Fully Internal Watercooled SLI system DIY project

I wondered why the most recent thing I could find was the Asetek M18x prototype from 2013 so I made one myself to investigate whether fully watercooled was feasible going forwards.


Preliminary testing:
CPU + GPU temperatures exceptional
Stock fans/rads not sufficient to keep loop temperatures acceptable.
This was something I was expecting since the P370EM has weak fans and small rads, and ran hot on its stock cooling even for the lower power hardware of its era (55W max TDP CPU, 65/75W GPUs) and had plans for.

GPU/CPU temperatures themselves are fantastically low though. About 10C over loop temp. CPU maxed at about 70C. Highest GPU temp I've seen so far at 115W is 65C.

But the loop heats up beyond >40C which is getting concerningly close to the 60C softening point of the cheap PVC tubing. (I have some more temp resistant ~200C silicone tubing en route)

With the external fans on and the 240mm rad cooling as well, loop temps stay acceptable <10C dT over ambient and the CPU sub 70C, GPUs ~60C.

The 3940XM somehow managed to drop dead during fitting and testing so the maximum CPU load from the 3720QM @ 3.8ghz is 65W so far which it handled at 70C

Fan control is through the RLECViewer fan control app - it is purely based on temperature for now in 5C gradients, which is good enough for now (although it does kill Fn+9 max fans functionality). In a perfect world it would be nice if the Obsidian fan control would work with the P370EM, its combination of load based and temp based schemes would be ideal.


Current status:
Connected to external pump/res (and optional 240mm rad)


TODO:
mod and add pump to ODD bay + wiring (slimline sata connector 5V -> boost converter -> pump 6V-12V)
fab and add res with bleed+fill port upstream of pump
redo GPU2 vrm/memory heatsink with right angle facing the other way for the line into HDD bay area
mod the case to fit the tube back to the ODD bay area going between RAM1/RAM2/msata and battery frame area
add 5V fan+rad+wiring to HDD bay (wiring: DC and ground lines spliced from GPU2 fan header)
add 5V fan+rad+wiring to ODD bay (wiring: DC and ground lines spliced from GPU1 fan header)
replace/tweak fan control of GPU1 fan? sounding a tad whiny/dodgy/occasional non-spin-up at low RPMs
System BIOS flash to the P370EM3 version and/or GPU VBIOS roulette to enable updates of 2+ year old drivers while retaining all display outs!!!
figure out how to route the external loop section via the detachable quick disconnects. Probably will have the inlet going into the CPU and the outlet after the slave GPU. That shouldn't cause too much of an issue with opposing flows from the dual pump/spliced lines.
replace PVC tubing with silicone (higher heat tolerance)

 

leak testing: post-double-45-U join soldering but pre-final assembly low temp/heat plate soldering

part way through the long and tortuous alignment process


Schematic of what is currently installed (except for pump/res):


Current final plan (except for the external-in to CPU heatplate, crossing that over the GPU1 heatsink is a no go):

 

Some testing.

Furmarking at ~250W load (2x115W GPU + part load CPU)
The temps below are after about 1/2 hour allowing the temperature to reach equilibrium
Loop temperature stabilised at a very toasty 50.6C which represented a delta T of approx 25C to ambient
The waterblocks themselves are working well, delta T of about 15-20C on the CPU and 10-15C on the GPUs (compared to loop temp)
Some throttling started on the GPUs which I wonder if it was the VRMs and the old thermal pads on them, combined with the plates getting toasty (GPU2 memory/VRM heatsink is literally the hottest point in the loop after all the heat generating bits)

This is absolute worst case scenario but still compares well with the stock results with both the 1070s at 115W, of low 80s master and high 80s slave.
The two additional fans and rads I'm planning should keep that loop temp down further.



This is after reaching that 50.6C high point, I connected the fans (1200rpm silent fans in push/pull so not that great) on the external 240mm radiator which brought the temps down to mid 30s over a couple minutes. The "70%" fan speed the laptop fans dropped to at this point is fairly quiet. I have the minimums set at "35%" in the fan control app which is almost inaudible.



There's a bit of a delta T between core 0 and core 1 on this CPU (not shown in the screenshots). Something I remember observing in the P170EM as soon as I liquid metalled this particular CPU. At times it's as much as 10-12C. Alternatively, perhaps it's because there's no fill in the gap under where the copper pipes are soldered to the heatplate (the low temp solder is very thin and flow-y). In any case I will redo the CPU heatplate at some point to incorporate the external quick disconnect line-in.

I tried to play whatever I could find on the data drive - Doom 2016 managed to run without steam even being installed (???!) - but it was rubber banding between 30fps and 70fps quite badly under DX11 SLI, damn these old drivers from 2016

 

I love this project. I featured it in the latest Forum Spotlight.

Great work!

Charles

 

Thanks. Haven't had a lot of spare time for progress of late but have done the wiring and fitment of the pump and everything in the ODD bay which required a bit of case grinding to make the space

 

Nice mod! I'm looking forward to see the final result

 

Awesome mod

 

So am I! I just came home after a week of holiday. Cramming everything into the ODD bay took a lot of time and a few different attempts at sealing to get it watertight.

Cheers

 

The water finally flows.

However during assembly and leak testing I've broken one of the seals I made causing a pinhole leak near the intake to the pump. Even though its so small I could probably deal with it coating the outside with liquid electrical tape I am being perfectionist with leaks, before I fasten everything with cableties.

I misordered the silicone tube, it is only 1mm wall and is too flexible and kink prone. Bummer. $20 wasted unless I spend time wrapping it in tape or something.

Pump works just fine on 5 volts even though it's spec is "6-12V", small bubbles take about 20 seconds to do a full loop which is sufficient flow. I've still got the boost converter in between the 5V source from the slimline Sata with the potentiometer screw accessible so I can turn it up should it be needed. That's all wired with 3 pin connectors which is how I can run the pump like this with an external power supply.

In hindsight I should have just got a sheet of 1mm or 2mm acrylic and cut and shaped it to fit.

 

Ugh. Looks like the EC panics on POST with the pump drawing from the 5V pins from the slimline SATA on boot. Upon pressing the power button, the pump fires up, but that's it, no display, and after about 20 seconds the fans go to max and the Capslock and Scroll lock LEDs start flashing. Disconnect the pump = boot as normal

Pump itself shouldn't be breaking anything, while it's rated for 0.35A at 12V I can't see it drawing more than 1A at 5V given how much slower the rpms are

I may have to integrate a switch, or get the 5V from elsewhere (like USB), or figure out how the enable pin in the slimline SATA connector needs to be altered (I think I read somewhere for the specification it's a 1Kohm resistor to ground, but will have to try and remember where I saw that.)

 

The two added radiators are installed and working.
(though intakes are not yet cut through the bottom for either of the added fans, a few more case mods to go)

Disconnecting the pump during POST is required for now, reconnecting straight after has no problems. I may add a teeny DIP switch for that in the meantime. Might get a double and have a speed selector as well - either bypass the boost converter for 5V, or through it for a higher voltage (up to 12V but it's really noisy higher)

The two fans running off each fan port seems to be working fine. The yellow (sense) line is only routed to the original GPU fans. The secondary fan on each line has the DC +V (red) and ground (black) lines connected in parallel, and that's all that's needed since speed control is through variable DC voltage in these old Clevos.

I haven't checked for voltage drop under load.

10 minutes of Kombustor at 115W per GPU
Loop temp approx 45C which is acceptable, right in the target range.
Maximum GPU load temp 65C
Maximum CPU load temp was 70C (this is 63W running TSBench - on the stock fan curve I have set up, if on max fans it would be in the 50s)
(eagle eyes may notice the 85C max on the Afterburner graph but that was when I disconnected the pump for a few seconds to see how quick temperature rises without flow - quickly at approx 5C/second...)

For comparison, the stock heatsinks would reach ~80C on the CPU with liquid metal, and 79C/86C on the GPUs at 115W each.

I redid the GPU2 VRM/memory heatsink with a different output direction and flattened the bottom of the copper tube more and used more solder, it seems to not feel quite as burning hot to the touch now. I think when I add the external lines I'll do the same with the GPU1 heatsink.


The dips in usage/power in the last couple of minutes were not causing dips in FPS reported in kombustor.

This is what it currently looks like:
I'm not happy with the 90 degree join near GPU2, but I stuffed up with the available clearance against the bottom plate (I wanted it closer to the corner of the battery bay but there's less clearance there and that ridge where the bluetooth addon module would sit looks rather... structural, so I want to avoid cutting it.) It still allows RAM to fit in the slots if I ever put more in.
There's a couple of other places where things don't quite fit yet, the housing for the fan on the lower right needs to be shaved as it's pushing right up against the frame on the top and bottom and that moves the vertical piece of frame where the HDD bay cover latches into. There's still actually a 2.5" SSD under that fan which is why it's rather crowded - airflow through that fan isn't the best since the copper tube is routed through the inside of the housing, and the case still blocks half of the radiator output (not shown) but it works.
The exposed wiring with 3 pin connector is of course so the pump can be disconnected for booting.

Here's whats crammed inside where the ODD bay used to be (lower left on the above pic):

There's a screw cap on the top left for filling/draining. the potentiometer screw from the boost converter is poking through where the ODD bay used to be anchored to the frame so it can be changed through the battery bay.
I have a wad of foam that I have sitting on top of the pump for some vibration damping (it sits right underneath where your right palm would), you can see I even had to shave down the sides for it to fit without bowing the wrist-rest part.

 

I was part way through cutting some holes in the case for the fan intake and pump control switches and - at nearly 3mm thick - I realised the base layer of plastic alone is about half the total thickness of some of the modern ultralight things these days haha

 

Behold, my wiring



Two selector switches.
One for source, either
- From the 3-pin source that will be exposed in the battery bay area, from an external power supply, so the loop can be filled and bled with the battery out and the laptop powered off (and in case there's a spill, avoiding letting out all the magic smoke!!), and for avoiding the EC panic on POST)
- or through the slimline SATA port when on

and another switch to either bypass the boost converter and deliver 5V (quieter), or go through it for a faster speed (used for bleeding, and maximal performance)

The hot glue is there to isolate the connections because they'll all be crammed in spaghetti style close together, and also to reinforce it as these super thin wire loves to break when bent...

 

Wish I had the balls to do this on my Ranger, would surely benefit immensely from water cooling

Love the work so far, thanks for sharing!

 

Thanks
Having spare parts around from buying up broken units cheap over the past couple years as they pop up on ebay/gumtree reduces the anxiety a fair bit. I had enough spares I was able to recently build up a complete unit as a replacement for my dad's old slow dual core HP, 3.8ghz quad core with SSD is light speed compared to that thing

Nearly done rejigging a few things and adding the external attachment bits, it'll all be back together soon and getting leak tested (2mm thick silicone tubing finally arrived) then benchmarked.

Last week I grabbed a 3940XM pretty cheap on ebay with a Dell Precision attached, it was missing a few cosmetic panels and a hdd tray ($50 tops on aliexpress) but I put it to use (with a lesser CPU ofc) to replace the the laggy Core 2 Duo Latitude sitting on the desk running the barcode scanner and business CMS. Each Windows 10 update and Intel bug fix slows it down more

 

Oh yeah having spare parts laying around definitely takes the edge off. The ease of replacement parts is what made me stay with Dell for the most part.

 

Installed and working.
The silicone tube idea was a bust, it's so much easier to work with and much more flexible, it went on nice and easily but wasn't tight and leaked immediately, I simply CBF doing cable ties for every. single. attachment. when they're on and off regularly for testing, so it's back to the cheapo PVC stuff for now until it demonstrates a problem.



The fine routing bits for my dremel are getting blunt and it was hard to make the switch cutout a nice square.
I had a plan to use my holesaw bits on my hammer drill for the fan intake hole but I just couldn't arrange my clamps and vice to hold the case still enough, so in the end I disassembled a busted fan and taped the metal housing against the inside, the sanding attachments die really quick against sharp metal but the hole is at least circular and fairly neat


*almost* able to fit the bottom plate on. This is a mangled one I've stuffed around with in the past (the ugly cutouts are from a while back when I had an idea of making see through sections abandoned that idea) and I have a spare I'll mod neater for a final product. There's still a couple of spots there's things with too much height. The tubes and joins down near the bottom right are a nightmare and stacking the tube on top of VGA1 heatsink doesn't fit, in both cases silicone would have squashed down without issue but the vinyl doesn't. It doesn't really matter but it would be nice to have it look "complete".


"Sleeper" lol. I got a bit too close when grinding making space for the pump, it didn't go through but I had the dremel too high rpm and the heat warped the plastic there, and combined with it being thin, it ended up breaking off making the hole just to the right of the wrist rest which looks white because I filled it with some epoxy

 

Copper desolder braid plus a fair bit of solder makes for good reinforcement when joining copper pipes together. This join cracked a couple times from the force upon it from the external connector.

 

Now, the problems.

The dual-loop setup just doesn't work well at all due to opposing flows. The external pump even though it's a tiny DC-LT is a lot more powerful than the even tinier pump in the laptop. I set up the external loop to combine its output with the output from the internal pump in a Y join, sitting on top of the GPU1 heatsink, then there would be a common path across the main heat generators being the CPU, GPU1, then GPU2, then it would split back into the two loops after GPU2. However water works a bit differently from air it seems and instead of combining, the output is pushed out both of the sides of the Y join at the input which even at a lower speed is fighting against the internal pump. If the internal pump is off, water actually flows backwards through it.

I might have to separate the two loops completely and have a separate line for the external, and just use it like a 'boost' to keep the internal loop cool.

GPU2 mount seems to have been misaligned somehow, temps are 10-15C higher than GPU1 now which wasn't a problem before.

The 3940XM is too hot for my bodged together CPU block. It jumps to over 80C immediately under 37x (about 65W). I have a spare and will make it better.

The higher speed setting for the internal pump (approx 8V) does make a constant difference of about 3C throughout vs the low noise/low speed 5V.

The extra rads do their job and the cold ambients of Melbourne winter at the moment mean the cooling is working well overall. I am yet to test it on a dual adapter and crank up the GPU power but ~115W per GPU is easily handled with low noise levels.

 

Gave it a quick test last night.

4.5ghz (with +50mV offset) on the CPU at ~80-85C around 75W on this 3940xm. A lot better than the other 3940xm which was constantly on the edge of thermal throttle requiring 90W and +140mV or so under the same TSBench load (have to dig up the old data) on aircooling.

Hooking up the external line the GPU temps drop greatly by >15C. It slowly cools down the internal loop but not by as much as I thought, but that's not so surprising I guess since there's no great amount of direct contact between them, more just pulling heat off the GPU heatsink plates.

I've only used it on a single 330W so no 150/170W GPU draw testing yet.

The long copper tube leading to the external qdc is causing issues with the GPU2 core mount. The oblong GP104 chip there is oriented at a right angle to the weak spring arms meaning it quite easily rocks back and forth. When in the right position temps are within 10C of GPU1, but sometimes shifting the laptop slightly will cause temps to shoot up to near thermal throttle. I might have to cut the pipe and join with tubing to be flexible. GPU1 doesn't seem to have this problem, either the arms are stronger on that heatsink or the die orientation prevents it.

The silicone tube is a huge PITA. It's supposed to be 6mm ID (1mm thick so 8mm OD) but it's still a bit loose over the 1/4" (~6.25mm) copper pipe, and especially where I've had to manhandle the pipe for tight joins and it's not perfectly smooth, circular and straight. I found pushing the tube on over the top of some epoxy would dry to give a reasonable seal, but still needed cable ties pulled as tight as I could (the only thing I could think of small enough to fit in the extremely tight spaces) to stop pinhole leaks just because the silicone is too flexible to stay in place.

Like all micro sized pumps, it loves to get some bubbles trapped in it and get noisy. Hopefully this is something caused by straight water and the green alloy automotive coolant mix I use on the other one with a bit of glycol in it will help that. I thought duplicating the pipe near the end would slow flow rate and I've got a bit of tube inside my home made res taking liquid from near the bottom but the problem is the pump flows fast enough to recirculate bubbles even on 5V and there's no high point air trap or bleed screw.

Pic is during leak testing (before I installed the last bit on the bottom left). There's about 3x as many cable ties now.

 

wow I would never have the balls to set up something like this. Well I dont have the knowledge or equipment in the 1st place.

 

Nice man! looking sexy!
I hope someone would do a Hybrid cooling mod somehow, probably in the future. When pumps are smaller.

 

About time I posted an update and some pics since this has been working pretty much flawlessly for months now.

Not a huge amount of stressful game loads to test it as I've been letting my son use it and he has been enjoying mainly Terraria, but a bit of Doom and Kerbal have put it through its paces. Unfortunately the ancient drivers needed for the bodgy vortex 1070s keep this from being more than a proof of concept.

But one time when the lady of the house was interstate at a seminar the pair of us played through Doom 1 and 2 (gzdoom / brutal doom) downstairs, I took my portable cooling station built off a 240mm rad, tiny DC-LT pump and res, and 12V power supply, hooked it up to both our laptops, and it was working fantastically. Very quiet setup.


Bottom plate fits on... almost, there's a couple of bulges. Not super happy with how far the connections protrude out the back but at least they are not fixed (like they were in previous versions) so are less vulnerable to being knocked and breaking something.
The tubing is 8mm OD / 6mm ID (1mm thick) silicone. It's a little loose over the 6mm copper tube for my liking so I found smothering epoxy then putting the tube on would give a great fit, and the epoxy wouldn't stick to the silicone. From there the dual cable ties (added after that pic above) would act like hose clamps for better sealing. The tubes are fixed in place pretty hard, they are not moveable by hand at all.

Visible on the bottom left are the two white dip switches, one toggles between 5V and whatever the boost converter is set to (the pump is rated for 6-12V but gets too noisy, so currently is at about 8V) which is adjustable by the tiny screw on the blue potentiometer housing, and the other switch toggles between internal (from +5VDC and ground of motherboard ODD SATA port) and external (1mm 3 pin poking through the same hole as the potentiometer). The battery hides these when in place. This system is fully functional and useable on battery power alone.


At a more "sane" 4.2ghz, a 1.2ghz OC over base 3940xm clock.


Long term CPU load @ 4.6ghz. I had never burst run that high ever.
Interesting to note the other CPU on the aircooled heatsink would report over 90W consumption during x264 at only 4.5ghz. Before it overheated and started throttling. (Here, ~78W at 4.6)
This was with the pump on high speed. At low speed it was a few degrees hotter.

Basically, 4.6ghz on the CPU long term stable under 90C, way better than aircooling could ever achieve, and because the cooling is distributed across 5 fans it could run that load at low fan speeds if I had more versatile fan control software (i.e. if Obsidian Fan Control worked)

GPUs run their load at 61C+71C where stock was 79C/86C from memory. With the external line hooked up they can run at 200W long term stable in the 60s at low noise levels. Clearly having the 2nd line running around the edge of the heatsink is more than good enough. Just as it was on my original hybrid setup.


Idling (if that's what you can call it with the behind the scenes Win10 crap being constantly active) in 16C ambient.


Kombusting at 100W each GPU on the older GPU2 heatsink. Those dips in GPU utilisation started after about 20mins and could have been a VRM thermal throttle as the heatsink would actually get pretty warm, but I couldn't see a fps dip that corresponded. So I redid the heatsink flattening the tube more and adding more solder for contact between the pipe and the aluminium and it seemed to help. Were I interested in improving it (!!) I would grind a path into the aluminium for the tube so the actual cooling was much closer to the VRMs.


I'm amazed at how my home made GPU waterblocks actually turned out. I wish I took more pics during building them but basically I cut and soldered the 6mm copper tube together, then ground one side off to expose the inside, then soldered that (with low temp solder so the tube structure didnt fall apart) to the plate from the disassembled stock heatsinks. I used desolder braid (fine copper mesh) to strengthen and gap fill where required.

The system does need a bit of topping of a few mLs every couple of weeks. I think the silicone tubing might be a bit permeable as I have never seen any leaks or dried residue anywhere.

 

FYI the WC setup adds about 500-600g weight all up, loses one 2.5" drive bay and the optical drive bay (oh no).
Left: WC setup, mSATA SSD + 2.5" SSD ~4.6kg
Right: stock P370EM, single GPU, mSATA SSD + 2.5" spinner hard drive, BRDVD drive in ODD, ~4.0kg

 

I have recently finished servicing this after it sat for a few months, very much unloved ...

Don't leave tap water in a loop for any amount of time, my stupidity, lesson learned. It turned a shade of orangey brown tat required a lot of flushing with vinegar to remove the gunk, the pump got fouled, that needed a thorough disassembly and clean as well.

It now runs with a single 980M (from a P870DMG) on stock Prema vBIOS because I'm sick of the driver BS, I have set this up for my son to use until it dies. It should handle Terraria, Kerbal, office 365 until that happens.

CPU is set at 4.2ghz on a slight overvolt (+15mV, sits on 1.26V load) and sits on around 80C, even under sustained load, I have the fan curves set low because the one GPU is laughed at by this setup that would cool >300W.

GPU gets the infamous 3-FET 980M blackscreen after about a minute of Kombustor at stock... the water line is too far from the VRMs and the aluminium part of the heatsink just heats up too quickly. I thought something was wrong until I checked HWinfo and it's drawing over 180W... not too concerned since it won't get anywhere near there in normal use. So I have a temp limit on the core of 75C to prevent this, should be rock solid.

This model was released over 8 years ago now. It's user is only a year older haha.