U.2 to quad M.2 carrier (2.5 inch form factor)?

[nexox]

Wiring and power consumption are blocking issues of course, but additionally M.2 only supplies 3.3V, while the SATA power connector provides 12V and 5V, which most devices with those connectors are going to need.

 

[slowlaris]

I'm having a hard time selecting the right cable. So most of the adapters look like this :

(disregard wrong Molex pinout, it's usually SATA power connector there)

Ideally I'd get something like this, but for a single U.2 obviously:


I've also found this, but it looks like 3.3V is made by this tiny onboard regulator and it doesn't look very reliable:




In my case I only have a single Molex power connector available on the motherboard and I'd like to avoid using Molex -> SATA if I can; for example with the latest adapter I can just cut off this SATA connector and crimp a Molex there.

Another problem is that I won't have much space in the case (ca. 5 cm) around the U.2 connector, so I'd like to avoid bending the cable too much. Adapters provide a way to just plug Mini SAS HD SFF-8043, and those are easily bent (and there angled ones) comparing to U.2 one.

Do you have any suggestions?

 

[UhClem]

I'm having a hard time selecting the right cable.
...
Do you have any suggestions?

A U.2-NVMe device (which is how this Viking cutie is expecting to be treated) only requires 12v. (All 3.3v etc. needed internally, and supplied externally [4x M.2], will be derived from that.)

Your real focus should be assuring decent airflow (top and bottom), so a vertical placement might be preferred.

 

[slowlaris]

A U.2-NVMe device (which is how this Viking cutie is expecting to be treated) only requires 12v. (All 3.3v etc. needed internally, and supplied externally [4x M.2], will be derived from that.)

Your real focus should be assuring decent airflow (top and bottom), so a vertical placement might be preferred.

Thanks, that is good to know. I'm planning to put Viking in the 3.5" slot (essentially acting as a rectangular pipe) and cool it with 2x20mm fans on full speed; I think it should work. Currently I'm waiting for all parts to arrive but I'll post it here once I assemble it all together.

 

[slowlaris]

Okay, check this out. 2.5" to 3.5" Goobay adapter, couple of M3 standoffs and two MOSFET heatsinks I had lying around (cut from old PSU board) which happen to be exactly 20mm in size, so fans fit very tightly. I will swap fans for Noctuas but you can get a general idea. And yes, apart from the bolt height (need to take shorter standoffs or use another bolt heads), this fits in 3.5" RDX drive slot perfectly. I also plan to put heatsinks on the drives for the ultimate cooling experience

It may not be obvious from the pictures, but standoffs are screwed directly in the Viking, as it has both horizontal and vertical mounting holes on its corners. I've also tried M2 standoffs because there are threads for them, but it didn't provide a stable mount; M3 worked perfectly.

 

[Mithril]

A U.2-NVMe device (which is how this Viking cutie is expecting to be treated) only requires 12v. (All 3.3v etc. needed internally, and supplied externally [4x M.2], will be derived from that.)

Your real focus should be assuring decent airflow (top and bottom), so a vertical placement might be preferred.

That seemed wrong to me so I checked the drives I dont have installed. All intel (so maybe specific to them) but both my Optane U.2 and intel DC U.2 specify 12v AND 3.3v on the label. However the 3.3 is in the 10mA or less, so a small 3.3v regulator *should* be fine. Going the other way, 3.3v from a m.2 slot to 12v (even from 5V honestly) would be fairly bad however

 

[slowlaris]

Finally Noctua fans are here. I've had a revelation that if I drill holes in a a side of these brackets, I could just mount them to an adapter, and so I did. I've also applied duct tape to make it smooth, and added heatsinks for the SSDs.

Mounting in the case is not perfect (I'm off by 1mm lol), and for that I'd had to make a cut in both brackets so they can slide onto the adapter and go underneath it (currently they sit on top)... but I think it's good enough. Most importantly, it fits inside of an empty RDX drive slot of my Fujitsu Primergy 1320 M4

I will also add a temp sensor since I have a spare fan controller I can use. I'm very curious to get the temps under load.

 

[slowlaris]

Here's some updates... as mentioned in this thread, power consumption by this card alone is around 7W, but I wasn't able to precisely tell how much more since my board is stuck in Pkg C3 state anyway (due to the BMC). Trying to optimize power states I've discovered that ASPM is actually supported, but not advanced L substates; e.g. PCI clock is always active and is impossible to switch off; this might or might not contribute to your system power consumption. In my case Viking is connected to the PCH and not to the CPU directly, so I guess it'll keep PCH always active. Given that for my particular configuration I cannot reach less than C3, it's not really relevant (apart from some amount of power the card definitely consumes).

Trying to look into it further I've stumbled upon this repo: GitHub - Microsemi/switchtec-user: Userspace code for the Microsemi PCIe switch . Switch is detected by the tool, but that's about it:

root@pve:~/switchtec-user# ./switchtec list
switchtec0              PFX 24XG3       RevB     1.09 B063     0000:08:00.1
root@pve:~/switchtec-user# ./switchtec fw-info /dev/switchtec0
/dev/switchtec0: No data available
root@pve:~/switchtec-user# ./switchtec status /dev/switchtec0
/dev/switchtec0: No data available
root@pve:~/switchtec-user# ./switchtec bw /dev/switchtec0
/dev/switchtec0: No data available
root@pve:~/switchtec-user# ./switchtec latency /dev/switchtec0
/dev/switchtec0: No data available

In any case, running it with Proxmox (kernel 6.14) it works fine but boots with lots of pci-related spam in dmesg (cannot assign address ranges etc). After some debugging it turned out switch chip tries to allocate all 24 ports as hotplug ports, and the kernel has hard time doing it (it takes up to 4 reallocation tries during the boot). So far the cleanest and fastest boot I get with these kernel params:

pci=realloc pci=hpmemsize=0

These instruct kernel to not rely on BIOS PCI allocation and do it manually, and also disables hotplug memory allocations (up to 2MB per port, which are wasted since you'll have a hard time hotplugging something into the card anyway).

Finally, temps. I've attached two probes, one directly to the chip under the heatsink, and another one on a side of heatsink (where screwholes are). I'm getting around 37 C on the chip and 30 C on the side of heatsink during fio randread test, both fans are at 2400 RPM (50% pwm).

 

[slowlaris]

Just found this and it could be a perfect usecase:

Forlinx FAI-ARA240-M Packs Ara240 NPU into M.2 2280 Module

Forlinx Embedded has officially launched the FAI-ARA240-M, an M.2-based AI accelerator built around the NXP Ara240 processor. The module provides a discrete NPU for offloading inference workloads from embedded host systems. The Ara240 processor was first seen during the launch of the...

linuxgizmos.com

• Processor:
  • NXP Ara-240 Edge AI processor
• AI Performance:
  • Up to 40 TOPS
• Memory:
  • 8GB / 16GB LPDDR4
• Form Factor:
  • M.2 2280 (M-Key)
• Interface:
  • PCIe Gen4 x4
  • USB 3.2 Gen1
• Mechanical:
  • 80 x 22mm