Fujitsu Esprimo E920 D3222-A12 12V/11Vsb to ATX conversion (#P28)
Recent hardware acquisition: Parts of an Esprimo E920 system (board, CPU, RAM and power supply) for replacing my main desktop system, a good old FX-6300 with 970 chipset board.
Why would I post about this? Well, it’s not a drop-in replacement. DIY required, since those OEM things always have some custom oddities. In this case it’s two (and a half) of them:
- Power supply has no ATX mounting bracket
- Power supply and board are not -12V/3V3/5V/5Vsb/12V units but 11Vsb/12V only
- Board has no I/O shield, just fitting case cutouts
Note that there are datasheets of this very Fujitsu board out there that show a regular ATX connector with all the standard voltages required. I would assume there’s some circuitry on these that is not present on my board, as there’s some things unpopulated close to the empty ATX mounting holes. Mine only has the custom 16 pin connector, no onboard USB port (headers only) and no open PCIe x1 slots.
Add in some issues with the case (e.g. the lack thereof) and it’s a total bargain for those that can handle the modifications. I paid the rounded up eBay value of the i5-4570 processor plus shipping, so with a power supply and the old board and CPU to sell, it’s an almost free upgrade. Plus I get to make a blog post out of it – priceless
So, here’s the battle plan: I never intended to run this thing with the original DPS-250AB-82 A power supply. Reason one being the non-standard mount, which would have been no issue with 3D printing parts, see the SFX transplant of the Supermicro X9SRW-F (#P23) blog post for reference. Reason two being the 250W power rating of that thing, which I BARELY fall short of when running games, so synthetic burn-in benches would overload that thing. It is therefore necessary to make the board compatible with the existing 450W ATX power supply. I don’t care about the I/O shield at all, so chances are I will never make a custom one. If there’s one to be found on eBay for cheap: Great. Otherwise it’ll stay that way.
In mid-2019, Intel declared a new motherboard supply standard called ATX12VO (O = only). The board pre-dates this effort by far, and, as I said, OEMs can do a lot of custom crap as they got the production volume to justify their island solution (or just want to be noncompliant with any reasonable standard, say Hi to Apple). While ATX12VO has this pinout, thankfully provided in color by some NG4hle guy on WP Commons:
(this is the view down to the female board connector), the Fujitsu uses this layout:
(rough sketch by me)
These are all Molex Mini-Fit connectors by the way, type 5557 and their mating partners like 5566. Apparently those are pretty pricey even from Chinese sellers, I wonder why. If you got a source for a nice 1-5pcs kit of every size, please let me know.
Interestingly the Fujitsu uses double the number of power pins (okay, that’s tricky to do on a 10 pin connector) and two 12V rails. Separate rails are no longer as common as they used to be in that ~2013 era, but there’s still a lot of 12V power consumed, and this has way more current capability despite being a low-power, OEM, non-overclocking machine compared to Intel’s new solution for everything. Well, both are an improvement over the current ATX 20+4 pin, but basically anything is, as it only carries 1+1 pins for 12V and the rest is legacy stuff.
The 11Vsb part however is important. I do not know if 12Vsb would be acceptable or over the limits of some components. I will also not check, as I found that anything above 7V is working just fine. Basically this is the voltage rail that supplies power to the LAN chip (certainly not running at 11V, so another DC-DC right there) for WOL services and powering the PS_ON soft-on/off power pin. People try adapting the 5Vsb rail to this and it doesn’t work. People are also stupid (me included) and try adding regular 12V, some drop a bit of voltage by adding a diode or two in series – and then they wonder why it doesn’t work. Well, regular 12V is only available once the power supply is “on”. If there’s no 12V to pull PS_ON to high in the first place, you’re out of luck.
Theory is all fun and games until someone has to make the adapter and sheds blood. Not much, but those things have rough edges. One connector didn’t fit my tool properly and I had to improvise. Pretty sure the connector was out of tolerance, not my tool (XSPC modding thing – as usual, Molex tools are priced beyond reason). Here’s the well used tool and a typical female pin removed from the connector:
Standard locking mechanism on both sides, so the tool just moves in from the front, keeps the two latches from locking, and it’s easy to pull the pin out.
Well, except for when it doesn’t and one has to make a DIY tool with slightly longer fingers that cannot be thicker than ~0.5mm or else they jam everything into place.
Here’s my quick and dirty solution after a few minutes of brainstorming: There’s this old pry tool back from WHL #2. Heavily used and replaced about a year ago by a pack of four slightly smaller ones. Sturdy spring steel (well, at least for the thickness), but I never had the heart to throw the fully bent thing away. That day, it served me for the last time:
I nibbled away a small strip that just statically held one latch in place, then used another strip still attached to the tool to wiggle both of them loose, and pulled (sometimes hard) the connector out. Two scuffed knuckles and a lot of swearing later, I had removed all 24 pins of my last 20 to 24 pin ATX adapter.
I did the same with another 8 pin EPS/ATX12V extension to get a bunch of yellow wires from another powerful 12V source. The ATX adapter yielded enough wires to make the mod, but to keep proper color-coding, I just had to harvest those. Putting it back into a connector housing was a breeze, and after I initially cut up the 24 pin one, I found a perfect 16 pin in one of my parts bins. Luckily that was before putting them in, otherwise that would have been it. These Mini-Fit connectors can be trimmed without having keying issues from one side (note the pattern above!); however, the outside latch is always centered, so a trimmed 24 pin does not latch onto a 16 pin, it’s just the individual pins that hold it in place.
Wiring according to the pinouts above is pretty easy. I made the split 12V cable from the ATX connector the center pin of both 12V 1 and 2 rails, and supplied the rest of them with the additional EPS connector. PS_ON and PWR_OK work the same and are also correct in color. Only thing missing is the 11Vsb, and this is an easy thing when having some DC-DC converters in the parts bin:
I know this is a 1W isolated DC-DC (Mornsun B0509S-1W) for 5Vin and 9V out, and for my application, that is all I need. 9V is sufficient to make everything work. 1W is never actually used despite the original power supply stating 11V/1A. Idle (off) power consumption is listed below, and it is below 0.5W for the entire thing, so far less than 0.5W are actually drawn on the standby rail. The fact that it is an isolated transformer is a waste of money, but not that prohibitive (around 1€ premium). I’ll order a couple more as apparently sometimes these come in handy. Of course 0V in and 0V out must be connected in order to work properly – that wouldn’t be necessary for a non-isolated one.
This is the final adapter cable – of course one could mod the entire power supply to not have an “ugly” adapter, but the Seasonic Focus Gold will likely be used long after the system has been replaced once more. Maybe that time ATX12VO will have arrived in the mass market…
And this is the entire Chenbro 4U rack chassis. A bit empty due to the µATX form factor. Note that the card holder / stability bar on the right side of the case isn’t mounted yet. It does not fit over the moved and also rotated Thermalright cooler (True Spirit 120 Rev. A BW). It also didn’t do that before, hence the few holes in it to make the heat pipe ends poke out juuust a little Making that thing fit is enough material for a follow-up post, as this will require a custom mount for the fan or cutting off the top piece of the fan holder. The card holder mod is straight forward and for pure desktop stuff passive cooling is adequate.
A few benchmark figures and a power consumption comparison:
Old system: AMD FX-6300 (3C6T, released October 2012), Asus M5A97 LE 2.0 ATX
New System: Intel Core i5-4570 (4C4T, released June 2013), Fujitsu D3222-A12 µATX
Shared components: 2x8GB DDR3-1600 ECC (not working in ECC mode on the Intel platform anymore), Radeon 380, BenQ EX3501R 3440×1440 monitor, Mellanox ConnectX-3 10GbE card, 256GB Samsung 830, Seasonic Focus Gold 450W, some slimline DVD-RW and a couple of USB devices.
FurMark 126.96.36.199 3440×1440 no MSAA:
Old: 78°C @ 34fps
New: 77°C @ 34fps
So the PCIe 3.0 connection instead of 2.0 doesn’t speed up anything in FurMark, this is purely GPU-limited. There’s also no cooling advantage/disadvantage with the rotated CPU cooler. Alright let’s start the CPU benches.
CPU-Z with Benchmark 17.01.64:
Old: Single-Threaded 205.9, Multi-Threaded 1022.2, Scaling ratio 4.96 of 6 = 0.83
Old2: Single-Threaded 205.9, Multi-Threaded 757.7, Scaling ratio 3.67 of 4T limited = 0.92
New: Single-Threaded 381.1, Multi-Threaded 1453.5, Scaling ratio 3.81 of 4 = 0.95
+85% single, +42% multi despite the 6-to-4 disadvantage on threads. That’s fucking impressive I have to say.
Handbrake 1.3.3, encoding Ask a Mortician from 720p base material into H.265 MKV 720p30
Old: “average encoding speed for job is 10.743363 fps”
New: “average encoding speed for job is 17.215359 fps”
+60%. Again, at a 4T to 6T disadvantage. Jaw-dropping. Maybe the AVX2 instruction set can be put to good use here?
Path of Exile, just idle in my hideout:
Old: 36-51 fps oscillating
New: 48-50 fps smooth
~+35% avg but no peak improvement. The frame times are noticeable more consistent, but there’s also higher network latency?
Path of Exile, spamming totems:
Old: 39-43 fps
New: 42-44 fps
Well, just +5%. That’s likely GPU-bound at this point. So to me that’s an indicator that the hardware is now evenly powerful and any major fps increases need upgrades in both directions. I’d much prefer better code, as PoE isn’t exactly highly optimized. Maybe the Vulkan API will allow them to step up a bit, it’s currently working in Beta but fps are like half of the DX11 render path, so not an option on 3440×1440 with the mediocre hardware of mine.
And some power figures as well:
Soft-off, so system shut down completely, but power supply still on and WOL active:
Old: 0.4 W
New: 0. 2W
-50%, but given the absolute values, any minor change, e.g. USB standby current, will skew that result. Still, that’s only like 40 cents per year for the convenience of having a WOL ready system.
Windows after booting, one minute to let it settle into complete idle (screen still on):
Old: 70-72 W (slight fluctuations)
New: 41-42 W (pretty steady)
-42%, really impressive once again. I always had trouble with the auto clock of the FX after overclocking and underclocking experiments, maybe it is held in high states longer and/or earlier than needed. Furthermore, the Fujitsu board is limited to 95W TDP (still covering all Haswell 1150 processors), while the Asus is capable of 140W for some overclocking headroom. That always causes higher idle load consumption – but still, 30W reduced base power consumption is a huge deal.
PoE during totem spamming:
Old: 255-265 W
New: 242-246 W
-6% – both GPU and CPU pushed to close to 100% utilization. i5-4570 TDP is 84W and FX-6300 TDP is 95W, go figure.
The custom CPU heatsink/fan mount will be part of a follow-up, so that’s it for today. If you can get your hands on cheap OEM hardware with little quirks that make it unsuitable for the masses: Maybe it is worth a shot. Maybe it isn’t. Google before you buy!