Log Monochrome 2: My custom fanless 7.5 L Strix Halo system (completed)

[TheJiral]

There is so little to find on Strix Halo Desktop implementations, even though the GMKtec EVO-X2 and the HP Z2 G1a workstation appear to be already on the market and more or less available. The latter has an impressive copper cooling system, looks from the images much more powerful than the GMKtec one. The absurd thing is that the EVO-X2 is offered at such high fantasy prices on Amazon, that its almost where one can find the HP machine, even though the latter appears much more robustly designed and comes with ECC memory.

Anyhow, for the EVO-X2 there is a new review online from TheServerHome (https://www.servethehome.com/gmktec-evo-x2-review-an-amd-ryzen-ai-max-395-powerhouse), of all places.


They also took a photo of the power brick. So that reconfirmed that it is not only merely a 230W power brick but also that the rated Ampere are also not higher than 230W would suggest (unlike for example GaN 360W power brick from Dell). So if the EVO-X2 can run on 230W, I am pretty convinced that my machine will be able to run on my HDplex PSU.


As a bonus:

Partial disassembly again. So there is the view of the PSU compartment. It is cool how well the PSU fits under the board area, staying clear of the areas below the heat sink. I'll just have to see how it will turn out wit all the cables and then maybe add some cable guide if needed.

The I/O mount is also the only big splash of colour, effectively completely hidden in the shadows, once assembled. Of all the build I am probably the biggest fan of the power button. It is a linear low profile keyboard switch (Kaihl "Ghost").

 

[TheJiral]

Next big step: Proper assembly of copper cold plate, adapter block and all the heat pipes.

The thermal paste (Polartherm X-10) application was messy, as expected, but my small printed application shape did a great job at creating a thin layer of paste in the circular outcroppings for the pipes. I needed quite a bit, just for this alone, around 10-15g of paste. Good that I got the 40g tube and for the die (and possibly for other things if I run out of Polartherm) I have another 6g of Duronaut.

I put the whole assembly back onto the heat sink for the time being but for the board assembly I have to remove it again anyway, so the heat sink side will be prepared once I really have everything for final assembly. I might clean up a bit before that though, now that I am looking at it

 

[TheJiral]

The stock cables of the HDPLEX GaN 250W are fine but as they'll be touching the backside of the heat sink and given that the interior could also be quite a hot box, the rated 80°C felt a bit tight to me. Also, the 25 cm for the 24-pin conneciton and 45 cm for the 8-pin connection take up a lot of space and block airflow to the card.

Therefore I decided to crimp the cables myself. 0.75 mm2 Heluflon-FEP for the 24-pin connector and 1.00 mm2 Heluflon-FEP for the 8-pin connector. That way, the cables are not the weak link anymore and the whole connection is designed for at least 105°C which should be fine.

Right: stock PVC cables
Left: new custom FEP cables

 

[TheJiral]

It is tape time.

Added Kapton tape (https://www.olimex.com/Products/Tools/Soldering/KAPTON-TAPE20MM/) for isolation and a tiny bit of insulation to the parts of the cold plate not needed for contact with components. (The light is really harsh on the copper surface)
Added aluminum tape and then Kapton tape to the interior of the side cover to improve the incomplete shielding (the mesh is ASA, can't be bothered to figure it out in metal.)

 

[TheJiral]

Framework Desktop reviews are out, after the media embargo had been lifted earlier today: https://community.frame.work/t/desktop-reviews/73250/15
Sadly it appears as if the bios does not have any power options.

However there may be means and ways. Less intrusive tools, which may or may not work for the aim of reducing APU power limits. I was told by lhl (https://github.com/lhl) however that there is ryzenadj (https://github.com/FlyGoat/RyzenAdj) which suports Strix Halo. Looks like a pretty invasive option that should be used only with proper care. However if I understand it correctly it should enable pretty far reaching reduction of power limits at wattage level resolution. It could also enable a lower CPU thermal limit. So an option could be to reduce that from 95°C or whatever it is to 90°C or 85°C where it would be a bit healther to run under thermal throttling.

Lots of options, if that really works. But let's see if I need it.

 

[TheJiral]

I have discovered a fairly old system, which is still being sold, with a somewhat up to date entry level GPU (RTX 4060 desktop version) but a pretty dated CPU selection (it appears to be rather aimed at industrial applications which might explain less of a need to be cutting edge on the CPU side): The Airtop3. Fascinating how it has some similarities to my project. It has the same volume (7.5L) and also vertical heat sinks along the sides, oh and it costs around 3000 USD, which ehem is not all to different from my bill so far. Which probably explains why this thing is not a high quantity product. It differs in other aspects quite a lot though.

Sadly I could not find any critical review of its cooling capabilities. They make the bold claim of it being able to cool 300W TDP, ie 150W on each side, one for the CPU the other for the dGPU, which have a TDP of 125 and 115, respectively. Now, their fairly slim flat chimney design in combination with a lot of flat heat pipes, look fairly capable for their dimensions. I estimated the heat sink surface roughly and indeed it is slightly higher, but not by much. But I have to manage "only"
120W TDP.

GPU performance should be fairly similar between 8060s and 4060 desktop. The Strix Halo CPU however should be 2-3 times faster, in single core and multicore.

https://fit-iot.com/web/products/airtop3/

 

[NinoPecorino]

I remember the Airtop3 getting a bunch of reviews. Someone must have done a review of it's cooling claims.

 

[TheJiral]

Indeed I found two reviews from Phoronix (not much else though). The one did not give me the info I was looking for but the other really showed long full CPU load that appeared to fully saturate the heat sink.

CompuLab Turns An 8-Core/16-Thread Xeon, 64GB RAM, NVIDIA Quadro RTX 4000 Into Fan-Less Computer - Phoronix

www.phoronix.com

CompuLab's Airtop 3 Is The Most Powerful Fan-Less Computer We've Tested Yet - Phoronix

www.phoronix.com

It does appear that 60-140W load (averaging a bit above 100W) did not push the sustained CPU temperature beyond 80°C. There were short bursts to Tjmax but that was also at power bursts beyond 140W. That means a single side of the Airtop3 heat sink can manage the CPU load. I am fairly confident that my heat sink should have at the least an equal heat dissipation capacity and then some. Heat pipe capacity is a question but theoretically it should be sufficient in my case.

Overall I find that a good sign and I also figured out an easy fairly safe way to limit peak power draw of the CPU, by reducing the preferred max frequency via cpupower command, which can be made persistent via cpupower.service. The nice thing about that should be that it is really only cutting back the CPU a bit without effecting the GPU power budget. In case it is harder to get the heat dissipated from the tiny CCD compared to the fairly large iGPÜ.

 

[TheJiral]

Finished!
I made some small modifications to my design but openSUSE Tumbleweed is up and running.
This little thing is crazy. I installed the entire OS without the fan turning a single time, other than for a few seconds during boot.

Jk of course
I forced myself to do a short function test before I'll get serious and unmount the stock heat sink.

 

[TheJiral]

Ok, maybe I stretched my luck too thin after all. Challenges with the cooling system aside. At regular use without too much stress things did work until I chose to restart openSuse. Since then it does not boot anymore. I might have tightened the springs too much of the cooling plate, and wrecked the APU. But then, I would have expected it to fail suddenly during operation, not as a consequence of a restart.

Behaviour is strange at boot. I connected the indicator lights to the 12V SATA rail, I'll keep this disconnected next time (in case that causes problems suddenly) but it did show me that SATA power turned on (so did keyboard lights) for maybe 10sec and then they went off and so did the keyboard lights. Some indicator lights are always on, even if the PSU is at standby. But after a much longer while, maybe a minute there is a blinking for a short while. So something isn't right.

I'll check tomorrow if there is something to rescue. Stock PSU cables instead of the own crimped ones, other PSU with higher Wattage, reformat the SSD.

I did not manage to get into Bios either, apparently it is F2 for Framework but thats for Laptops, maybe I don't know the right key. Anyhow with an empty SSD it should boot into the USB stick, if it boots.

If I am lucky, that boot failure is due to some change I did in cpupower.service which is started by the bootloader. Might have wrecked the boot process but that bootloader and the settings should be gone if the SSD gets reformated.

... Well, I knew the risks. Let's see how it goes from here.

 

[TheJiral]

Small update. To get to the SSD I had to disassemble the whole damn thing again. But I had to do so anyway as something was clearly not ok with the APU/cold plate interface. This was confirmed. No risk, of cracking. My installed springs are so soft that the large area of thermal pads were just mildly dented. They were surprisingly stiff. So while the bottom (the iGPU/IO die) did contact the cold plate somewhat the top with the CPU dies probably did not at all. That explained the really poor heat transfer with not even the heat pipes getting properly warm.

Luckily the boot failure was indeed likely a corrupted bootloader that possibly killed the booting before even getting to a boot menu. Removing the SSD let the thing boot again. So I'll reformat and then I'll test without assembling everything right away. I'm done with optimism

 

[TheJiral]

Build complete!

I left the the sidecover open for the moment so that I can have a look at temperatures and I decided against including the front light. I think it looks better as monolithic strictly no-RGB block.

 

[TheJiral]

Deviating from the given specs the Framework Desktop Strix Halo is not running at continuous 120W TDP but at 100W TDP. This is very good for me indeed as it is way better balanced in terms of efficiency and a bit more manageable. One has to consider as well, that the memory and VRM also create heat. From what I could find it might be up 10W for the memory and possibly also for the VRM. Than we would be at the 120W.

I let the game Eriksholm run for 2h and I should have gotten at least close to saturation. The heat sink reached 65°C in its core and around 60°C at the tip of the fins. The APU was straight at 80°C. 65-25°C= 40°C deltaT. 40K/120W=0.33 K/W
That would be slightly worse than the 0.23-0.28 I was estimating but overall still good enough.

 

[TheJiral]

Here are some benchmarks of my System.

I found a special condition where the APU heats up beyond 80°C. Maybe that happens at mix loads of CPU and iGPU. Anyhow in game at Cyberpunk it was similar to Eriksholm but when entering the game menu overlay of Cyberpunkt temperatures of the CPU die immediately jumped up to 88°C. Tjmax of the 395 is 100°C, thanks to its mobile heritage. When leaving the menu it immediately jumped back to around 78°C.

I haven't been using FSR so far but when comparing 1440p with and without FSR2 quality, I have to say that I could not identify any visible artifacts and the FPS were flat out increased by 50%.

All in all the system is running within expectations I have to say.

Monochrome (fanless Framework Desktop AI Max+395 64GB)
OS: OpenSuse Tumbleweed with alternative scx_lavd scheduler
(Cyberpunk (GOG) was run via Lutris, stock settings)

Benchmark​	Settings​	​	Average result​
Unigine Heaven​	1440p​	​	118​
​	​	​	​
CacheBench​	Read​	​	19582​
​	Write​	​	146814​
​	Read/Modify/Write​	​	157176​
​	​	​	​
Geekbench 6.4.0​	Single-Core Score​	​	3116​
​	Multi-Core Score​	​	20785​
​	​	​	​
Cyberpunk 2077 Liberty​	1440p Ultra but FSR off​	Average FPS​	39​
(in-built game benchmark)​	​	Min FPS​	32​
​	1440p Ultra preset (FSR Quality)​	Average FPS​	60​
​	​	Min FPS​	51​
​	1440p Default (RT Ultra preset, FSR Quality)​	Average FPS​	23​
​	​	Min FPS​	18​
​	1080p Ultra but FSR off​	Average FPS​	77​
​	​	Min FPS​	63​
​	1080p Ultra preset (FSR Quality)​	Average FPS​	98​
​	​	Min FPS​	78​
​	1080p Default (RT Ultra preset, FSR Quality)​	Average FPS​	31​
​	​	Min FPS​	24​

 

[TheJiral]

Sidecover is now on, build is complete. I thermal stress tested it under full iGPU load (100W) again for 2h. Results are similar but might 2-3°C higher on the APU. Heat sink looks similar. It does noticeably move air when holding a hand above it, nothing like with a fan of course, rather like a hot radiator... which it happens to be.

The second image shows a different perspective from before starting the high load, during light desktop use with some video streaming.

And as a bonus, an image from before closing the side. My initial plan was to reuse the springs from the stock heat sink. Oh, why does Framework have to construct even that thing so repair friendly and with captive screws? I did not manage to release the screws and with it the springs but I still had some other springs in stock. Seem to do the trick.

 

[TheJiral]

Last but not least. Some images in operation (I found I actually prefer the look without the indicator light, staying fully no RGB) so you have to believe me .

 

[nrp]

Looking great! Note that 100W is with the default “Balanced” power profile. The “Best Performance” setting will take you to 120W sustained.

 

[TheJiral]

Looking great! Note that 100W is with the default “Balanced” power profile. The “Best Performance” setting will take you to 120W sustained.

Thanks!

Alex on the Framework reddit suggested to test also the Performance mode at 140W/120W, so I did.
Testing thermals is a tedious thing on such a passive system with 4.5 kg of heatsink So I did "only" test for 2h at which point it seemed to stabilise.

When starting from idle temperatures it did not thermothrottle even at Performance, still pulling 120W after 2h ... but only barely, stabilising at 98.8°C after those 2h. With the heatsink reaching crispy 70-74°C. That is too nerve racking for me

So to be honest, I think my system is really made for those 100W TDP. There it is rock solid even under mixed heavy load.


EDIT:

140W (10 min)​	82​	°C CPU4C​	50​	°C Heatsink thermal camera​
140W 10 min +120W 20 min​	91​	°C CPU4C​	63​	°C Heatsink thermal camera​
140W 10 min +120W 35 min​	95​	°C CPU4C​	69​	°C Heatsink thermal camera​
140W 10 min +120W 50 min​	97​	°C CPU4C​	72​	°C Heatsink thermal camera​
140W 10 min +120W 80 min​	99​	°C CPU4C​	74​	°C Heatsink thermal camera​

 

[miskol]

If you want to play more with overclocking/underclocking
try https://amdaputuningutility.com/
or on linux https://github.com/FlyGoat/RyzenAdj
You can change cpu Curve optimizer on all cores to -50 that will change cpu voltage -50 mV som it will be cooler

 

[TheJiral]

To be honest, the CPU is plenty performant for me, but I'd be really interested in undervolting the iGPU. I'll check if this is even possible on the Strix Halo in Linux, like it is with dGPUs.

You are right though, that saving power on the CPU side would be interesting, too. But that would require undervolt at unchanged frequency settings.