Log Monochrome 2: My custom fanless 7.5 L Strix Halo system (work in progress)

[TheJiral]

This is my custom project of a completely passive Strix Halo system based on the Framework Desktop Mainboard featuring an "AMD AI MAX+ 395" (how I hate that name) with 64GB unified memory. It is definitely not a rational or cost effective undertaking but rather a tinkering project, trying something new and see how it turns out.

​

Currently I have to wait, the case is ready, the custom cooling solution, including custom bent heat pipes is as well. It could take a number of more months though until I can finally get my hands on that Framework board. I read rumours about Strix Halo supply being scarce so maybe it will take longer than "Q3" that is officially the shipping window. We will see. Once I have the board I can see if my contact plate is actually working out or if I have to last minute modify it and then I can finally attempt assembling the heatsink and heat pipes with proper thermal compound.

 

[TheJiral]

Will the thermal concept work out? Only one way to find out. Theoretically, the heatsink should be capable of dissipating 140W when I manage to get it up to 60°C, which of course necessitates that the APU is running very hot. I would hope it will work also under load without too much thermal throttling but that is what I shall find out.

To illustrate it better that is the actual cooling system without the side cover:

And some more pictures with the side cover assembled, the PSU in place but obviously without the board:

 

[Sicaris]

Wow, this is truly a work of art. Can you show us some of the design and manufacturing steps? I'd be particularly interested in how you determined the size of the heat sink and how it was manufactured.

Thanks!

 

[TheJiral]

Well, it is a bit early for that I guess. I don't even know if it will work out yet.

What I do know is how my fanless Mini PC is working like: That system has a 28W APU (7640U). With two flat heat pipes and those heat pipes have not too efficient contact with the heat sink via 1 mm thermal pads. That is good enough. There is thermothrottling during full single core load and its hitting thermal limit during lasting full multicore load. Interestingly with full iGPU load during games temperatures stay pretty decent at less than 70°C (even though RAM is hitting thermal limit). From that system I also know that the heatsink can manage to reach 55-60°C or so (just a bit too hot to touch for longer than a few seconds).

The Mini-PC heat sink is actually a LED heatsink (PADLED-13080) and has a technical data sheet: https://wakefieldthermal.com/content/data_sheets/PADLED_WT_v2.pdf

At a deltaT to the room temperature of app. 30°C, the data sheet would specify a power dissipation of 33W. Close enough I think to what I see in reality. The data sheet is listing a thermal resistance of 1.5 °C/W. But based on the actual numbers it would look more like 1.0 °C/W. Anyhow, it is a starting point.

I found extruded heatsinks with technical data sheets and took the profile of one as inspiration for my heat sink. The commercial one would have a width of 200 mm and be 5 mm less deep with a thermal resistance of 0.3 °C/W. My heat sink is on one side 250 mm long and on the other 70 mm long. That should estimated get me somewhere around 120-140 with 30°C dT. If I can keep thermal losses low enough from the die to the heatsink it should theoretically work out.

What should work in my favour is that my heat pipes are distributing the heat along the height of the heat sink and there also 3 dedicated heat pipes to the front where there is more heatsink surface. What works against me is that some heat pipes are rather long at 250 mm. However there is additionally also a pretty large direct aluminum link to the heat sink which should also transfer some of the heat.

That's my reasoning. How well I estimated it all i will only see once I have it completed and I can test it.


Manufacturing

Ok, I possibly did not go the most effective route and most definitely not a cheap one. But I have to say I did enjoy to design CNC parts and see them turning out just as I wanted them. I was already aware that my design with slanted fairly narrow fins (for CNC) would probably need more advanced tools. I was ordering it form JLCCNC and they used wire cutting to cut the heat sink fins and then milled the rest. That came at a rather steep price admittedly. But at least I got app. 5 kg heat sink in return for it.

 

[TheJiral]

The heat pipes are another story. Those I bent myself and every individual heat pipe is bent differently. This was a real pain.
In the end I needed 3-4 bending shapes per tube to get a somewhat decent curve, while veeeery slowly bending it.

Those 3d printed shapes were screwed into the base and I had a metal tube as lever.

A first test assembly:

I replaced some of those pipes in the end. Lessons learned: Go for heat pipes with mesh if you want nice curves. Sintered pipes do still work out but grooved pipes are an absolute nightmare. Those groves stabilise the pipe and make it very challening to bend without knicking them, at least at minimal radii.

 

[Sicaris]

Wow, such an impressive endeavor! Even with the final test with components still to come this is such a well thought out project. Thank you for sharing your thoughts and process. Bending heat pipes is something I've always been too afraid to attempt, your 3D printed guides sound like an excellent solution. I look forward to seeing it all assembled, I'm keeping my fingers crossed that your board arrives in the first wave.

 

[TheJiral]

Thanks.

I have to make a small correction. The commercial reference heat sink with 0.3 K/W thermal resistance was 250 mm wide, not 200 mm. So my 250+70=320mm width would have roughly 0.234 K/W if I assume that it all simply adds up. So, to cool 140W the dT of the heatsink to room temperature would have to be 33 K. In other words, at 25°C room temperature, a heat sink with 58°C.

It is a bit on the tight side but I the reference was under the assumption of black anodized surface and vertical orientation but no heat pipes and with 5 mm less fin depth.

The real question mark is how high the temperature loss will be at the interface of die to copper spreader, from there to the heat pipe, in the heat pipe and then from the heat pipe to the heat sink.

 

[TheJiral]

Fanlesstech reported about this thread and there was a question on how I designed the cold plate without having the board in hand. The answer is that Framework has published a 3D model and some technical drawings here:

Framework-Desktop/Mainboard at main · FrameworkComputer/Framework-Desktop

Documentation for the Mainboard and printable mechanical parts in the Framework Desktop - FrameworkComputer/Framework-Desktop

github.com

That did help a lot but for some reason the VRM is not defined there. So my cold plate does involve some educated guessing based on those files and the available photographs. If it is just a minor issue at the VRM I will probably just wing it, cutting, sanding some minor details. In the worst case I'd have to reorder the cold plate but I do not expect that to be necessary.

 

[AlexTSG]

Your project has been reported on at Videocardz too. Well done, it looks great.

Custom 7.5L fanless case built for AMD Strix Halo MAX+ 395 Framework system - VideoCardz.com

Hopefully Minisforum will bring out one of their MoDT boards with Strix Halo soon, although I'm sure that would require some reworking of your current design.

 

[TheJiral]

I've seen it. Maybe a bit premature, given that there is still no "PC" in the SFF-PC.

I didn't know about a Minisforum Strix Halo MoDT. It would certainly make some sense. I wouldn't hold my breath though for it to be dramatically cheaper than the Framework one. It is really hard to say how much reworking of my design it would need. If the layout and especially the clearances are very similar to the Framework board, maybe just the cold plate but it also could be that all the heat pipes and everything but the black heatsinks would have to change.

The main challenge is that one would probably have to buy the board first as they will hardly upload technical drawings and a step file like Framework did.

 

[nrp]

This is awesome, and we'd love to help out on it. I'll send you a DM!

 

[windv]

I love your endeavour and I am looking forward to seeing it finished!
Have you thought about anodising your heatsink?

It inspires me to finish mine, even when using a discrete GPU. I hope I can cook down up to 150w for the GPU. Those are my heatsinks that I got from a bigger device and peltier elements.

 

[TheJiral]

Thanks. All the exposed heatsink parts are anodized black, internal parts are not, because their main purpose is transferring the heat and from what I have understood, mating untreated aluminum surfaces have slightly less thermal resistance than anodized ones. Also, I actually do like the more industrial look inside.

Those are some beefy heatsinks you have there but 150W needs beefy.
I assume the challenge there will be to get the heat efficiently to the larger surface of the heatsink backside.

 

[TheJiral]

I think I found the perfect SSD for my build. The WD SN7100. It has a really decent performance, while being among the more reasonably priced but is an absolute power efficiency king. The idle power draw is mediocre at best but no matter how hard you push the drive, it doesn't go beyond 4W, in gaming it should be closer to 2W. So with the Stock board cooling plate, this should also have in such a fanless build no problems with overheating.

I got the 2TB drive for a start.

 

[TheJiral]

My PSU is a bit of a gamble and clearly does not follow the recommendations from Framework. It is pretty hard finding fanless PSUs in the necessary compact formfactor. So the best I could find was the 250W HDplex PSU, which is actually 240W on the 12V rail.

This should be theoretically enough if one ignores transients, especially as I plan to have no Wifi card and only one super efficient SSD which doesn't go beyond 4W. I think the GMKtec EVO-X2 has also only a 250W PSU, but it could be that it is rated higher for short bursts. Anyhow, I'll give it a try. It does have overcurrent protection so the worst that should happen is system shutdowns. When I finally get my hands on the system I'll check what my real load will be and if the system runs stable also under full load.

I do have a plan B though. My PSU compartment is already designed that way, that I could also install the 500W DC/DC converter from HDplex. The downside there is that I would have to use an external power brick. There is a DELL Alienware power brick that is rated for 360W and is actually rated for high enough current to supply 500W at least during transient bursts (which should be fine even according to official Framework recommendations). It's just means burning another 300 EUR

PS: If it is possible with the Framework system, I also consider setting the power limit lower than 120/140W. I do like the idea of having power optimised systems and the golden balance seems to be around 80W. But let's see.

 

[Arboreal]

Do you have room for the newer 500W HDPlex GaN PSU?
I've heard a lot of reports about the 250W GaN model, and how capable it is.

 

[TheJiral]

I did consider it but the 500W HDPlex GaN PSU is not fanless. It is also a lot larger, especially longer. In that case I could have stuck with the stock Framework semi-passive PSU. But the whole point of this build is being fanless. So I can live a lot better with an external passive power brick.

I also did consider a MeanWell passive PSU but that would have also needed an ATX DC/DC converter, would not have fitted either.

 

[Arboreal]

That's fair enough, interesting to hear your rationale.
I'd forgotten the 500W wasn't passive, and it is a bit bulkier.

I hope you get a board soon, it was good to see that Nirav from Framework posted on your thread

 

[k0n]

It should be OK with the 250W GaN... I didn't even think about the board not having the option to set the power limit lower, but yes, my Framework 13 (11th Gen) is actually disappointingly limited in the BIOS! No power limit, no fan curve :/

Fantastic work

 

[TheJiral]

I am not an expert on that matter, so I am not sure what one can do in Linux in case of the bios options being limited.
My guess would be that the CPU which is according to the Techpowerup database limited to a TDP of 55W, shouldn't be the issue. It would be more about the iGPU, but a global power limit, like in the GMKtec EVO-X2 would govern both. Let's see what I'll get with the Framework Desktop mainboard and its bios.

I am hopeful it won't take too long for the board. Might have sped up a bit in the meanwhile.