Test of the AI PC kit

#17

@willy OCuLink is an open source analogue of Thunderbolt. It is a cable that channelizes 8 or 4 PCIe lanes to an external PCI enclosure or external PCI peripheral.

For example, see this external GPU which takes an OCuLink connector: https://www.gpd.hk/gpdg1graphics . Also here https://www.amazon.com/GPD-Docking-Station-Charger-Perfect/dp/B0CHMXSG8Y?th=1 and here https://gpdstore.net/kb/accessories-support-hub/kb-article/getting-started-with-the-gpd-g1-egpu-docking-station/ . Note “SFF-8612” is synonymous with SFF-8611. In this enclosure I think the OCuLink is 4 lane only. Not sure if it’s PCIe v3 or 4.

image
image.jpg1492×904 334 KB

image
image.jpg1210×1008 343 KB

As for the Orion O6, of course the OCuLink supports PCIe v4 which is double speed. What I like to understand is that all 8 lanes are connected i.e. it’s “OCuLink 8i” (x8 lanes).

If you like, have a look at the PCB if it looks like connectors from the whole 8 lane width of the PCI socket are connected to the SFF-8611 connector. Anyhow I’m sure @jackw will let us know next week.

@willy If you want, can you share some close-up photos of all details of the PCI to OCuLink adaptor included in the AI kit?

1 Like
#18

Thanks for explaining! So it’s 4 lanes here, as I’m seeing ~9-10 pairs routed to the connector, hence 4 RX, 4 TX, and a few clocks etc:

IMG_0719cs
IMG_0719cs.JPG1920×1014 135 KB
IMG_0720cs
IMG_0720cs.JPG1343×1080 104 KB

1 Like
#19

@willy Can you send a photo looking into the socket too?

@willy Also is the 12V lane connected?

Are you sure only 4 lanes are connected. I wonder why that would be the case as this adapter is custom made for this PCI slot which has 8 lanes. @jack

#20

In the Orion o6 diagram, in the “PCIE_X16_SLOT” socket, they are only connected halfway down the socket, so the socket is only half occupied.

#21

@blondu It makes perfect sense that connections only go halfway in the X16 socket, because it’s a X8 (8-lane) PCI connector installed into an X16 socket (made for 16-lane PCI).

If the PCI-to-OCuLink has coords halfway, it means it is connected to 8 lanes.

#22

No it’s quite less than half way, for me it matches the length of an x4 connector.

#23

Waiting for the AI ​​PC Development Kit, it’s on its way!
Also solves the need for two M2 slots for two systems.

1 Like
#24

Yes, I found their low-profile PCIe-M2 adapter very nice. It’s even thinner than the one I modified myself for my macchiatobin, because the M2 connector is placed so that the SSD goes over the PCIe connector. And this will indeed permit you to have two M2 in the same machine and without changing the form factor.

The only thing is that is condemns an x8 slot for only x4. The top would be to have a second M2 x4 that would support an FPC cable to an auxiliary connector but if the bus doesn’t support bifurcation, a PCIe bridge would be needed, and a Gen4 x8 to 2x4 might not be cheap at all. Or we could have a 10/25 or 40G NIC + M2 adapter on the same board, but that wouldn’t fit in such a small form factor!

#25

The problem occurs when I want to use the video card and the second SSD. I use two operating systems: Fedora 41 and Windows 11.

#26

The PCIe x16 to Oculink adapter board we made is x4, this is because Oculink x8 connector is much wider, you can check:

https://www.amphenol-cs.com/product-series/oculink.html?page=1

2 Likes
#27

What type of cable is used: SFF8611 or type?

From the construction of the cable and the plug, it can be seen that it is x4.

#28

I ordered an Expc Oculink Gpu Dock until I get your AI ​​PC Development Kit and I will be able to test the Oculink port.

Oculink dock

I don’t know if it will work.

#29

The AI ​​​​PC Development Kit arrived and I installed it.
It is very silent and the temperature rises to 60 degrees Celsius at the most at the LAN chip. I encountered two problems, one when installing the screw on the motherboard in the area of ​​the Oculink socket I could not install it (it does not fit next to the adapter socket and cannot be threaded). And the second is the wireless card antennas I do not know where to put them, the casing is metal.
Otherwise it behaves well and the design is beautiful.

2 Likes
#30

@willy What millimeter dimensions does your copper plate have exactly?

So after adding that copper plate and thermal grease below and on top of it, the O6 has nice cooling and noise properties, no configuration needed?

What about taping some air vent holes on the bottom of the fan, did you keep that tape, why?

If you have some purchase URL (on Aliexpress Taobao Amazon Teemu etc.) for the copper plate please share.

#31

It’s really not convenient to disassemble it all, so I restarted from the photos to take the measures. On this page, there’s a high-resolution photo of the board: https://docs.radxa.com/en/orion/o6/hardware-design/hardware-interface
I’m getting 965 pixels for 75mm spacing between holes. The die is 214px * 203px hence 16.7 * 15.77mm, and the inner space covering it is about 278px hence 21.6mm. On the photo above my pad is slightly too large, maybe 22mm. I think you should take 20x20mm and you’ll be fine. And it was 1mm thick BTW.

In my case, no config needed, though I slowed down the fan a little bit at rest (echo 15 > /sys/…/pwm) because I was hearing a faint buzz in my office that was attracting my attention (I’m extremely sensitive to fans and recurring noises). I also noticed that once it starts running faster due to temperature, it slows down at the default PWM speed which is a bit higher than 15, maybe 20 or so, so some times after running a build on it, I’m thinking “what’s this humm? ah yes of course” and I just ssh to it and echo 15 again. To be honest, 99.99% of the people won’t hear it at all.

I didn’t find it logical to suck hot air from the SoC to blow it through the heat sink, so I preferred to leave the tape. I have not made any measurement to compare with/without. Maybe it has no effect, maybe it’s counter productive, but the combined changes worked for me.

Unfortunately I don’t, I salvaged it from a larger one a long time ago (it has cut traces on two sides) and I had to flatten it a little bit (copper bends easily and when you pick it from your salvage tray, it’s rarely completely flat).

I think you can really take any copper pad of 1mm thick and around 20x20mm. Often they are sold in bags of 3, 5 or 10, and are aiming at GPUs, northbridges etc.

The thing I have not tried was to file the threaded holes to diminish their height and make it possible to press more on the SoC. I think it could be sufficient, but most likely the copper pad helps here by being larger than the SoC and spreading the heat into a larger area on the heat sink which likely has a moderately high thermal resistance due to not being very thick metal. But we’re talking about optimization and not something necessary IMHO.

1 Like
#32

Thanks again for all your detail here. I put my AI Kit together yesterday and I think you’re right - the contact is poor: Very little heat exhausted from the fans when going full pelt.

Have ordered some 20x20x1mm copper pads. When they arrive, I’ll give them a go and let you all know if they appear sufficient.

1 Like
#33

Do you have a file to try to shorten the legs under the heatsink ? I think that removing 0.1mm should give a first idea of what’s to expect (I think there’s at least 0.5mm too much).

#34

Just reporting back: I installed a 20x20x1mm copper pad under the heatsink and I do think this made a big difference for me too.

Where previously, it would heat up to 60-65C while doing some Llama.cpp stuff, it’s now usually hanging in the 40C’s.

I’m wondering if this might also reduce power consumption a tad due to the fan not needing to blast as hard: Previously, consumption would hit ~20W, where as my tests now have it sitting at around ~18W or so.

1 Like
#35

I hadn’t thought about that but it totally makes sense since the fan itself draws a few watts when spinning. Anyway, since I value my ears way more than a few cents by year, I was mostly focused on the fan’s noise :slight_smile:

@jack someone at Radxa should definitely look into this. I think you’re risking a number of failure reports due to the heatsink not touching the SoC well by default. The modification is not hard (adding a 1mm copper shim) but is not trivial for end users. I think it can be easier at the factory to just file the fan’s fixations by 0.5-1mm to make it touch well. Otherwise adding a copper shim would do the job.

#36

I think we had looked into this since your previous post. I will get back to you on this once confirmed with our mechanical engineer.

1 Like