Thread 106342137

Got my Framework Desktop set up (395 model, I don't know if any of this applies to other models).
You need r8169 for the Ethernet driver.
You need to use F2 to get into the Administer Secureboot page to turn that off if you want to do that. Booting from the SD card module works, but it wasn't even showing up until I turned secureboot off.
You can set 96 GB VRAM, but you need to turn the option off of auto, I forgot what it was called but it's right next to the 0.5 GB that it's set on by default. Only 0.5 GB, 32 GB, and 64 GB will show up until you set it to custom.
I don't know how to get the wifi working, it uses a different chip from the HP Z2 Mini G1a so I need to figure that out still.
Mine came with 2 heatsinks for the NVME drives even though I didn't order any. The HP one will not come with heatsinks if you don't populate the NVME slots.
You need to unscrew the fan in the HP workstation to get at the NVME drives. I think you can get to the drives without tools in the Framework but you'll still need a screwdriver (the hex driver) to replace the drives.
Do what you will with this information, maybe it will save you some time if you bought one of these things.

I love framework

>>106342137 (OP)
Correction about the wireless. It's a Mediatek MT925B22M wifi card.
Actually, the HP one also used a Mediatek card. I just booted it up again and it uses the same mt7925e driver. I had issues with mixing the 2025 version of the firmware that came with Ubuntu with Linux 6.17.0-rc1, and I had to download a 2024 version of the firmware from github. The wlan0 interface just went into hard block mode whenever I tried to bring it up. The firmware/driver are working with this Framework desktop too.
Ubuntu was working out of the box, in case you use that instead of trying to use some custom OS.
There's an NVME slot in the back and the front of the motherboard, both are accessible without tools but require the hex driver to replace.

>>106342190
I personally had good luck with the support, but I've seen complaints about it. No idea if those were legitimate or just trying to get free replacements though.

>>106342137 (OP)
>>106342291
Oh right. I tried to change the GTT or whatever module load parameter to try and make the shared video memory higher, but when set to 512 MB dedicated VRAM it doesn't look like it allows you to use more than 64 GB of dynamic video memory. I think that if you want to use 96+GB of video memory then you'll need to set the dedicated memory to 96 GB in the BIOS, unless you're able to get rid of that shared memory limitation by modifying the driver. It would be nice to have the full 128 GB available for the CPU, then 96 GB available for the GPU if you need it instead of limiting the CPU to 32 GB at all times, but I'm not sure if that's possible.

why wouldn't you just buy the mainboard and put it in a $40 case with a normal ATX power supply

isn't the case + psu like $300 if you break it down by component? that's ridiculous

>>106342426
Yes, you can just buy the mainboard if you want. I got the full thing because I didn't want to get a power supply or case. I'm lazy and make a lot of money so I don't care.

Should I buy a framework 13?

>>106342584
I don't think that the thermals are very good. If I was going to get a laptop I'd probably get the HP ZBook Ultra G1a.

no one has ever set up a computer before
thank you for saving us

>>106342676
Glad to help anon. I didn't include instructions for setting up the initramfs and putting modules in there since I'm assuming anyone doing that would know how to do it already.

Use case?
Just get a normal desktop instead of an overpriced ITX meme

>>106342853
>normal desktop
I have a normal desktop. This has 96 GB VRAM and 2x40 gbps Thunderbolt so I can cluster them together. It's also ~180 watts max so I don't need to worry about trying to stop my breaker from tripping.

My desktop is more modular and upgradable than most modern desktops including this.
What's the use case for this for real?

>>106343091
What all did you do to build it? This is essentially a prebuilt. I installed the fan and if I was using Windows then it would be 100% plug and play after that. This is a low power AI computer with a retard tax. It's basically a more affordable (and higher performance from what I've seen for compute-heavy workloads) Mac Studio.

>>106343091
the ai max 395 is not available in a socket, so you need to buy one of these oem reference model rebrands to get it because amd is gay now
computing sucks in general

>>106343364
>ai max 395 is not available in a socket
It's supposed to be a mobile chip. Are any of these AI things supposed to be desktop chips?

>>106342891
Does Thunderbolt actually support PC-to-PC connections? How do you cluster PCs over Thunderbolt, like with what software supports that and for what use case?

>>106343091
Use case for this sort of thing seems to be getting a lot of memory on a GPU on the cheap. This APU in general just seems to be a high-end mobile chip with an especially fast iGPU, so beyond this kind of thing you'd probably use it for various kinds of portable gaming devices.

>>106343415
I haven't tried it in Linux yet, but with Windows you just plug the computer into each other and it creates a 40 gbps network connection. I have the thunderbolt/USB4 and networking drivers set up, I just haven't tested it yet, but it works fine if you load the modules (or probably if the distro you're using has whatever services running to automate that).

>>106343428
>various kinds of portable gaming devices
The actually mobile ones start at like $3k and use like 150+ watts. Not sure how much of a portable gaming device you'd be able to make, but I guess you could drop the wattage by a lot and still come out ahead of the competition for performance.
I think Asus made some tablet or something? There's also the HP laptop. I thought GPD or whatever was also making a handheld with this chip as an option, but it's going to be expensive as hell.

>>106342137 (OP)
looks cool op, please report back on large llm performance with it, curious if its worth it

>>106343438
Yeah there's an Asus tablet, a few days ago I read that One X or whatever they're called are also making a tablet. GPD was making a small handheld with a fucking external battery that clips on the back.

>>106343452
It's not worth it if all you want to do is run LLMs. There was a page that had all sorts of different environments (ROCm, Vulkan, flash attention options) for different models, but I don't know where that went. I think the link was posted on one of the older /lmg/ threads, but I lost it. The compute is good, but the memory bandwidth is like 256 GB/s or something, and that's the same whether it's dedicated VRAM or not.
Here are some resources with some benchmark data. I don't know what the second one is, but it might contain info.
https://github.com/lhl/strix-halo-testing/tree/main/llm-bench
https://strixhalo-homelab.d7.wtf/

>>106343502
>GPD was making a small handheld with a fucking external battery that clips on the back
It will probably still get 30 minutes of runtime, assuming it doesn't catch fire.

>>106343452
>>106343512
It was in my history:
https://kyuz0.github.io/amd-strix-halo-toolboxes/

>>106343512
I'm pretty sure you can run LMStudio on it and nothing else so it has very limited use case.

>>106343549
ONNX Runtime (ROCm) works with it, so I have lots of use cases.

>>106343512
I'm pretty sure you can run LMStudio on it and nothing else so it has very limited use case.
AMD might officially support this chip with ROCM in the future so less fiddling will be required.

>>106343570
ROCM seems to have some issues with this chip for now. Vulkan compute runs perfectly, at least on Windows.

>>106342137 (OP)
>>106342137 (OP)
>>106342137 (OP)
ANON I need Wan2.1 14B 720P GGUF Q8 benchmarks, stat! Or whatever the biggest WAN model you can fit in there is.

>>106343599
Does that run in LM Studio?

>>106343616
I don't know! I usually run WAN models with https://github.com/deepbeepmeep/Wan2GP because I'm a poorfag and I have no idea if it'll work!

>>106343632
fucking python, I hate python more than anything

>>106343599
>>106343632
I'll spend the next week trying to get that to work. I'll put results in the /lmg/ thread if I can get it working. This thread was just to put details about the issues I had with setup.