S4MAX: Brickless S4M w/ 3090 FE and R9 5950x - 800W, 5l, water cooled

petricor

Airflow Optimizer
Original poster
Bronze Supporter
May 12, 2018
265
1,396
@petricor And what's the tool btw?
It's a Dremel 8200 with Dremel 220 workstation (that's a stand, really) - incredible value for money. Next time i'll probably go for a wired one - since I have the stand, I'm not really making much use of the "portable" feature, and I run into the odd charging break every now and then...
 
  • Like
Reactions: Fitchew

petricor

Airflow Optimizer
Original poster
Bronze Supporter
May 12, 2018
265
1,396
I was also surprised to see such milling was doable with what looks like a Dremel (4000?) and a Dremel Workstation!
I guess that such a static tool must imply to manually move the piece you want to machine, thus the use of some guide and doing one direction at a time. Right?
It takes a bit of of practise - and it won't win you a beauty contest, so a good old sanding block is indispensible for post finishing. Stability of tool and stand directly correlate to the price point - again, great value for money, but nothing a professional would use for milling.
 
  • Like
Reactions: BaK

petricor

Airflow Optimizer
Original poster
Bronze Supporter
May 12, 2018
265
1,396
I'm guessing that guiding is what the clamped-down aluminium bar in the pics is for? Either way, pretty impressive to do these kinds of mods with such "lightweight" equipment.
Yes - well-spotted. There is a reasonable degree of tolerance in the stand though that will get the milling head to move under pressure - patience is of the essence for a straight cut.
I'm a spatially constrained urban dweller (that's how I got in to SFFing in the first place!), and my workshop is really a bunch of pop up tools spontaneously appearing in my flat - with some dedicated space becoming available in the near future I'm eyeballing things like a Pocket NC...
 

petricor

Airflow Optimizer
Original poster
Bronze Supporter
May 12, 2018
265
1,396
@petricor

Can I ask you who your 'preferred supplier' is for your custom CNC Alu parts? Got a few things I would like making myself but hard to know who to use from the AE suppliers.
Hi Ditlan, that's QC Mold ("Qichuang Mold") of Shenzhen. They have excellent engineering support and turn things around swiftly. Shipping from China, of course, takes a while.
 
  • Like
Reactions: NinoPecorino

petricor

Airflow Optimizer
Original poster
Bronze Supporter
May 12, 2018
265
1,396
Time to return to the desk for some updates - I'll have to break this down into a few instalments as a lot has happened!

So, first chapter is cooling: Despite commonly accepted rules-of-tumb, my 140mmx25mm radiator has been able to cope with a 3950x + 2080ti to north-of-stock performance (involving some dark science like undervolting), and that has been using a single Prolimatech Vortex UltraSleek Fan with merely 0.9 mm H₂O equivalent pressure and 166 m³/min airflow at full blast. Also, it had about a quarter of the fan's cross-section covered by the protruding full-length card.

I did some calculations for the initial 2080ti setup, correlating manufacturer data on fan and radiator efficiency: They seem to indicate that most water-cooled builds have overspecified cooling, and that OTS rules of thumb for radiator and pipe sizing are not meaningful in the SFF space when airflow is well-considered.

Re-iterating over them, I am pretty confident that my new setup will actually run cooler and below thermal limits:
The short PCB of the FE 30x0 cards allows to avoid an overlap of the full GPU volume and the fan, which enables me to have almost the entire cross-section of my 140mm radiator clear for bottom-up airflow, and it also means that I can use a full-height 25mm fan rather than the ultra low profile ones I had to resort to previously.


And this is where the mighty Noctua A14 industrialPPC-3000 comes in: Performance-wise it is an absolute beast. It pushes 10.2 mm H₂0 at 269.3m3/h airflow - this is an order of magnitude more pressure compared to my previous setup which you see to the right of it. And as the S4 Mini's case airflow relies on elongated holes making up perhaps 40% of the case surface over the fans, pressure is what matters, really.

On the other side of the equation, we have thermal output:
My previous setup was around 424W nominal TDP: 145W for the 3950x, plus 279W for the 2080ti. My new puppy will blast out around 488W: 141W from the 5950x, 347W from the 3090. Leveraging the 3090's undervolting potential, I should be able to get it down to 297W, resulting in a total load of around 438W.

Of course, nominal TDP is controversial in terms of absolute output - but assuming that it is advertised to consistent standards between generations (3950x->5950x; 2080ti->3090), it should still give sufficient indication of relative change, and on that basis, I expect an uplift of only 3.3% in heat emission compared to what I had before.
And in the light of a better flow cross-section and the massive gain in fan performance I am introducing now, the new design should actually run cooler than the old one!

We shall see.

The Noctua fits nicely and reasonably air-tight onto the radiator - and has exactly the height permissible to fit inside the case:

On the right side of the fan you see the S4M's side panel still screwed to the radiator, and it sits pretty much flush to the upper edge of the fan. Space well used!

Inserting the cooling unit (that's fan, radiator, pump) into the case...

...shows how the shorter PCB footprint of the 3090 FE (compared to the full-length 2080ti used previously) enables the taller fan: There is a minimal overlap between GPU and Fan which can be dealt with by taking out material as sketched above, resulting in fan and GPU sit flush under the case's cover.


Test-fitting my ADT-Link HDMI and Displayport cables, I should be able to fit one each hassle free, and potentially even a second DP by cutting out the corner of the fan - that's provided I find enough space for it at the cases' rear. As DP 1.4 (note the shielded cables) should allow me to daisy-chain anything I could ever think of to a single port, I do not really need more than one each.

The next mod is required to route the AC feed past the fan: As there is no space left between fan and side panel , I will need to route the cable over the fan plane near the tangent point of fan circle and side panel.



I suppose this makes it a bit clearer:

Resolving this requires milling in two notches into the fan's frame to bring the cable up and then down again.

With that planned, I can go ahead and start milling the fan's frame to fit:


Ten minutes later, the fan looks like this...

...featuring a pocket for the connectors and a potential second display port extension...


...and the notches for the AC cable bypass.

After test fitting the fan, I decide to remove another corner to make more room for pulling air into the PSU...

...which will sit pretty much where the AC cable ends in the picture above. The Intake slots of my custom bezel are placed infront of the fan, and removing the fan's corner should allow more air to pass though towards the PSU’s pull fan.


This should do the trick!

Next up is the other end of the GPU port extenders.
I have been buying a bunch of perforated PCI slot covers off eBay which should make good raw material for my cover plates...

...and using a sharpie, I mark the position of the display port connector and fastening holes on the mesh.

After a little drilling and cutting with a Dremel (4.5mm holes), I get to something like this...

... still being on the crude end of things, but after treating it with different sanding disks...

...it starts to look civilised enough:



That's a fit. Two M4 Allen screws and some touch up with a black permanent marker around the cut edges make for a finish. Displayport good to go!

For the HDMI port, I need a slightly more complicated backplate, as I don't have enough internal clearance (will become apparent further down) to fit another port onto the S4M's PCI slot covers:

This piece here (cut out of another slot cover) is intended to fill the slot right of the PCI slots when seen from the case's rear...

...and will be mounted vertically. Repeating the process above...

...this is what the finished part looks like.

And here is where it all goes:

Two new ports in my case!

The Displayport "hangs" a bit to the left- should do the trick though.

Reason for placing the HDMI port vertically next to the power plug is pretty much everything else behind the PCI slots being taken up by pump and tubing...


...as it gets pretty busy in the case once everything is mounted.


A view from underneath shows how little space is left when the cables are routed and folded.


With the cables in position (I have installed the tube from pump to CPU to make sure the DP cable can stay clear of it), they should be good to connect to the GPU - one more box ticked!

Next is getting the 3090 watered - and in order to do that, I need to build in a U-Turn for my tubing as the CPU outlet and GPU inlet are more or less on top of each other.

As a C-fitting is voluminous and I barely have space for an 8mm soft tube between CPU and RAM, I need to re-route the problem to the periphery where I have enough space for a fitting to do the trick.

The most compact solution to this appears to be a Festo QSMY adapter with 8mm push-in fittings for tubes - using a blanking plug in one of them results in the flow reversal required...


...connecting CPU to GPU with connectors stacked on top of each other:

(that's the underside of the 3090 in case you lost orientartion!)

The counter part looks like this:


Ultimately I settle for routing the CPU outlet tube under the water block's bracket to make sure there is enough vertical clearance for the return pipe:


WIth that sorted, it's time to look at the PCIe connection: Meet my new ADT-Link male-to-male extender, this time, the RX/TX signal swap variant which should address the pinout issue frying my first 5950x along with the board.

First up is cutting it to fit.

The "ears" would clash with other components and cutting them off with a cutting disk is pretty straightforward.


This looks better...

...and fits without clashes.


With the flex cable mounted, I can move to a test fit of the 3090.


Squeezing it in, I note that one more mod is required to the fan:

I need a little pocket in the frame for the PCB to sit flush with the fan as otherwise the case would not close.

After a little milling the frame looks like this...



...and makes for a compact fit of the 3090 on top of the motherboard stack...


...which you can see here with the RAM removed. Note how the PCIe interconnection between the flex cable and the extension slots in behind the CPU - that's a pretty tight fit.

The port extensions seem to fit...


...so its time to prepare powering the 3090:

It will require a rather beefy, yet compact solution to feed >350W peak power to it, and a way to deal with that ridiculously long RGB cable.

But as I am still astonished that the 3090 survived shortening the PCIe port when I used an extension cable that effectively reversed the pinout, I start by measuring my new signal-swap extension cable for a pinout match, pin by pin - I have little appetite to push my luck again!


And: Bummer. No game.

BOTH ADT-Link male-to-male extension cable types (TX-TX and TX-RX) don't reproduce the pinout 1:1 mirrored when "unfolded". I have no idea what these cables do and what they are good for - but they are NOT working as risers.
In all fairness, they are not labelled or sold as such, but as "extensions". From what I can measure, the difference between the two types appears to be reversing the order of pins within each PCI lane - but in the grand scheme of things they are still leaving me with the "mirrored" commector problem I get when looking at two PCIe spec male connectors used for an extension, and, most importantly, they would still make my +12v motherboard power feed end up in the GPUs GND - so: DONT TRY THIS AT HOME.

The bad news: Plan A is dead - at this point I have no solution to connect GPU and board as I am not aware of a longer flex cable as thin as the HDPLEX, and as I have established previously, there are little more interconnections that a daisy chained PCIe cable could take. And any thicker cable could not fold between CPU and GPU without the stack exceeding the case's height...

The good news: I found this out before frying more components.

So this throws me back to about here:


Next episode: Plan A is dead - long live Plan A!
 
Last edited:

Valantar

Shrink Way Wielder
Jan 20, 2018
1,965
1,898
Ultimately I settle for routing the CPU outlet tube under the water block's bracket to make sure there is enough vertical clearance for the return pipe:
Me: nodding, thinking "yes, that is clearly the reasonable way forward."

Also: "Wow, how is your GPU mounted?" "Well, the water block kind of rests on the folded PCIe riser cable, the PCIe slot rests on the trimmed-down rear I/O of the motherboard, and the edge of the PCB rests on a notch I cut in the frame of the radiator's fan."

This build is completely redefining SFF. Even with the constant setbacks, this is so damn impressive. I hope I'm interpreting that Next Episode teaser as a "don't worry, it's figured out" and not a "well, I had to custom design my own riser cable, which will arrive by canoe from a boutique cable weaver on Fiji in the next 6-12 months."
 

Vaderator

Average Stuffer
Bronze Supporter
Aug 7, 2017
58
116
This may be a long shot, but maybe the hdplex gold finger insert isn’t a straight pass through. It strikes me that it should be your reference for whether the adtlink does what you need. That key is directional so it can’t just be a straight pass through.
Other than that, haven’t you already demonstrated that it works with a normal extension and the gold finger inserted. I remember it wasn’t ideal to have the extra connector, but it didn’t seem unworkable.
 
  • Like
Reactions: petricor

petricor

Airflow Optimizer
Original poster
Bronze Supporter
May 12, 2018
265
1,396
What about loque riser with many individual strands, you can bend that one any way you need?
It wouldnt fit with my current layout at the GPU side connector is much deeper than the HDPLEX - it may indeed be an option when resorting to (slow) ULP ram and sacrificing speed...
 
  • Like
Reactions: riba2233

petricor

Airflow Optimizer
Original poster
Bronze Supporter
May 12, 2018
265
1,396
This may be a long shot, but maybe the hdplex gold finger insert isn’t a straight pass through. It strikes me that it should be your reference for whether the adtlink does what you need. That key is directional so it can’t just be a straight pass through.
Other than that, haven’t you already demonstrated that it works with a normal extension and the gold finger inserted. I remember it wasn’t ideal to have the extra connector, but it didn’t seem unworkable.
It works with the HDPLEX insert, indeed, and that is a 1:1 pass through, pin by pin. And yes, both variants of the ADT Link extensions have a different pin mapping when compared to the HPDPLEX inserts.
Another interesting thing I noted when getting curious and measuring various PCIe cables is that many appear to "cluster" pins where the PCIe spec allows and interconnect them to save strands - that's mostly GND pins of which there are many.
All HDPLEX products, however, that's the long flex cable, the short riser, and the "gold finger inserts", keep each and every PIN separate and map them 1:1 from end to end.