voltages were you using? Sorry if you already posted it earlier.

 

I was using stock voltage.
No flashed Vbios.
both cards clocked at between 600-700 memory and +175 on core giving me 2088 boost.

in 4k on witcher after around 3 hours , both cards would sit max around 63-65

 

That's a very healthy core clock

 

http://forum.notebookreview.com/threads/p870km1-g-cpu-heatsink-waterblock.829341/page-12

See above - GPU's peaked at 70 in SLI at the same time as maxing out the CPU using Aida64, the gpu's under full load then dropped to 63 when i added a laptop cooler.

You guys have the TM setup which should in effect not allow heatsoak to get to your GPU's like my setup.

 

Thanks bro - stresstech told me my 1080's in the laptop were hand picked from over 28 samples.

 

It does make it a little unfair to do a comparison against regular cards. Though normally a well fit heatsink should maintain 2ghz clocks without throttle.

 

Actually, the OP @artpra purchased one of my 2 1080's, so I am giving him the cards performance when it was in my machine.
But yes i get your point.

 

At stock clocks and voltages, I'm averaging in the high 70's and peaking in the low 80's after mere minutes in game. Prior to all the mods I did, it would bounce off the thermal limiter within minutes. Boost throttles down to 1810mhz.

This is my third brand new TM VC heatsink. All three have exhibited the exact same issues. What are the chances I got 3 bad factory new heatsinks in a row? I also know someone with a KM with SLI 1080's that has the exact same issues with the KM VC heatsink. Really the only way to tell for sure is with pressure paper (which I have on order), but I'm starting to question if it's an issue with the heatsink.

What are the chances the chassis is skewed or the motherboard is skewed in the chassis resulting in the GPU dies being very slightly tilted in relation to the heatsink?

Honestly while the VC is a very cool concept, it's execution in these laptops is extremely poor. If it's true that a 0.5mm gap would result in a 12C temp increase, then you have a ridiculous amount of dimensions that have to be toleranced to within thousandths of an inch to each other: The z-height of the both MXM PCB's, the pitch of both gpu die packages, the z-heights of the packages themselves, the z-heights of each mxm slot, the z-height of the motherboard, the z-heights of both contact patches on the VC, the pitches of the contact patches, the mounting points of the VC to the chassis, the chassis dimensions, all have to be ridiculously precise. They're all referenced to each other. Any one being off a hair will mess up the entire system.

Even if I managed to track down what the offending dimensions are, I have absolutely no idea how I'd go about fixing them considering the level of precision required.

A much better design would have been two independent VC heatsinks (one for each gpu) mounted directly to the MXM cards with no interface to the chassis or each other. In that setup the only dimension that would matter would be the flatness of the contact patch which could be easily machined. This is extremely frustrating.

 

Possibly your posts either side are off, this is my second heatsink without issue.

 

"Posts" being the MXM standoffs on the motherboard, or the places where the VC screws into the chassis? I suppose it could be both . . .

 

More likely the chassis connection.

 

Guys,
any input about high clocks/GPUs usage with 144MHz screen?

 

Not on the latest drivers pushing two 1440p 120hz displays.

 

Should've been fixed in drivers years ago. Used to be that 144Hz would cause the GPU to stay in P0, only 120Hz and below would allow it to downclock unless Nvidia Inspector Multi-Display Power Saver utility was used.

 

IIRC it came back during the RTX launch for RTX cards in the first driver packages?

 

Doubt it, considering most higher end gaming laptops since early last year ship with 144Hz panels, it would've been a widely reported issue, but it hasn't been.

 

Figured I'd update this thread. I think I've solved my issues. I bought some pressure paper, but did not get around to testing the die contact patches. I'll probably due that the next time I need to do a repaste.

However, I also bought some K5-Pro thermal Pad replacement: https://www.amazon.com/viscous-thermal-replacement-60g-Aspire/dp/B00K04D3UK

This stuff is like a putty that you can form into the shape of thermal pads. It is viscous enough that it will not disperse like thermal grease, and it is formable enough that the pressure from the heatsink will force it to compress easily unlike thermal pads. I replaced the 0.5mm and 1.0mm thermal pads I was using on both cards with the K5-Pro. I did not replace the 3.0mm pads for the voltage regulators.

Since doing this, at stock voltage and a 155mhz overclock, with both cards pegged at 100% usage in games, I've yet to see temps over 72C. Usually they sit in the high 60's. I guess the issue, (or a major part of it) was insufficient clearance between the heatsink and the components that require thermal pads. Back when I first got the machine, I changed out the stock pads for the 0.5mm and 1.0mm pads, and that improved temps significantly. Not sure why they degraded from there, but eliminating the resistance from the thermal pads fixed it.