This is great 
I’m back on the Queue after seeing this is possible, would be amazing if we get the Noctua 3D panel with the hole for the Oculink
On a peripherally related note, anyone had much luck running llama.cpp in a multi-GPU setup with an nVidia dGPU? I’ve gotten it to load, but it desn’t make full use of either dGPU or iGPU.
Sorry about the lack of replies here. I’ve been dealing with a hardware issue that requires RMA. That issue is completely unrelated to the problem we’ve been experiencing with dGPUs. Since I’ve worked around the problem for now, I’ve been doing my typical workload with VMs, web surfing and gaming and having a mostly smooth experience in Arch Linux.
I do know that some other users, while they can get the dGPU to boot, have very long boot times where it seems to boot, restart and then will display a UEFI loader (Grub, Luks login, etc.) I don’t believe that issue is a configuration problem but more so a problem with the pcie riser being used. If you have this problem and are able to build a UART debugger, lemme know. I’m collecting people in a discord to assist with any problems directly.
I’ll want to start doing more AI workloads once I get the RMA processed.
So my riser finally came in
And I picked up the Intel Arc B50 Pro
I Frankenstein-ed my set up back together, I’ve been working on a custom cold plate.
And it looks like it’s just working happily
I’m about to grab my other monitor and plug it into the intel card and test some more stuff like under load and whatnot. But I was just so excited because I believe that terminal output means that even resize-bar is working.
Annnnndddd: I don’t have a mini display port cable or adapter…
3 Likes
I got my mini display port adapter in and it works just fine. In playing around for a few hours I couldn’t get steam on linux to use the dgpu over the igpu to load test it… but with the display coming out of the dgpu it was able to game off the igpu which was a neat trick, event if it’s not what I wanted at the time. Other quirk is that somehow having the dgpu installed has made the system very mildly unstable, if I leave the computer alone long enough for it to go to sleep I instead come back to a wall of fullscreen terminal error text. (I’m testing on cachyOS right now)
So aside from the mild instability, it “seems” to be working? I need to play with it more and make sure the gpu can be fully loaded up happily. Also the reason I got the gpu was not to use it for gaming, I bought it for SR-IOV for virtual machines… which that feature for the card is so new that the only people that have it working got it working “early” so… hoping that soon official support hits and guides appear.
To be clear my usability goal is to use the igpu for everything on the host, and to just have really strong virtual machine performance with the 16 cores and being able to pass through chunks of the dgpu. (I regularly use CAD and old adobe applications, but those don’t play nice on linux even with wine, so I run them in a VM but that VM needs gpu acceleration to be smooth)
Are you able to post any of the errors you’re seeing? I thought I saw errors with when adding the dGPU but those ended up being memory corruption due to a bad trace to one of the memory chips.
Might also try adding “pci=realloc” to your kernel boot parameters. This didn’t really help my issue but you might have better luck as this parameter will attempt to reallocate BAR space and could fix issues with the firmware being wonky about the dGPU bar allotment.
hi. would you mind providing a link for the raise please
depends on which model you want: https://www.adt.link/x4.html
1 Like
I’ll try again and with your recommendation and report back
1 Like
Also passed the intel arc b50 pro through to a windows VM went to install drivers and it decided it needed the new firmware… hopefully that will add to the stability…
I also found before adding the kernel parameter and before updating the firmware that it only liked to crash with a monitor plugged in. So let me spend some time fiddling with it and see if I can get it to crash again…
So not sure why but even before tinkering with things today I did have it crash twice if a monitor was plugged into the intel card… But I added the kernel parameter anyways… it still crashes randomly with a display plugged into it on the host…
but unplug the monitor, boot the host, configure the VM with it passed through, then run the vm…
Resizebar confirmed:
and I can even load it up fully with a game in the VM with no issues
My guess is that the card has some kind of driver bug with my kernel/firmware/whatever on cachyOS, but I can at least confirm that everything works with the GPU running on the framework motherboard, nothing fancy needed… just a Adt-Link riser that has power injection.
Now to be a monster and run another PCIe card off the bottom M.2 (not GPU) and my dream workstation config will be a reality…. And then you know design the entire case and all that, because I have plans for this contraption… and even on a surface level no one makes a case to support M.2 → PCIe risers
1 Like
Sounds like your use case is baby-threadripper. Glad you’re able to figure it out. Let us know if you have any issues with SRIOV.
Baby Threadripper but with strix halo’s gpu memory 
Yeah full plan is:
- 16 core platform
- Strix halo igpu for models/games/etc
- Intel arc b50 pro for multiple gpu accelerated vms
- Another PCIe usb-c card (ideally with multiple controllers… because passing through usb devices is annoying but passing through controllers gives the VMs hot swap, I’m looking at this one: Amazon.com: Sinefine USB PCIe Card Power Delivery 4 USB Type-C Ports, PCIe USB3.1(USB 3.2 Gen 2) with 10 Gbps Motherboard Card for PC Desktop : Electronics but waiting for someone to answer me about if it indeed has 2 usb controllers on it or not)
Edit: Yay the seller got back to me, they say it is indeed 2 usb controllers, so I placed the order.
The bonus fun part is in addition to this eventually being my main computer, I’m designing and building my own case and custom watercooling set up for it for Quake con’s annual case modding competition… So if you see a manta ray as a computer with a pumpless watercooling system then you’ll know I’ve finished…
Also my main computer was a threadripper but threadripper got too costly for me and honestly 16 cores is all I “need” but I do miss the memory bandwidth and PCIe connectivity.
2 Likes
Hey, random question: do we think resizeable BAR affects compatibility of dGPUs? I’m debating whether to update my dGPU BIOS.
compatibility? no, without it the gpu will still boot and work… however some newer cards, especially intel’s, will perform terribly without it. Right?
1 Like
So I know my use case was a bit different than why others wanted to use a dGPU with the Framework desktop, but I wanted to share a video of it working.
This is me running Cyberpunk’s benchmark on the iGPU, then alt-F4-ing to open up the VM to show armored core 6 running. You can see the GPU information in linux on the left side and you can see the gpu information in the VM when I open up device manager.
Edit: The Intel dGPU in the VM even shows rebar enabled. I’m not sure why the framerate in the cyberpunk demo was a bit low at the start, I close it right as it came up to the normal framerate. And the stutters in Armored core at the beginning probably have to do with the fact that I only had 2 cores in the VM
And to be clear, I don’t use the VM for games… but that is the easiest way to showcase that it’s working. I use the VM for Autodesk Fusion and a few other windows only applications.
2 Likes
I might be a noob. but i just got my framework desktop like 2 days ago. 128GB option. ive tried configuring different splits of cpu vs gpu ram allocation in bios, but it never gets seen by any OS. ive tried omarchy, fedora, ubuntu, cachy. please help i feel dumb =)
You might have better luck making a different topic or posting in a different topic, this one is specifically for getting a dedicated GPU working in addition to the integrated one.
i did make a specific post about it. 0 hits. lol.