A lot of catching up with progress is due - this is a long one, even by my standards - TL'DR: It powers up!
What's clearly missing so far in this story is one of the key enablers of the build: the water loop, and a radiator beefy enough to cope with 4090 and 16 cores of CPU.
Whilst the 7950X3D should perform better in a heat-constrained environment than the 7950X I have been considering originally, it still comes at a higher TDP than the 5950X I have successfully been using in an S4Mini - they compare as below:
5950X - 105W
7950X - 170W
7950X3D - 120W
...so that's 15 W more compared to what I got to work in this form factor so far.
All values are nominal TDP though, which is not the full picture, but in the case of AMD probably closer to the truth than with Intel, where power draw flies off the charts when hitting their high-end CPUs hard.
The 4090, again, comes with a nominal power draw uplift compared to the previous model I have been using:
3090FE - 350W
4090FE - 450W
...that's 100W more than before for graphics, so it looks like I have to tuck away 120W somewhere in my new build.
Well, TDP is one thing, but When looking at actual consumption, however, it doesn't look that bad.
There are quite interesting tests measuring actual power draw under load, and here it would appear like the gap is a bit smaller.
The 3090FE appears to pull a bit more than the advertised TPD under load -
whilst the 4090FE appears to be more conservative than advertised:
3090FE IRL - 361.3W
4090FE IRL - 436.2W
Actual delta: 75W
Interestingly, the deviation between TDP and IRL performance gets even more extreme we look at the CPU - it appears that the 7950X3D would pull less than the 5950X despite the higher advertised TDP when
putting it onto a test bench - the gap is negative:
5950X IRL - 168W
7950X3D IRL - 99W
Actual delta: -69W
So I'd be left with a mere 6W uplift to deal with - doesn't sound too bad? Well, the thing is, full CPU and GPU loads very rarely occur simultaneously, and as my application bias is heavily towards 3D, there will be a bit more power to deal with in real-life scenarios, which has already been showing when my old power wiring got to the brink of cooking off when connecting it to the 4090 (further up in this thread).
To address this, I have two things to play with:
The 4090 appears to be taking significantly lower power targets without losing much of its punch - I may just get away with 80% which would bring my IRL power draw down to 350W so effectively giving me a system pulling 80W less than my previous iteration.
The second optimisation may be in the radiator - which gets us to the next part of the story, and a full test fit to adjust all components where necessary, including the radiator I will be using- meet the Corsair XR5 140:
In the box, we find the part itself, a bunch of screws - and that's it - well, it is what it says on the box.
And you may note that it's not what I have been using before (an Alphacool Nexxxos 140.v2) - and the reason, as someone in this forum pointed out (cannot find that post or author anymore - thanks regardless!), is a very valid one in the world of SFF: It's smaller!
Next to the Alphacool, you may already note that the protruding bulkheads at the top and bottom are a bit shorter, and more importantly, it's about 4mm less wide...
...with the size differences becoming most obvious in the top view.
Very interestingly, it does that while having the exact same clear fin stack projection in plan - it simply has a thinner frame, you can think of it as a screen with a thinner bezel and the same visible diagonal...
...and most interestingly, the fin stack is finer/denser and about 2mm higher, so with a beefy fan as I intend to use here, it, in theory, it should deliver better performance, contributing constructively to my thermal envelope challenge!
Further, you see how the G1/4 terminals are placed slightly differently: They sit a tiny tad closer to what's the upper edge on this picture, and, most importantly, the bulkhead is a bit lower when compared to the AlphaCool rad, allowing the XR5 to use standard screw ports rather than the low-profile ones (requiring special seals) in the AlphaCool. As a result, the Alphacool 40 DC-LT pump/res package sits perfectly perpendicular to the radiator, whereas with the Alphacool it would sit at a small angle as the ports are further in. Nice!
Fitting it to the cases' frame, the slightly smaller dimensions result in a perfect (formerly rather tight) fit of the radiator/pump/reservoir combo...
...and a nice little gap between the radiator and the motherboard base frame which gives me extra space for wires, and most importantly, a little more breathing room for the PSU.
Adding some 8MM OD push-in fittings to the radiator and pump/res combo...
...I identify a location for the Aquacomputer Quadro fan controller that should now fit between the pump and board...
...and continue with the DisplayPort/ HDMI extenders which are supposed to go under the board.
...and yes, it's getting busy in there - and whilst those wires are all working and should make for an initial validation, I'll ultimately need to lengthen them to fit nicely into the case.
The final step before installing the board is running the pump cable under if...
...to then place the board on top - that's a fit!
The PSU goes in next...
...with the GPU going in on top.
All wires hooked up - looks like a computer!
With everything put together, I take notes of a few adjustments I need to make to the parts:
First, the radiator wants to be secured against the outer frame of the build.
To do so, I need to drill in a hole at roughly this position here (yes - I know, drilling into a radiator is generally a bad idea, so I better get this precisely...)
...and with a 3mm carbide drill that's a pretty easy one.
Should do the trick!
Then, next, I note a clash of the radiator and case frames: Aim is for the rad to be aligned to the cases' front edge, and flat to the bottom cover panel, for maximum contact and heat transition (it's the beauty of the S4M: Everything is pretty thick aluminium and can add up to a giant heat sink): And for that to work, the parts have to interlock. That in turn means, some of the case frames' bottom flange has to give:
I first start with two notches...
...but they don't give me the full bottom contact I seek...
... so I end up recessing the flange along the entire length of the radiator:
This makes for a nice and flush fit...
...and here you see where the radiator is secured against the frame using a trimmed M3 nut inside the flange over the fin stack.
The second fixing point is a steel bracket between the motherboard frame which will ultimately be sandwiched between the fan and radiator...
...which gets me to the next mod which is trimming the frame of the main fan back a bit further for the 4090's PCB to fully dovetail and the GPU not to stick out at the top (I have recycled the fan it from my last build, the big opening you see to the bottom left is
generating a bleed air stream extracting hot air from the PSU through the main fan - this remains to be a useful feature and is the main reason for the fan coming across like a bit of a skeleton).
This should work!
The final major mod is the PCIe cable.
The LinkUp PCIe 4.0 Ultra I finally opted for has a GPU connector with a tiny piece of PCB and a latch which are both sticking out above the GPY backplane by a couple of mill (as established earlier in the thread).
Clipping off one side of the latch is the easy part...
...whereas for trimming the PCB, I need to resort to a mildly nerve-wracking manual trimming procedure using a cutting disk. All sorts of things can go wrong here: Apart from having to trim just above the solder points (I have to cut it back to about 1mm away from them), I trim off a piece of multilayer PCB with vertical trace connections (they show in the photo - it's the metal traces running normal to the PCB plane), and if some are cut off fully and non-redundant, I may lose PCI lanes - let alone creating a short in the connector by bending the traces in the cut and accidentally connecting them.
So next thing I do is install the cable in a donor PC with an older Quadro card on the other end, which should allow me to run Nvidia-smi -q for a PCIe connection check. Worst case, I fry an old piece of hardware...
But to my great relief, I get 16 PCIe lanes happily reporting back to the service management interface, so the part is alive, functional and good to go!
Finally, I finish the PCB edges with some lacquer to avoid shorts in the case.
So I should be good for the final assembly, and hopefully a power-on test!
In go the fan and pump - and as everything on the top half of the picture is in its final position now, I also go ahead and screw in the HDMI and DisplayPort extenders along with some slot covers.
Next, I can install the motherboard for good (hopefully)...
...and provisionally sort the wires under the motherboard to leave clear space for the SSD.
In goes the power-on switch...
...and I can finally mount the CPU block "for real" and thermal-paste it to the processor:
Mildy regretting that I got a bit obsessed with avoiding excess thermal paste when applying it to the GPU (I may eventually redo this - let's see how temps work)...
...I go for the geometrically more logical variant with one drop in the centre, and four drops at the 2/3rd points of the diagonal: Whilst this is likely to push out some MX-4 at the edges, geometrically, the five drops should not create any enclosed air pockets when expanding concentrically.
We'll never know precisely what's going on underneath, so here I go:
Done!
And now to some tubing and a feature I always wanted to have: Ventilation- and fill ports, allowing me to bleed air from the build whilst it's running and without having to take stuff apart. This has been a real pain in my last iterations - getting it to run air free was really hard without a reservoir proper. And this is where the T-fitting in the top-right corner comes in:
...and before putting things together for a fill, I cannot forget about replacing the Corsair RGB connector with something that talks to my Asus-5V-RGB header:
Trimming the cable and crimping on a connector is pretty straightforward...
...and an hour later or so, I am looking at my planned loop IRL - looks closed, and using the highly scientific blow-into-the-tube-and-see-whether-it-holds-pressure-method, it appears to be airtight!
I decide to do a quick power-up of all components t check on power and cabling before filling the loop:
Power delivery looking good...
...and, connecting fans, pump, and CPU, I dare to quickly power up the build for a POST, relying on the mass of the CPU block to run the machine for a few seconds and read out the board's POST LEDs, albeit without GPU and relying on the Ryzen's APU for an image...
Moment of truth:
happily POSTing through to the moment it doesn't find the SSD I have not installed... IT'S ALIVE!
And yeah - that's an image projected to the rear of a chair. I'm somewhat allergic to plastic which makes monitors difficult, and this project is intended to be run headless - my entire venture into SFF started with looking for a way to run VR on an IMac to which the answer was "PC"... but that's a different story. I may end up buying a Studio Display when this build is finished...
Oh, and did I say "it's alive"? Well, that's but for one distinct sound I'm missing: The one of an unhappy, unfilled water pump squeaking away
. It's rather dead. But then, that's what the test was for - finding stuff out before flooding the build with coolant.
Turns out that when pulling the pump cable under the board, I may have pulled a bit too strong and the wire gave in.
Bugger.
Nice twist: note the load relief rubber block dangling off the cable there - obviously designed for a larger pump enclosure, Alphacool didn't really provide an option to fit it in the 40 DC-LT pump/ res, so all cable action is transferred directly to the PCB...
Let's see whether it's salvageable:
Taking the top of the pump off and removing the glued-on cover, I see some gunk over the solder points, probably to make it a bit more robust against water ingress...
...and I decide to clean the PCB of all glue and re-solder the cable entirely and in situ as removing the pump would mean taking the entire build apart again.
I re-seal the solder points with some double-sided ultra-sticky silicone tape...
...that also works rather well to hold the cover in place.
Nice - and another power-on confirms it works. And yes, pumps without water really sound horrible and are about as bad an idea as running the CPU without coolant.
So next: Let it flow!