Fitting a proprietary Supermicro X9SRW-F into a standard 3U 836 case (#P23)
Blogging’s a bit tedious right now, as shipping parts is rather slow at the moment, and there’s so much going on at work…
This has been in the making for the past two months now. Some very nice dude gave away “old” servers at a popular German hardware forum, but requested local pickup due to the difficult and expensive shipping of entire server chassis. These were decommissioned colocation servers from his own business, aged 7 to 12 years and no longer viable for 24/7 operation – still decent machines for occasional use. As at the time I needed to visit family which is a 700km round trip in that general direction anyway, I decided to slap on another 200km detour around Munich (and the Tegernsee) to pick those up. Supercharging is free! He listed older systems up to a Dual-1366 HP system which I was interested in, as well as two mystery boxes which turned out to be the very best of the entire stack. One other guy also wanted to pick up an identical HP which got some RAM from mine, those were equipped in dual-channel style despite being triple-channels. Well, he’s obviously not a hardware guy, but rather into software.
Here’s how I left (tons of rails scattered in the rear footwell, as there’s no transmission tunnel for obvious reasons):
So aside from the huge mess the disassembly of those six server chassis caused, the 2U mystery box was promised to another guy, and I picked the rather unpopular 1U chassis with close to identical hardware. A Supermicro CSE-815TQ-R500WB equipped with an X9SRW board. Bingo! Single-socket 2011 system with Sandy/Ivy Bridge support, 8 DIMM slots for up to 256GB of DDR3 ECC Reg (or 512GB or LRDIMM), up to five PCIe 3.0 x8 slots depending on riser, and IPMI support. A clear successor of my X8DTL and by far the most powerful piece of computing hardware that I own!
Now the X9SRW is listed as proprietary form factor board. Which, in general, is true, as it’s 331mm x 207mm in size, absolutely not conforming to any nearby ATX standard (345×262 XL-ATX, 305×244 ATX, 284×208 µATX [not a rounding error]). It does, however, have a similar placement of mounting holes. I checked with my older gen 836 chassis and only two of the nine were off, plus the back I/O panel would not only fit exactly into the standard cutout, but there’s also a shield (MCP-260-00042-0N) available for cheap. That one however took around 6 weeks for delivery from Russia…
…and didn’t need to. Turns out this slightly newer revision 836 does have the required mounting holes, I think labelled H and U. Well – perfect! So much for the proprietary boo-hoo…
The X9SRW however has some downsides, one being its riser-only expansion slots. There’s one x8 PCI Express connector on the board that’s hard to access (riser much recommended), and there’s a proprietary x32 connector that needs one of several risers – dual x16, quad x8, and some oddballs, including support for port bifurcation that’s needed for like quad x4 M.2 adapters. I pinched the quad x8 riser from the other chassis as my 1U chassis obviously didn’t have physical support for it. But those cards flap around in the breeze, they do need some support.
A couple of design iterations later, I made this:
The quad riser card holds two IBM BR10i cards on the bottom (8 SAS ports each) and the Mellanox ConnectX-3 card above (already shown in WHL #42). Still looking for a cheap small NVMe card as OS disk that might run in the top slot. So only the 10GbE card requires an outside connection, all of the others only need support structure in the back. I linked them with simple spacers for some extra rigidity of the Mellanox card.
The top has some notches for the former vertical card mounts that I simply removed with the Dremel. I do have another back piece from the other 836 case that is compatible, so in the event of going back to an ATX compliant board, I can simply swap those. It’s just a couple of screws. Oh, and for the extra HP-esque look, I tapped the printed part and put in a lot of harvested HP torx screws. Which doesn’t exactly hold it in place rock-solid, so one more screw from the bottom or side is required. I’m still waiting for that to arrive (extra-long countersunk UNC #6-32), so for the mean time, I cheated with a M3 one that’s barely a friction fit. Screw holes are integrated into the 3D model. Hexagon size however isn’t perfect for the print, as the wall thickness is 1.0mm and my colleague printed it with a 0.6mm nozzle, hence it’s not gotten perfect – but it’s REALLY sturdy despite a couple of local issues. Adding the coarse UNC threads is also no issue, even for structural purposes. Just make sure your print is somewhat dense locally – you do not want to thread through a touch of support material.
DAC cable fitting nicely:
That’s the required piece done – now for the optional stuff. I’ve removed the redundant power supply bracket a long while ago, I simply do not need and do not want dual 800W+ power supplies with 40mm fans in my living room. An ATX supply will fit barely – problem being a tiny piece of the bottom chassis (motherboard mount) that’ll tilt the power supply, making it stick out of the chassis envelope. I went for a SFX unit instead. Power’s sufficient (450W), it’s semi-passive and the short wiring loom is basically perfect for the tiny space that doesn’t allow for modern cable management.
I’m sorry for not having a proper photo, and the unit is already back in the the rack. I’ll revisit it one day, I promise!
This bracket likely is adaptable for other ATX to SFX conversions, only the part that’s sticking out to the side and the raised drill holes on the side are custom to the Supermicro chassis. But there’s also plenty of other examples on thingiverse.
This addresses two issues at once: First, the power supply is now somewhat close to the end of the case and in a fixed position. That wasn’t the case before, see the picture with the old board still installed. Unplugging the power chord is now much easier. Second, as the chassis is U-shaped in the back, it requires a continuous piece mounted perpendicular to that in order to gain some stability. That was only partially the case before, as taking out the entire dual power supply tunnel made the remaining blank plate quite weak. Now with the beefy printed part in place, it’s almost back to original spec.
I do remember the day when I bought my first computer and it also showed 131072 of those rad RAMs in the BIOS. Only back then it was Kilobytes…I still have that DIMM around, in the “good old memories” box.
Bonus footage: Supermicro was founded by an Asian couple. It still shows…
Welcome back to Wan Hung Lo, where safty is number one prawowity