-
Save 15% on ALL SFF Network merch, until Dec 31st! Use code SFF2024 at checkout. Click here!
Log *complete* S4MAX'23 / World's smallest 4090 build: Brickless 5l S4Mini, 4090FE, 7950X3D, 800W, water cooled
- Thread starter petricor
- Start date
We really need to mmake an adapter board for the SuperMicro PWS804-r......
I thought about it, even found the name and part number... just need to learn a little bit of kicad to make it happen.
Molex EXTreme LPHPower Receptacle Header (both available through mouser and digikey)
Horizontal: 45984-4343
Vertical: 46114-4241
This is what I did
It adds a significant amount of extra length onto the end of the power supply so might not work in every circumstance.
Nice one. I believe that that's as small as it gets without soldering onto the PSU directly... have you measured how far it projects beyond the front edge of the PCB? Pretty sure it wouldn't fit my build, but it looks like something I may want to plan around in the future!This is what I did
It adds a significant amount of extra length onto the end of the power supply so might not work in every circumstance.
I can't measure it right now but from what I can guess, it is about 10mm past the end of the psu pcb. With a minor redesign, it could probably be made to accept all the wires on the psu side of the connector. The 12 volt 12ga wire for the gpu connector is partially behind the psu side connector. It could probably be made to be around 6 or 7 mm tall then.
All things must come to an end - and here that means some snags, tweaking some settings, and a few thoughts on performance:
A few weeks into running the build, I can confirm that everything runs (surprisingly) stable - clearly, it has one giant thermal bottleneck which is the 140mm rad - with that said, both GPU and CPU are designed to work in constant re-evaluation of their thermal environment and testing the thermal limits through constant frequency adjustments anyways, and the system is equalizing nicely in the mid-fifties for water and case temps (yep, it's designed to run "hot"), with GPU and CPU surfing just below thermal limits when under load.
Whilst these latest gen GPUs and CPUs have much less scope for manual overclocking or undervolting, they appear to also very well adapt to their thermal and max out performance with whatever cooling it available.
The benchmark results I get are pretty well-comparable to typical builds whilst slightly on the below-average side of things, so the cooling solution seems to be saturated - in a way that's what I have been after for this SFF build as radiators come at an immense space tax.
Overall, it seams to stabilize and handle the "hot" environment much better than my previous 5950x/ 3090 combo: I have not had a single thermal shutdown, and that's torturing it for hours with ultra settings in MSFS2020 VR until my headset batteries were drained. I had crazy ram temperatures in excess of 105C in my 3090 setup whereas here everything is high but within safe bounds here, and things appear to be well balanced even at simultaneous high CPU and GPU loads.
But more on that later - first to some final touches on the build:
First, I have treated myself to a new Noctua NF-A14 Industrial: The one you see in the build went through numerous iterations and turned into a bit of a skeleton along the way.
With a new one I expect to get a bit more pull and a tighter fit with the radiator - and potentially a bit less noise as the various cuts on the current one all contribute to turbulences.
And this is what will have to give on the new one:
A long slot in the corner for my "bleed air" intake from the PSU, next to it a pocket for the overlap with the 4090's PCB…
…and a pocket on the long side as a conduit for the AC power cable.
About one battery charge of my Dremel later…
I look a this - the minimum viable modification to the NF-A14 with the pocket for the 4090 in the background, and the space for the AC cable in the foreground. You may also note that I have sleeved the coloured fan wires running along the frame for improved blackness. Not entirely sure why Noctua doesn’t do it by default as their factory sleeving does fit nicely in the conduits of the fan.
Tha't how my bleed air port now looks like - a bit smaller than the pervious iteration and precisely matching the PSU's air outlet.
Next, I prepare the edges of the radiator with some super sticky double sided silicone tape (its turning into my universal favourite for fixing anything!) to get things reasonably air tight between fan and radiator…
…and in goes the fan and a cover plate to form a duct between PSU air outlet and the bleed air port of the NF-A14.
Nice!
Another thing I really wanted to add this time is a dust filter:
When taking my old build apart, I noted quite some dust on the radiator's fins- and they are particularly hard to clean.
And as these are the most critical parts to keep clean for good heat dissipation…
…I intend to protect them as good as possible: This generic carbon filter fleece for cooker hoods should do the trick: Not only is it catching a good spectrum of particle sizes whilst having a reasonably low flow resistance, it also comes in a very satisfying shade of pitch black!
This piece of 130x140mm should nicely tuck under the flanges on the frame of the radiator…
…and yes, that's a fit:
Should do the trick!
Last finishing touch is applying some sleeving to my 12AWG power cables:
Whilst they are already reasonably black, they kinda stand out in what has turned out to be a surprisingly clean build given the density of stuff within.
This should be reasonably presentable - I failed to get both wires into one sleeve- the bullets are just a tiny tad too wide…
And ta-daa: Build complete!
Performance
Reading a bit into the behaviour of the 7950X3D, I quicky realise that undervolting it like the 5950X is not really a worthwhile pursuit - the way AMD has made the 3D cache work by is pretty much undervolting the chip automatically, and that way more efficient than I could to it. This also shows in the X3D's significantly lower TDP compared to the 7950X.
So I settle for the full complement of AMD's out-of-the-box tuning kit for the CPU: Expo profile on (my 6000Mhz CL32 ram should pay off here), Precision Boost Overdrive on, ASUS's voltage profile applied, slightly elevating the temperature target to 90C - and that's it!
So what I'm left with for performance tweaking is the GPU: And when I say performance tweaking, it's attempting to drop the power consumption in order to raise performance, as in a thermally bottlenecked system, less load on the cooling solution should increase the frequency boost headroom.
Like with the 7950X3D, the 4090 appears to do much more automatic and on-the-fly optimization for frequency boost targets compared to its predecessor, so manual voltage curves appear to be of little effect, whereas the global power limit would be the more efficient (and also much simpler) way to adjust its behaviour.
I started out with a PPL score of 977k - and going through a number of iterations, I indeed see the GPU performance go up when setting a GPU power limit of <100%.
The sweet spot turns out to be at 85% - see my test table below, benchmarking both CPU and GPU performance with varying settings:
And finally, with a 85% GPU power limit and ASUS' CPU voltage profile applied, plus a CB23-biased boost profile, I get to a PPL score of 1.07M!
Time to post my score… thanks everyone for following here!
A few weeks into running the build, I can confirm that everything runs (surprisingly) stable - clearly, it has one giant thermal bottleneck which is the 140mm rad - with that said, both GPU and CPU are designed to work in constant re-evaluation of their thermal environment and testing the thermal limits through constant frequency adjustments anyways, and the system is equalizing nicely in the mid-fifties for water and case temps (yep, it's designed to run "hot"), with GPU and CPU surfing just below thermal limits when under load.
Whilst these latest gen GPUs and CPUs have much less scope for manual overclocking or undervolting, they appear to also very well adapt to their thermal and max out performance with whatever cooling it available.
The benchmark results I get are pretty well-comparable to typical builds whilst slightly on the below-average side of things, so the cooling solution seems to be saturated - in a way that's what I have been after for this SFF build as radiators come at an immense space tax.
Overall, it seams to stabilize and handle the "hot" environment much better than my previous 5950x/ 3090 combo: I have not had a single thermal shutdown, and that's torturing it for hours with ultra settings in MSFS2020 VR until my headset batteries were drained. I had crazy ram temperatures in excess of 105C in my 3090 setup whereas here everything is high but within safe bounds here, and things appear to be well balanced even at simultaneous high CPU and GPU loads.
But more on that later - first to some final touches on the build:
First, I have treated myself to a new Noctua NF-A14 Industrial: The one you see in the build went through numerous iterations and turned into a bit of a skeleton along the way.
With a new one I expect to get a bit more pull and a tighter fit with the radiator - and potentially a bit less noise as the various cuts on the current one all contribute to turbulences.
And this is what will have to give on the new one:
A long slot in the corner for my "bleed air" intake from the PSU, next to it a pocket for the overlap with the 4090's PCB…
…and a pocket on the long side as a conduit for the AC power cable.
About one battery charge of my Dremel later…
I look a this - the minimum viable modification to the NF-A14 with the pocket for the 4090 in the background, and the space for the AC cable in the foreground. You may also note that I have sleeved the coloured fan wires running along the frame for improved blackness. Not entirely sure why Noctua doesn’t do it by default as their factory sleeving does fit nicely in the conduits of the fan.
Tha't how my bleed air port now looks like - a bit smaller than the pervious iteration and precisely matching the PSU's air outlet.
Next, I prepare the edges of the radiator with some super sticky double sided silicone tape (its turning into my universal favourite for fixing anything!) to get things reasonably air tight between fan and radiator…
…and in goes the fan and a cover plate to form a duct between PSU air outlet and the bleed air port of the NF-A14.
Nice!
Another thing I really wanted to add this time is a dust filter:
When taking my old build apart, I noted quite some dust on the radiator's fins- and they are particularly hard to clean.
And as these are the most critical parts to keep clean for good heat dissipation…
…I intend to protect them as good as possible: This generic carbon filter fleece for cooker hoods should do the trick: Not only is it catching a good spectrum of particle sizes whilst having a reasonably low flow resistance, it also comes in a very satisfying shade of pitch black!
This piece of 130x140mm should nicely tuck under the flanges on the frame of the radiator…
…and yes, that's a fit:
Should do the trick!
Last finishing touch is applying some sleeving to my 12AWG power cables:
Whilst they are already reasonably black, they kinda stand out in what has turned out to be a surprisingly clean build given the density of stuff within.
This should be reasonably presentable - I failed to get both wires into one sleeve- the bullets are just a tiny tad too wide…
And ta-daa: Build complete!
Performance
Reading a bit into the behaviour of the 7950X3D, I quicky realise that undervolting it like the 5950X is not really a worthwhile pursuit - the way AMD has made the 3D cache work by is pretty much undervolting the chip automatically, and that way more efficient than I could to it. This also shows in the X3D's significantly lower TDP compared to the 7950X.
So I settle for the full complement of AMD's out-of-the-box tuning kit for the CPU: Expo profile on (my 6000Mhz CL32 ram should pay off here), Precision Boost Overdrive on, ASUS's voltage profile applied, slightly elevating the temperature target to 90C - and that's it!
So what I'm left with for performance tweaking is the GPU: And when I say performance tweaking, it's attempting to drop the power consumption in order to raise performance, as in a thermally bottlenecked system, less load on the cooling solution should increase the frequency boost headroom.
Like with the 7950X3D, the 4090 appears to do much more automatic and on-the-fly optimization for frequency boost targets compared to its predecessor, so manual voltage curves appear to be of little effect, whereas the global power limit would be the more efficient (and also much simpler) way to adjust its behaviour.
I started out with a PPL score of 977k - and going through a number of iterations, I indeed see the GPU performance go up when setting a GPU power limit of <100%.
The sweet spot turns out to be at 85% - see my test table below, benchmarking both CPU and GPU performance with varying settings:
And finally, with a 85% GPU power limit and ASUS' CPU voltage profile applied, plus a CB23-biased boost profile, I get to a PPL score of 1.07M!
Time to post my score… thanks everyone for following here!
Last edited:
I really don't want to be "this guy" given how awesome this build is, but doesn't the front panel (heatsink) add some amount of volume?
Well-spotted - and a fair point. Taking the dimensions, it adds 0.26l to the bounding box so with that bezel it's 5.26l- gets the high score in my table above to 933k.
In this configuration I suspect the bezel only marginally contributes to the cooling though. I have originally designed the bezel for my first S4M build with two passive PSUs, where the PSU's PCBs have been resin-immersed and bonded to the case for thermal conductivity, and that's where it really matters- the Supermicro PSU is actively air cooled and pretty isolated from the case. I'll run a few benchmarks with the vanilla NFC bezel and see what happens - I'm curious as to what the effect is...
It may show in the 30 min test, but Interestingly, and that's perhaps something to consider for future PPL rounds, none of the performance tests stress GPU and CPU simultaneously (well, technically time spy does, but it's not pushing any RTX or tensor cores, so the GPU may be at full CUDA load but still has other areas dormant, and it's far from saturating the CPU). So the PPL test is not seeing the machine thermally maxed-out.
What really pushes the machine to its limits is MSFS and Cyperpunk 2077... it's what I optimise my fan curves for.
Also, full disclosure, in daily use I run it horizontally as I simply prefer the way this looks, and that of course adds feet to the equation: It would choke without them in horizontal orientation. When benchmarking, I put it vertically on its short side (radiator side) where no air goes in or out for it to run without the feet.
I don't normally go for the plastic sleeving on the cables; I tend to either prefer none or cloth, however those look really nice. I also think those gpu power wires look nicer individually run rather than together in one sleeve. As for performance, it looks pretty good as is, however I think there is still some room for tweaking, depending on how much additional time you want to spend on it. The 4090 definitely responds fairly well to a ram overclock and is fairly easy to do. Mine is fairly bad and only goes to about +1400 but I have heard of plenty that go much higher. This tends not to generate much in the way of additional heat but can pick up a percent or two here or there. System ram also benefits from some manual tweaking beyond just the expo profile while not impacting heat generation particularly much.
Actually - checking my own test logs, I have been homing in on 80%, not 85% for the GPU- interestingly, this gives me higher boost clocks.
Bottom line, I'm pretty certain that it draws a little less than my previous build:
Nominally, the 4090FE pulls 100W more than the 3090FE (450 vs 350W) - but real world testing at techpowerup suggest its only a 75W difference (436 vs 361W) - take 80% of that, and you get to 349W for the 4090FE, which is 12W less than the IRL values for the 3090FE.
Similar picture for the CPU: Whilst nominally, the 7950X3D pulls more than the 5950X (105 vs 120W), IRL tests by techpowerup suggest the opposite: 99W for the 7950X3D vs 168W for the 5950X, that's a -69W delta.
So, all in all, I believe to be running around 55W lighter compared to my previous build - and it kinda shows in better PSU temps.
Of course, that's ignoring the rest of the system, but differences should be marginal.
I think I still have some off-the-wall readings from my old build somewhere and may compare them to the new setup at some point.
I was finally able to measure the height of the connector off the end of my power supply
It is approximately 13.5 mm past the end of the metal casing of the power supply if you include the cables and zip tie. Approximately 9 mm for just the pcb and the pins that stick through from the connector itself.
It is approximately 13.5 mm past the end of the metal casing of the power supply if you include the cables and zip tie. Approximately 9 mm for just the pcb and the pins that stick through from the connector itself.
A brief update on performance - after retraining the memory (I had skipped this in the bios to economise on boot times...) I am now beyond the 1M mark in the PPL round 4 score...
with the vanilla bezel coming with the case. Not as pretty as my own design, but I consider this a track suit for particular purposes:
And just as I have completed this, I came across an interesting guide on undervolting the 7950X3D - particularly promising as potentially helping with temperatures which is of particularly interest in my case - I'll report back once I had time to test this! I'd like to think I can still squeeze out a little more...
with the vanilla bezel coming with the case. Not as pretty as my own design, but I consider this a track suit for particular purposes:
And just as I have completed this, I came across an interesting guide on undervolting the 7950X3D - particularly promising as potentially helping with temperatures which is of particularly interest in my case - I'll report back once I had time to test this! I'd like to think I can still squeeze out a little more...
The disassembly of this RTX 4090 really made me think this GPU card is actually not that big and this whole "using three slots because it holds so much power" is just plain stupid marketing ... a bit of the same thing goes for the mobo blocks over lower rads. All this wasted space reducing air flow!
...in comparison to the original EK Annihilator...
Hi
2 questions:
1. Where did you get such a detailed 3D model of the new EK AM5 block? Did you make it yourself? If yes, can you please share it?
2. Have you considered this one (also from EK) instead? It is 8mm narrower and 1.8mm lower. The mounting mechanism is pretty much the same, so should be easily adaptable.
Here is the comparison of the EK CPU block drawings for AM5 (left) and LGA 4677 (right):
EDIT: also here is a trick from der8bauer/Thermal Grizzly that can save about 1mm of the whole package height: YouTube
Last edited:
Was following your build through the year, but back end of 23 was crazy so missed your posts about finishing the build. Really innovative use of space, and looks gorgeous. Raw understated power.
Performance is also looking really good. It feels like there is a real difference (results and mindset) when you tune for max performance (overclocking) vs efficiency, and for me the performance you have achieved in a fixed thermal envelope is more impressive and useful than outright speed. Clearly depends on your use case and objectives, but your results are great.
Not sure why you bothered with a dust filter, is there any space left in that build for even a spec of dust??? :-)
Performance is also looking really good. It feels like there is a real difference (results and mindset) when you tune for max performance (overclocking) vs efficiency, and for me the performance you have achieved in a fixed thermal envelope is more impressive and useful than outright speed. Clearly depends on your use case and objectives, but your results are great.
Not sure why you bothered with a dust filter, is there any space left in that build for even a spec of dust??? :-)
Similar threads
