LocalHake

Tutorial

Strix Halo + ROCm 7.1 + Ubuntu 24.04

Install ROCm 7.1 on the AMD Strix Halo (GMKtec EVO X2, AI MAX+ 395) running Ubuntu 24.04 with kernel 6.14 upgrade, AMD GPU drivers, and shared memory (TTM) configuration.

Hake HardwarePublished Updated ~6 minintermediate

This post contains paid links (affiliate) — how that works.

▶ Watch this build on YouTube
Contents

I picked up the GMKTec EVO-X2 with the AI MAX+ 395 and the initial experience with the AMD platform for AI was rough. But things seem to already be improving with ROCm 7.1.

This guide focuses on Ryzen AI CPUs. If you are running an older AMD integrated GPU like the 780M or 680M, this will not work. You'll still have to go with the unsupported methods of ROCm if you want to proceed, but those work pretty well as well. It's nice to see the EVO X2s actually having some support out of the box for this type of stuff and not have to do any janky methods.

Installation

Update the Kernel (Required)

bash
sudo apt update && sudo apt-get install linux-oem-24.04c

Reboot:

bash
sudo reboot now

Confirm the kernel update:

bash
uname -r

You should see version 6.14.0. Then perform a full system update:

bash
sudo apt update && sudo apt upgrade -y

Download and Install ROCm

Download the installer:

bash
wget https://repo.radeon.com/amdgpu-install/7.1/ubuntu/noble/amdgpu-install_7.1.70100-1_all.deb

Install it:

bash
sudo apt install ./amdgpu-install_7.1.70100-1_all.deb

You may see a permission warning — this is safe to ignore. Now install ROCm:

bash
amdgpu-install -y --usecase=rocm --no-dkms

If you also have a monitor hooked up and want a graphical user interface with AMD drivers, you can add graphics to the usecase parameter (e.g., --usecase=rocm,graphics). I'm just doing ROCm for my headless system.

This is a large install — around 23 GB — so it does take a little while (about 5-7 minutes).

Add your user to the required groups:

bash
sudo usermod -a -G render,video $LOGNAME

Reboot:

bash
sudo reboot now

Validate Installation

Run the validation command:

bash
rocminfo

Look for Agent 2 in the output. You should see AMD Radeon Graphics with the GFX 1151 or 1150 identifier.

Terminal output of rocminfo showing Agent 2 with gfx1151 GPU, Marketing Name AMD Radeon Graphics, Vendor AMD

Configure Shared Memory

Install pipx:

bash
sudo apt install pipx

Add pipx to your path:

bash
pipx ensurepath

Exit and re-open your terminal, then install AMD debug tools:

bash
pipx install amd-debug-tools

Check current settings:

bash
amd-ttm
Terminal output of amd-ttm showing current TTM pages limit of 16311946 pages (62.23 GB) and total system memory of 124.45 GB

The default allocation is approximately 62.23 GB. In future releases of Ubuntu, this will likely be handled automatically, but for now we're going to set this manually.

I'm going to use 115 GB for the shared memory. If you go over 112, you get a warning saying that you have more than 90% assigned to it. Lots of people have run 110 GB just fine. If you're using this system for other things as well — like a Ubuntu desktop you're daily driving — you probably don't want to go this high. Maybe around 100 or 110. I'm only using this for AI, so I just want to dedicate as much as I can to it.

bash
amd-ttm --set 115
Terminal output of amd-ttm --set 115 showing a warning that 115 GB exceeds 90% of system memory, followed by successful configuration written to /etc/modprobe.d/ttm.conf

Reboot for changes to take effect:

bash
sudo reboot now

Verify the new settings:

bash
amd-ttm
Terminal output of amd-ttm showing current TTM pages limit of 30146560 pages (115.00 GB) and total system memory of 124.45 GB after reboot

And there you have it. You are now running the latest ROCm with your TTM set correctly for maximum AI shenanigans.

Next Steps

From here you can move on to whatever AI workload you want:

  • Text inference with llama.cpp or Ollama
  • Image/video creation with ComfyUI

Gear used