My framework desktop mainboard (128 GB, batch 18) has a strange memory allocation issue:
The BIOS does not provide an option to set more than 24 GB of iGPU memory. In Windows, I can set 96 GB in the Adrenalin driver, after this, if I go back to BIOS/Advanced I see this message: “Question value mismatch with Option value! (iGPU Memory Size: mismatch value is 98304.)” but still no memory options over 24 GB.
And also if I try to set any custom memory values when iGPU Memory Configuration is set to Custom, I get this message after saving and rebooting: “Cannot configure preferred UMA size setting correctly, so UMA mode will be changed to AUTO”
Support is clueless, they already replaced my mainboard. The two mainboard behave exacly the same.
I tried with the following:
- updating BIOS multiple times (EFI Shell and from Windows), currently I’m at latest, 3.04 version (shipped with 3.03)
- mainboard reset
- RTC battery reset
- I even switched my power supply (currently I’m using a Seasonic Focus GX-750 ATX 3.1)
- OS-s I tried: Ubuntu 25.10, Windows 11, Fedora Live
This definitely needs to be looked at. Based on what you’re experiencing, this is something our Support team is best equipped to look into, since they can check your hardware details and guide you through the right steps.
To get in touch with them, you can do the following:
From the forums, click Support at the top of the page.
Scroll all the way down.
Click Submit support request and fill out the form.
Then Submit it at the bottom.
Once that’s in, the Support team will review everything and get back to you as soon as possible. If anything else comes up or you need help finding the right page, just let us know.
It’s very counterintuitive, to allow use of the maximum possible GPU memory, you need to set it to “Auto” (512MB) in the BIOS. This works for Linux but I’m not sure if Windows supports it yet.
iGPU memory limit is set by firmware and hardware, BIOS cannot exceed 24 GB on your 128 GB desktop board. Windows/Adrenalin can map more dynamically, but BIOS only shows max allowed. Custom UMA settings fail because firmware enforces limits. No software workaround; contact Framework support with full logs to check if board revision or BIOS update can lift the limit.
To elaborate on this a bit, the amd driver has GTT for shared memory architectures, which controls the amount of shared system memory which may be used by the GPU if and when it needs it.
So in Linux, often setting a large amount of “vram” (and I put vram in quotes because it’s all the same memory) can even be inefficient if your apps run in gtt and effectively leave the “vram” pool untouched.
I believe by default GTT is set to 50% of system memory, so 64GB on a 128GB system, but you can increase this using kernel parameters amdttm.pages_limit and amdttm.page_pool_size, where you set both to desired memory in GB x 1024 x 1024 x 1024 ÷ 4096.
I would think the only reason you’d need to set high amount of vram in bios while using Linux is if you have an application that has hard coded vram checks to run and is not gtt aware. Ollama for example used to be this way, but newer versions are gtt aware so it now works fine with amd shared memory platforms. Otherwise it’s better to set vram to minimum and let the driver handle memory management.