that’s awesome, can you please try MSFS 2024 and MSFS 2020?
These have been my results with 4 lane oculink/5090
MSFS 2024
Ultra settings, 10-20 FPS (and randomly crashes)
High-End settings, 30 FPS but dips to 10 (and randomly crashes)
MSFS 2020
Ultra settings, 30 FPS but dips to 10
High-End, 30-60 FPS and very much playable
sorry but i don’t have that game…
Yes, its niche. But the hotplug side is not electrical. Its software. Electrical can be done hotplug, as its supported by the PCIe.
But I assume its hard to get AMD or NVidia to fix their drivers…
I mean, Framework is supposedly to have great relations with AMD, perhaps some AMD designers helping them as part of their job or something. So, how about the GPU hotplug?
Is hotplug really that important for eGPU? I can stomach having to wait the 20ish seconds it takes to reboot.
Yeah I dont consider it that big of an issue either. Using USB4 on my FW16 currently and I can plug the GPU in after turning on, sure. But if I try to unplug it without first ejecting it, the laptop immediately bluescreens.
Thing is however that the hotplug is not a hw issue, as many would like you to believe.
It works with TB docks, so…
Yes, that’s true, no GPU hotplugging and no power delivery, but I personally am willing to trade it for a couple of fps ;). But indeed, that’s what makes thunderbolt so impressive (and difficult to do, which is the reason why thunderbolt 5 took so long). But I am already used to it, so my preference is OCuLink 8i.
Problem is, as far as I know, it potentially involves software that neither us nor Framework directly control, like the BIOS.
I’ve tried to get some semblance of hotplug working, but to no avail. Rescanning the PCIe Bus does not seem to work, so I assume somewhere on the firmware level (likely BIOS) there’s something not implemented. To be fair, PCIe hotplugging is only really used in some server motherboards, and I’ve heard it’s not uncommon for desktop motherboards to not support it in the BIOS. Not sure about laptops.
Maybe, perhaps, that is why Framework is holding back. They know hotplugging will be an issue, and they have no direct control over fixing it (neither on OS side nor BIOS side), so they don’t proceed. Not saying that’s a good way, but could be.
IF it was implemented in BIOS, I do think there’s a good chance you CAN get it to work on linux at least. Even if it may require logging in and out again.
The best I’ve gotten is that linux won’t crash when you unplug it (provided no monitor is attached anymore, didn’t test with one connected). But it certainly is not a clean state (GPU is still shown in PCIe List, for example) and I wouldn’t trust it to run forever (e.g. after resuming from sleep, forgot to test)
In Thunderbolt Dock NVidia drivers support plug/unplug as it almost instantly show that NVidia connected / disconnected overlay window. So, maybe could be used for OcuLink as well somehow… But not a big deal - I unplugged/connected it just once - just to check how it works and type here this message 
@Gu_tally
Are you planning to make them available somehow? Looks like you did a great work and I definitely would be interested in getting something similar…
Yes, and I fully understand that. My hunch, although of course I do not know, is that management at Framework believes that OCuLink is simply, as a technology, not mainstream/mature enough. And this is where I would seriously consider Framework to be one of the few companies that could and should contribute to making it more mainstream. It’s an incredible chance to do something new. It would make Framework the company that built the incredible Framework 16 with new features. And it would allow it to be the only company that makes an OCuLink 8i EGPU connection possible.
And I get it, hotplugging, sure. But this did not prevent ASUS from offering a much celebrated EGPU solution for years successfully. And other companies (e.g. Minisforum) are building entire product lines around OCuLink nowadays (albeit still 4i only, hence the chance for innovation).
OCuLink 8i just needs a little push, and they could do it so well. Custom cable solution that’s more professional, custum EGPU (allowing to also connect their expansion cards), custom 8i board for the dual m.2 board… Companies like Minisforum are already successful with the DEG1. I am sure the 8i EGPU could be sold successfully too. People would take it due to the speed, as it will be simply faster than thunderbolt 5, so competitive for the foreseeable future.
PS: I am writing this mostly as Framework themselves mentioned they would like to enter the EGPU business.
it is going to be. but for now development will pause for a while because I have a dissertation coming up… and I aim to resume it after June. But a second version of the board is already sketched.
Great!
Good luck with your dissertation!
Just to add to my 8i point, OCuLink 8i is starting to get traction, other fairly small companies like Xyber are implementing it already… I am not saying we need 3 screens any time soon, but I believe it would have been possible for Framework to be the first here. Such projects certainly show that the competition for the Framework 16 is on.
Why does everyone get so gun-ho about Thunderbolt 5?
It’s just an implementation of USB4v2 that is identical except its limited to Intel.
Intel have mandatory certification for Thunderbolt and don’t certify non-Intel platforms.
USB4v2 has 80Gbs/80Bbs symmetrical, 120Gbs/40Gbs asymmetrical and 240w Power Delivery - the same as Thunderbolt 5.
I think people should stop pushing for TB5 and instead be asking for USB4v2 at least AMD can officially support it (I know there are workarounds, but it doesn’t seem necessary).
Occulink on a Framework would be amazing, the main reason I want a Framework laptop is to use it with an eGPU.
But I feel it would be easier to get USB4v2 support (and I suspect in future more EGPUs will support it if mainstream vendors continue to ignore Occulink, which we all wish they wouldn’t). I am not very familiar with Occulink though, would there be that much of a difference between USB4v2 and 8 lane Occulink?
I think most people mean USB4v2 when they say Thunderbolt 5. It’s just more clear to express what you mean with Thunderbolt, because USB spec versions are unclear and hard to remember.
Thunderbolt technology has typically been a forerunner to more advance functionality than USB introduced by Intel and their special interest group around thunderbolt products and the vendors that use them and make hardware. More better, so everyone gets excited. TB has long since offered more bandwidth and more features (such as pcie and displayport extension) than plain old USB iterations offered at the time, but originally started off using their own proprietary connectors beyond what the USB (and Firewire) spec offered (see old macs).
Eventually everyone that isn’t Intel and their sycophants eventually catch up with a more broadly available USB spec that isn’t license and royalty-encumbered as TB is to Intel directly (thus over AMD’s cold dead body will you see a TB port on an AMD board unless a mobo vendor goes rogue).
Looking further back, USB (Universal Serial Bus) came about to overcome bandwidth limits to at the time traditional Serial Bus could offer over any distance more than a few inches, with a better smaller connector than an old DB9. It’s just evolution of one external bus or another for connecting things.
USB is also designed to be a sharable bus with a certain amount of oversubscrption and buffering built in where PCIe is not, but this results in potential overload and performance will suffer as a result to all shared services when overloading a buffer. These use dedicated controllers and drivers (software) to provide an interface to hardware connecting, so tend to vary in quality and capabilities as a layer of complication PCIe doesn’t have with set 1:1 bandwidth and lanes built for close on-board transmissions.
The best bandwidth and latency for a PC to connect anything has always been their local motherboard busses such as PCI and later PCI-E (an evolution of ISA and others prior), but distance and cable limitations haven’t traditionally allowed for an extension beyond a few inches, thus the aforementioned busses like Serial and USB were more suitable, but slower and less bandwidth as a sacrifice.
Evolution in high-speed busses have evolved in general with PCI-E (and newer CXL) busses to be transportable in longer distances at full lane speeds, but still shorter than USB/TB. Occulink branding is simply a suitable pluggable cable interface (male+female ends) format that’s been standardized for use, and when combined with suitably shielded wiring to prevent interference, can be transported up to 1 Meter theoretically. The SFF connectors used began life for SATA 4-to-1 breakout cables, but proved useful for PCIe as well.
The ultimate evolution of pluggable busses (currently) is Nvidia’s NVLink where they’re using 800Gb interfaces to interconnect GPU clusters between racks and rows in data centers to NVLink switches, or basically just low-latency Ethernet (or Infiniband to be specific).
… or close enough. At least that’s my 2 cents on PC history.
I’m more sad manufacturers seemed to figure out 4x PCIe lane extension was possible (and desired by consumers) on given external connector and cabling lanes, but then got lazy and stopped there beyond extending to 8x or full 16x lanes in an external cable/connector/docks to crank out low end disposable and limited low-end egpu’s and mini-pc’s. Gimme the full lanes dammit, especially when the GPU’s are larger than the PC’s now and likely to snap the socket off!
I think the eventual evolution will be from PCIe towards something like infiniband.
It is also easier to pass over optic fiber, and thus eventually have direct chip connections, resulting in the length limitations of copper being eliminated.
Expensive data center AI computing is moving towards direct chip optics, so it will eventually move to the desktop/laptop.
This thread s incredibly long, can it be bought? Thanks
No.
You can, however, get it working with the nvme expansion bay module, though it requires a bit of DIY. See OcuLink eGPU works with the Dual M.2 expansion bay module.
I posted on the other thread, in regards to being able to buy the oculink 8 board and cable and have it working with the oculink 4 connector for now. so once either board is released you dont have to repurchase boards and cables.
“I went ahead and ordered all the parts. I can confirm that oculink 8i dual split cable and 8i board work with only one 4i connection. Got the error 43 after hooking everything up and applied the patch and it now works no problem with my RTX 3070. Here are the parts I ordered. Cable, Board, M.2. The M.2 that I received was one with 4 holes not 2. I got a sonnent thunderbolt egpu enclosure off ebay that had a borken thunderbolt port and just removed that board. I ordered a 3d print of the back panel for the framework expansion module. I didn’t take any pictures during the process but if anyone wants after pictures i can post some.”