[HN Gopher] Not all OCuLink eGPU docks are created equal
___________________________________________________________________
Not all OCuLink eGPU docks are created equal
Author : speckx
Score : 92 points
Date : 2025-09-29 14:46 UTC (8 hours ago)
(HTM) web link (www.jeffgeerling.com)
(TXT) w3m dump (www.jeffgeerling.com)
| privatelypublic wrote:
| Are they actually extra? Or does the other cable simply route
| them over coax? Answer here is to grab a couple SFF-???? Breakout
| boards, a scope, and a signal generator.
|
| Another thing to consider is isolating all but the first PCIe
| lane.
| geerlingguy wrote:
| > Answer here is to grab a couple SFF-???? Breakout boards, a
| scope, and a signal generator.
|
| Yeah, I've considered buying an extra adapter and just probing
| pins -- I don't have a spare M.2 to OCuLink adapter but will
| probably order a few different models to see if there are any
| differences there too.
|
| > Another thing to consider is isolating all but the first PCIe
| lane.
|
| The Pi nicely does that for me ;)
| privatelypublic wrote:
| Haha, touche- forgot the Pi5 only has a single lane.
|
| The comment came out of knowing if the cable's a little dodgy
| the timing could be off- scrambling the PCIe signals (easier
| to just tape it off than look up if its _supposed_ to matter
| and then tape it off anyway), and the fact my first HBA had
| the "QC pass" sticker prevent some lanes from connecting and
| the whole card was unresponsive as a result.
| SilentM68 wrote:
| In my experience, incompatibilities like that have for the most
| part been due business partnership decisions. Let me give you an
| example. A company that designs motherboards may state that one
| board should be compatible with a full range of other hardware in
| the market, e.g. RAM, PCIe, M.2, etc. The reality is that they
| usually partner and/or are locked into certain vendors,
| contractually, that supply them with their products for the
| board. You can see this in their products spec pages, in a
| section titled, "Support List," with a warning that should you
| deviate from this list, they will not guarantee 100%
| functionality. Basically, the company tests each product that
| they list, and if does not work properly, they either tweak the
| board, its firmware, or communicate this to the third-party
| vendor to see if they can provide a solution. Most companies that
| sell these boards or devices more often than not, are small, do
| not have enough R&D funds, or manpower to do the proper research,
| testing and adjustments to their products, since devices are
| already 6 months behind current technology when they are released
| into the wild. Usually, customers can call into the company's
| support line and report the problem. The company may or may not
| address the incompatibility of their board with the third-party
| product if they see that there is potential for increased sales
| should the third-party device prove to be a trending product, or
| the customer is a high-volume order customer. More often than
| not, the incompatibility is not properly addressed.
|
| Just my thoughts :|
| cmcconomy wrote:
| i had a nuc with an eGPU, connected via a simple usb/thunderbolt
| connection, and I recall it was a nightmare to setup
| coredog64 wrote:
| Same for my wife's old Mac Mini. Finally gave up on it and
| bought her a new M4 Pro
| dylan604 wrote:
| I just bought an external thunderbolt eGPU box (even thought
| it'll never support a GPU with its mini form factor) to host
| a blackmagic 4k display card. Luckily, I'm still on the last
| gen i9 CPU so it worked right out of the box once I found the
| slightly older software. I've read people have issues getting
| it to work on the M* series chips though.
| monkmartinez wrote:
| I have read that thunderbolt and oculink are very different in
| this regard. Whereas thunderbolt devices can be plugged in at
| anytime, the oculink needs to be plugged at boot time. This
| seemingly innocuous detail is the catalyst as to the reason why
| oculink is better performing. It comes down to PCIe vs
| Thunderbolt in general.
| arcfour wrote:
| While PCIe as a standard allows for hot swapping I would be
| quite surprised to learn that any motherboard or GPU
| supported it. At least in the consumer space
| geerlingguy wrote:
| Lenovo's TGX 'extension' (I guess? It was for their eGPU
| solution) allowed hot swap, but support for it is
| definitely not very broad.
| officialchicken wrote:
| I understand the focus on the cables, and as an owner of several
| Minisforum devices, IME they don't really do compatibility very
| well (e.g. linux on V3). Their specialty seems to be taking
| reference PCB's, down costing them by changing/removing
| components, and then putting them in very unique enclosures.
| They're one vendor I would avoid considering - unless it's x86.
| jauntywundrkind wrote:
| I love the hardware Minisforum makes, but man, this company
| ongoingly sets consumer expectations lower and lower.
|
| I haven't checked in the past year, but historically Minisforum
| has been awful about having bios updates, doing so seemingly
| only when truly needed.
| int0x29 wrote:
| The one AMD x86 box I got from them occasionally randomly
| segfaulted and would blackscreen and hang if left on on after
| anywhere from 45 minutes to a few hours. I read that it might
| be a driver issue in proxmox debian and tried running windows
| on it and had the same issue. Oddly memtest passed on it while
| running repeatedly for 3 days strait before shutting down
| unexpectedly.
|
| Its kinda made me hesitant to trust any of these cheap minipcs
| ever again.
| Cieric wrote:
| I've actually been dealing with this myself recently, but only
| the gpu portion. I've been trying to get a gpu working on the
| orangepi RV2 just to see if I can. The tools for building the
| kernel for orangepi are kind of bad though.
|
| But no, the 2 old amd gpus I have just don't show up at all in
| lspci, testing the RTX 2080 and RX 7800 both did show up though.
| syntaxing wrote:
| I can't seem to find the answer about this, is there anyway to
| get an oculink through the x4 slot on framework desktop?
| Palomides wrote:
| oculink is basically just pcie, you can get a dumb adapter card
| for like $20
| 0x457 wrote:
| There a cables/adapters to use NVMe slot for oculink. It should
| "just work".
| chao- wrote:
| If you mean "can it be done with the connections" the answer is
| yes, as there are many PCIe x4 add-inboards for an external
| oculink port on the market.
|
| As for framework's desktop specifically, their site says this
| about the PCIe x4 slot:
|
| _> 1 x PCIe x4 slot (not exposed on default case)_
|
| No opening in the back of the case for an expansion card that
| exposes ports. My guess is you would be best off putting the
| mainboard in a different case, but alternatively you could
| either:
|
| (A) Cut a big hole in the back that can expose an AIB with an
| external oculink.
|
| (B) Use an internal oculink, and cut a smaller hole (anywhere
| you want) to run the oculink cable through.
| monkmartinez wrote:
| You mean like fish a cable through the case? It looks like it
| might be doable based on the pictures. If its in the right
| spot, you might even be able to 3D print a cable gland or
| something at the case. Hard to tell based on the pictures.
| bigwheels wrote:
| I like how the comments section on Jeff Geerling's website now
| looks exactly like an HN thread. I did a double take, haha!
| geerlingguy wrote:
| It's spartan, text-focused, and threaded; I've enjoyed having
| nice tidy threads, and the comments often contain a lot of
| great nuggets of wisdom over the years (which is also why I
| leave them open as long as possible).
___________________________________________________________________
(page generated 2025-09-29 23:00 UTC)