Log Deskmini x300 "Turbine", a 12vDC brickless Vehicle build

[msystems]

A few days ago when running performance tests in the stock case, I noticed that the memory could get quite hot and needed some kind of solution. It might be acceptable at stock speeds with 1.2v but at 1.35v and clocked higher, it can make the memory unstable.

This is because Asrock has it located at the bottom of the case with absolutely no ventilation.

Spoiler: x300 modification #2

So I decided to use these SODIMM heatsinks with graphene coating.

I don't believe it dissipates much heat, but it does spread the waste heat evenly, to reduce hot spots.

Thanks to @HydrAxx747 for the tip and also @LukeD for the helpful memory information in various places.

I also drilled some ventilation holes parallel to the ram, which would be on the floor/side of the case depending on orientation.

It is also possible to fit a 40x10mm fan on the inside right here to put air directly on the memory. The x300-stx board has an extra unused fan header as well just waiting to be used. I tested it with a Noiseblocker XM2 (14dba) which ran full speed at about 3300rpm and had a noticeable result. The only downside was it could not be controlled, since it was a 3-pin fan. The Noctua NF-A4x10 is 4-pin and would be ideal here.


Back to the case, I attempted an acetone vapor bath for the ABS duct, but it was too cold outside and didn't work well. So I just ended up painting it over and then a bit of poly.

Stock faceplate went back in. I did not paint it. It would be ruined with the crude spray paint. It needs a nice enamel & small brush.

AsRock put spaces for the feet mounting on two sides. It's hard to choose which side to put them or to use them at all.

Final assembly:

The black faceplate matches the fan perhaps?

Also working great in horizontal placement:

The fitment of the duct onto the case is very good thanks to the tension force possible from the threaded inserts. It sits on the top of the NF-9 snugly and will not budge at all.


And finally here is the placement in the vehicle cabinet:

Much more spare room than I imagined was possible.

Currently running just fine through the inverter. Just waiting on the parts to modify it to run direct 12v now and get rid of that brick for efficiency.

 

[SFFMunkee]

Any chance you’d bother with a dust filter for that duct? I can see a lot of dust getting pulled in there

 

[msystems]

Ok, so to get the Deskmini to run from the vehicle directly, the input has to be rewired.

This is dangerous because instead of passing 6-7 amps through the barrel, the barrel and wires need to be able to handle as much as 10-11 amps due to the lower voltage.

Since @Curiosity actually had the DC barrel melt on feeding the x300 12v, I was extremely fcking nervous about this step and wanted to try to do the safest thing possible. @REVOCCASES suggested soldering to the board to avoid the failure point, or finding a good quality barrel. I looked at doing the soldering, but just don't have the right tools & skills to attempt it. So I just tried to find a decent looking DC barrel. Thanks to their advice in @Curiosity 's log.

The cheap DC plugs on Amazon were really only good for 3-5 amps and typically come with 20-22awg or oversized aluminum wiring inside of unrated, silicone insulation. This is guaranteed to melt.

So I ordered a nice looking DC barrel plug from China, slightly better looking quality. That finally arrived.

Spoiler: 12v Conversion

I used some 105c UL rated pure copper 16awg and soldered onto the DC plug. Since the wire is quite a bit oversized for the plug, I was unable to use the plastic screw-cap to close the end of the plug, so I just used layers of heatshrink and tape to strengthen the soldered tips against mechanical failure.

I wired the other end to an XT-60, to reduce wear on the barrel so I can just unplug the XT-60 if it needs to be unplugged, instead of the barrel, & so that the XT-60 contacts will eat any electrical arcs when plugging it in to a hot circuit, instead of the barrel tip.

The XT-60 is wired to my DC-sub panel in the cabinet which is now getting a bit crazy...

Fusing the x300 wire @ 15 amps. The monitor is on the same XT-30 connector type, so it is color coded RED and x300 BLUE so no accidental blowing of fuses.

This sub panel is now feeding 5 things- 2x USB hubs, the Monitor, a cigarette socket with on on/off switch for a cell booster, and now, the x300. The total of everything on this sub panel is about 20a. It is connected via 10ga wiring on an XT-90 to the main panel.

First boot - it runs! Ditch the brick!

For safety, did some monitoring of the DC barrel socket temperature while pulling max load, in case there was an issue with the electrical contact.

The highest temps I saw were just a bit over ambient, so I am hoping that the wiring is sound.

With the 120w brick removed from the equation, the real power of the 5700g and x300 can be unleashed...

Pulling 11.04a x 13.2v = 145.728 Watts

Assuming enough cooling, it might be possible to pick up some of the performance left on the table by the stock brick.

How much?

How about 15,000+

Stock = ~13,500

Fun experiment but it is unsustainable thermally.

What is next? I would like to keep the memory cooler so am thinking of doing something like this with 40x10mm fans.

I can already screw these in sideways, but they work better when sitting on top of the memory.

They really work awesome:

Peak temp of 32.5c during being hammered by a full system load of Heaven + Cinebench.
Rapid drop into the 20c range within seconds of stopping the tests.

So what do you think? Mod it more or stop here?

 

[SFFMunkee]

So what do you think? Mod it more or stop here?

Mate, you already know the answer to that one

 

[aromachi]

I don't think it's been mentioned, but have you thought about using a car audio capacitor to stabilize the current coming from your battery?

 

[msystems]

I don't think it's been mentioned, but have you thought about using a car audio capacitor to stabilize the current coming from your battery?

No but that's a good idea. It is an isolated battery system, but there are many things on the circuit. I am a little nervous about the fridge compressor on that same circuit causing a ripple. It has no slow start. Its not a big compressor... about 6 amps, but still. The other motor on the circuit is a DC ventilation fan, but it has an integrated slow-start so its safe.

 

[timginter]

great thread! thanks for documenting in detail!

 

[msystems]

About a month ago, I ordered an x16 to M.2 adapter for fun, after seeing the successful eGPU implementation by others on here.

The adapter finally arrived, so it was time to make an eGPU Frankenrig.

Spoiler: eGPU musings

Because it omits the 24 pin connector, the R43ML adapter is much sleeker than the commonly used R43SG.

Thanks to @REVOCCASES yet again for the good find.

ADT-Link R43ML

Getting the Frankenrig to work was actually pretty easy. There was no configuration required on the a/x300 whatsoever, it was just a matter of connecting the m.2 adapter and powering up the donor PSU with the shorting trick. Thanks to @REVOCCASES and @Curiosity again for the tips.

The Frankenrig. An x300 pancaked on top of a Streacom FC5 with semi-passive GTX 1080. A chopped-in-half and duct-taped Enermax 650w powers the GPU.

I ran a few performance tests on the Frankenrig and compared it to the baseline system (Intel 6700k, 16g Ram). Stability was tested by pushing the same 2037mhz GPU overclock.

The x300 Frankenrig is pretty strong!

In theory there should be a bigger penalty for using an eGPU of around 10-15%*, but the 6700k (Skylake) system is also on a riser.

*On Thunderbolt with external controller. Using m.2 seems to have a lower baseline penalty.

So there isn't a riser-specific performance penalty, just the PCIe x4 lane penalty. And since the 6700k is a much older processor than the 5700g, the lead shrinks a little more. [Edit- there is likely small driver penalty as well. X300 on newest drivers & 6700k kept on older ones for stability.]

Although the FPS was similar, the x300 system responded in games better. There seemed to be less input lag, better smoothness, and no micro stutter. "Minimum FPS" numbers were slightly higher compared to the Skylake system.

Imagine a custom eGPU enclosure?

It shouldn't be too hard to strap the 1080 onto the x300, but the power design has to be sorted out.

I was hoping that the HDPlex 400w DC-atx would work a bit out of spec, because it is a great piece of hardware to drive a video card. It can deliver 35 amps on 12v rail!

Might as well try it... and retire Skylake with a bang?

It didn't seem to like the 14v power supply though (simulating vehicle DC power). A bit too far out of spec.

So that means I would need a wide input 12v dc-dc regulator (decently sized to handle 200+ watts on 12v), or use a 12 to 19v step up and enclose the HDplex dc-atx internally in the DIY eGPU case. Then there is also going to have to be a new, 8-10ga wire run in the vehicle to handle 20 continuous amps with low voltage drop. And the enclosure has to be designed and printed. So I don't plan on doing an eGPU any time soon.

The 5700g is actually pretty decent once tuned...

 

[Arboreal]

That's brilliant @msystems!
I am starting a voyage of discovery in the Deskmini arena having managed to buy a well priced A300 on UK eBay last week.
The CPU is on order for v0.9, and I chose the A300 over an H110 version (which i have a 'T' CPU for...) specifically as it has 2 M.2 slots.
The second slot is calling to be repurposed as a GPU feed, and could well end up driving my modded GTX 1050 Ti as an experiment at leat.

Thanks for the inspiration and showing the proof of concept works, not just on YouTube!

Your X300 build is great, and I think there will be a lot of fun to be had with this A300.

I have a couple of thin ITX kits and a couple of NUCs, but Mini STX was definitely an itch that was needing to be scratched.
I must start at thread of my own, don't hold your breath!

 

[timginter]

In theory there should be a bigger penalty for using an x4 riser of around 10-15%, but the 6700k (Skylake) system is also on a riser.

So there isn't a riser-specific performance penalty, just the PCIe x4 lane penalty. And since the 6700k is a much older processor than the 5700g, the lead shrinks a little more.

Good performance may be due to Nvidia Optimus drivers - from what I understand, there's lossless data compression and x4 M.2 bandwidth limit is not an issue until roughly 2080 Ti levels. In general, performance loss on 4-lane M.2 may be around 5% compared to x16 PCIe

 

[Arboreal]

Good performance may be due to Nvidia Optimus drivers - from what I understand, there's lossless data compression and x4 M.2 bandwidth limit is not an issue until roughly 2080 Ti levels. In general, performance loss on 4-lane M.2 may be around 5% compared to x16 PCIe

That's really interesting, i didn't know that.
I though 8x was minimal loss and 4x took a ~20% hit...

There's hope for my potential A300 / 1050Ti plan then

 

[msystems]

There is another factor I forgot which is the difference between m.2 and TB. They are not equal, and the implementation of the TB controller matters. Also, the OS also has an impact, such that a gtx 1080 egpu over thunderbolt on Mac OS yielded horribly worse performance than windows 10 over m.2.

I was looking at this thread below in particular, & that is where my baseline 10-15٪ estimate came from: https://egpu.io/forums/mac-setup/pcie-slot-dgpu-vs-thunderbolt-3-egpu-internal-display-test/

But these results are all with thunderbolt. The worst performing set ups in that thread were on TB and a Mac.

Revoc has tested TB vs m.2, where the only variable was the connector, & using M.2 was a few percent faster, everything else being equal.

 

[Curiosity]

That's really interesting, i didn't know that.
I though 8x was minimal loss and 4x took a ~20% hit...

There's hope for my potential A300 / 1050Ti plan then

Definitely, I'm using a 6600 and havent noticed a big drop compared to review benches

 

[msystems]

I have been tinkering around with the x300 for a few weeks to try and find optimal settings, and came up with the following:

Spoiler: Tuned x300 Settings

CPU side
Unless it is hidden somewhere, the x300 bios seems to provide no way to offset the voltage of it's boost to make it more efficient.

For finer control of max power draw and temperatures, I used manual settings, turned SMT off and settled on a 4.4ghz clock @ 1.3v Vid (1.24v under load), which results in a "CPU package power" draw of 54 watts during Cinebench, with temps in the low 70C range. This consumes about 65mv less package power under load than the automatically assigned boost voltage at the same frequency, which is 1.306v under load, PP=62w. A savings of about 8 watts.

In sustained workloads it should be more efficient, and also faster since the stock boost settles @ around 4.2ghz. In light and medium (gaming) workloads, it should be about the same performance, but with less heat and power.


Gpu/Memory side
This was really an interesting challenge. It seemed to be about striking a balance between the GPU and memory controller.

I found that 2000mhz FCLK with good timings and the 2200mhz GPU @1.15v Soc was a good compromise. The Corsair bdie kit here could probably keep going, but vDimm is limited at 1.35v.

2000mhz CL17 is about as fast as the 2100mhz CL18 level, but more stable for the FCLK & GPU.

Unfortunately, the memory gets really hot... and the memory heatsinks + holes I drilled in the side just aren't good enough.

Asrock DID include a second fan header on the x300 though, if it can be used...

Spoiler: More modding required

I decided to completely revise the duct design to incorporate 2x 40mm fans for active memory cooling.

To simplify the design, instead of adding 8 new screw holes for the 40mms, I designed some "guides" they can be placed on and only need 2 screws.

This was my first foray into shortening a fan cable, and possibly the most sadistic and tedious thing I have done in pursuit of SFF. As you can probably tell, I did not have the right tools for this job and soldered them instead of crimping. It probably took more than an hour to get all these 14 tiny wires stripped and soldered.

The fans fit, but I did not have the correct screws to attach them. And the "guides" did not hold them like I hoped. So this turned into an experiment of whether marine grade adhesive sealant is any good for attaching fans.

Very messy on one side, but (mostly) clean on the other and can't see it. Some got on the fan cable, but just on inside.

Putting it together. The 40mms sit flush with the CPU fan on one side and chassis on other side (just enough room for fan cable). Clearance over the memory is 1mm.

The paint could be better, but the fitment it turned out well. The x300 outer case is starting to get a bit of metal fatigue from all the abuse.

To test the results I loaded up Testmem5 to hammer the entire memory bandwidth continuously.

With RAM fans off, I had to stop the test when it climbed over 60c.

With RAM fans 100%, it is stable at 40c, and this is the "worst case". It is about 30-35 during normal use. So the fan curve can be set nicely.

With the modding and tuning complete, this build is done.

Regarding the gtx1080 as an egpu - I forgot that I put some copper heatsinks on the memory with permanent arctic alumina adhesive. So the stock cooler can't go back on unfortunately.

With the modding and tuning done, here is an updated performance comparison:

It is generally able to run 1080p acceptably with settings turned down.

I thought to run Cyberpunk as a joke, but then realized it actually runs pretty well.

Spoiler: Cyberpunk???

Cyberpunk is surprisingly playable at 1080p Low. It is even playable on Ultra after enabling Super Resolution, which gives it a few more FPS.

Super Resolution downgrades the resolution of objects at long distances. It increases the rendering speed without losing much visual quality.

Cyberpunk Image quality comparisons:

1) 1080p Ultra Preset - 14fps

2) 1080p Ultra Preset with Super Resolution set to Ultra Quality. Objects at long distances have less detail but it is not that obvious. - 21fps

3) 1080p Low. Still looks acceptable. - 29fps

4) 1080p Low, Super Resolution "Ultra Quality", Motion Blur on "Low" - 38fps

The image is clearly deteriorated at this setting, but it's a fantastic preset.

There is not really a reason to need to drop to 720p.

 

[Elaman]

Great job with the performance comparison. I wonder if there is a way that we could visualize the performance-per-watt, being this a 12V build and all.

 

[msystems]

Great job with the performance comparison. I wonder if there is a way that we could visualize the performance-per-watt, being this a 12V build and all.

Its about 10 watts idle, 70w gaming, and screen is about 14 watts.

It can't beat a laptop, but it is much more efficient at idle than my desktop 6700k Skylake gtx1080 build, which was about 35 watts idle.

Its sort of in between a laptop and desktop.

What I might want to try next time around is a Laptop board (NUC or frame.work) paired with an undevolted desktop eGpu. This should sip power but be able to hit hard when it needs to.

The x300 doesn't have Ryzen curve optimization - so it is a little hamstrung on max efficiency.

 

[aromachi]

I'd play so much Skyrim on that thing.

 

[SFFMunkee]

The x300 doesn't have Ryzen curve optimization - so it is a little hamstrung on max efficiency.

That said, you're already saving between 5 and 15 W by having the Knoll instead of a chipset

What I might want to try next time around is a Laptop board (NUC or frame.work) paired with an undevolted desktop eGpu. This should sip power but be able to hit hard when it needs to.

If you set a lower power limit, your CPU is basically the same as the mobile Cezanne SKUs anyway, I think. So I'd try that before replacing any more gear as if buying a new board and planning to use an eGPU you'll be dependent on finding one with Thunderbolt, which isn't available on all options unfortunately.

Using an M.2 adapter opens up more possibilities, but it would basically be the same as what you have now just with a mobile chip (and lots of options have only one PCIe x4 M.2 so you lose out on NVMe).

 

[Elaman]

What I might want to try next time around is a Laptop board (NUC or frame.work)

Is it that the recent news about framework releasing a standalone laptop mainboard is giving some SFF ideas?
It's definitely very interesting: it's compact and it's modular. But yeah, no standalone laptop GPUs out there except for those weird external ASUS-only ones.

 

[SFFMunkee]

Is it that the recent news about framework releasing a standalone laptop mainboard is giving some SFF ideas?
It's definitely very interesting: it's compact and it's modular. But yeah, no standalone laptop GPUs out there except for those weird external ASUS-only ones.

There are MXM cards but good luck finding one, and especially good luck finding one at a 'reasonable' price