Step-by-Step Guide: Install AMD ROCm on Ubuntu with RX 6600 GPU

Learn how to quickly and easily set up AMD ROCm on Ubuntu for your RX 6600 GPU, enabling powerful machine learning, AI workloads, and GPU-accelerated computing right on your system.

Dedicated Servers vs Peer To Peer - CTCservers

AMD ROCm on Ubuntu

Setup Guide for RX 6600 / 6600 XT

What is AMD ROCm and Why It Matters for Machine Learning

AMD ROCm (Radeon Open Compute) is AMD’s open-source platform for GPU computing, designed to bring high-performance compute capabilities to Linux systems. Unlike traditional graphics drivers, ROCm provides a complete ecosystem for running compute-heavy workloads such as machine learning, data science, and scientific simulations directly on AMD GPUs. With ROCm, developers can leverage GPU acceleration for popular frameworks like TensorFlow, PyTorch, and HIP, enabling faster model training and inference without relying on cloud resources. This makes ROCm particularly valuable for users with AMD consumer GPUs, like the RX 6600 series, allowing them to run modern AI workloads locally with efficiency and control. By combining ROCm with Ubuntu and tools like Ollama, you can deploy large language models, experiment with AI, and harness the full potential of AMD hardware in a cost-effective and flexible way.

Running modern machine learning workloads on AMD consumer GPUs is no longer a fringe experiment. With AMD ROCm, Ubuntu, and tools like Ollama, you can run large language models such as Llama 3.1 locally on your own hardware — ideal for AI development, inference workloads, and GPU-heavy tasks without cloud costs.

This tutorial is written for users who want a clear, practical, and copy‑paste‑ready guide to set up ROCm on an AMD RX 6600 / 6600 XT GPU using Ubuntu 22.04 LTS, then install Ollama and run a real LLM.

⚠️

Important: RX 6600 GPUs are not officially supported by ROCm. This guide uses environment overrides to make ROCm work reliably. Follow each step carefully.

System Information (Reference Setup)

Below is the system configuration used for this setup. Your hardware can vary, but results are best if you stay close to this environment.

username@hostname

----------------------------

OS: Ubuntu 22.04.4 LTS x86_64
Host: redacted
Kernel: 6.5.0-45-generic
Uptime: 2 hours, 40 mins
Packages: 1820 (dpkg), 11 (snap)
Shell: bash 5.1.16
Resolution: 2560×1440
DE: GNOME 42.9
WM: Mutter
WM Theme: Adwaita
Theme: Yaru-blue-dark [GTK2/3]
Icons: Yaru-blue [GTK2/3]
Terminal: gnome-terminal
CPU: AMD Ryzen 9 6900HX with Radeon
GPU: AMD ATI Radeon RX 6600/6600 XT/
GPU: AMD ATI e8:00.0 Rembrandt
Memory: 4023MiB / 31356MiB

System Update & Headers

Install the specific headers for your currently running kernel. This is critical for driver modules to build.

BASH

sudo apt update && sudo apt upgrade -y
sudo apt install linux-headers-$(uname -r) linux-modules-extra-$(uname -r) wget curl nano -y

Set Permissions

Add your user to the render and video groups so you can access the GPU without using sudo.

BASH

sudo usermod -aG render,video $LOGNAME

Restart your computer now for these permissions to take effect.

BASH

sudo reboot

Install AMDGPU & ROCm

This step uses the correct "Jammy" (22.04) repository instead of the incompatible "Focal" (20.04) one.

1. Download the Installer:

BASH

# Example of a valid recent path (Check repo.radeon.com for exact current version)
wget https://repo.radeon.com/amdgpu-install/6.1.3/ubuntu/jammy/amdgpu-install_6.1.60103-1_all.deb

2. Run the Installer: We will install the ROCm usecase.

Note: We attempt to use the DKMS (kernel driver) first. If this step fails with "building module errors," you can try running it again with --no-dkms added.

BASH

sudo amdgpu-install --usecase=rocm

3. Configure Library Paths: Link the ROCm libraries so the system can find them.

BASH

sudo tee /etc/ld.so.conf.d/rocm.conf <<EOF
/opt/rocm/lib
/opt/rocm/lib64
EOF
sudo ldconfig

Verify Installation & Identify GPU ID

This is the most important step for your specific system (Ryzen 6900HX + RX 6600), as you have two GPUs.

1. Run rocminfo:

BASH

sudo rocminfo

If this command is not found, verify /opt/rocm/bin is in your PATH or run /opt/rocm/bin/rocminfo.

2. Find your GPU ID: Look through the output. You will see "Agent 1", "Agent 2", etc.

Look for the Agent that says "gfx1032" (This is the RX 6600).
Note if it is the first GPU listed or the second.
Usually:
- Device 0 = Integrated Graphics (Rembrandt)
- Device 1 = RX 6600

Install Ollama

BASH

curl -fsSL https://ollama.com/install.sh | sh

Configure Ollama Service (The "Override")

1. We must force ROCm to support the RX 6600 and tell Ollama exactly which GPU to use.

BASH

sudo systemctl edit ollama.service

(Note: This might open a blank file or one with comments. If systemctl edit is confusing, you can use sudo nano /etc/systemd/system/ollama.service but editing via systemctl is cleaner).

2. Paste the following block exactly. Check the ROCR_VISIBLE_DEVICES value based on your findings in Step 4.

Ini, TOML

[Service]
Environment="HSA_OVERRIDE_GFX_VERSION=10.3.0"
Environment="ROCR_VISIBLE_DEVICES=1"

HSA_OVERRIDE...: Tells ROCm to treat the RX 6600 (gfx1032) like the supported RX 6800/6900 (gfx1030).
ROCR_VISIBLE_DEVICES=1: Tells Ollama to ignore the iGPU (0) and use the RX 6600 (1). If you only have one GPU, change this to 0.

3. Save and exit (Ctrl+O, Enter, Ctrl+X).

4. Apply changes:

Ini, TOML

sudo systemctl daemon-reload
sudo systemctl restart ollama

Final Test

Run a model to verify GPU acceleration.

BASH

ollama run llama3.1

How to know if it's working: While Ollama is generating text, open a second terminal and run:

BASH

watch -n 1 rocm-smi

You should see the Power usage go up (e.g., 100W) and SCLK (clock speed) increase on the RX 6600.
If the "VRAM%" stays at 0 and only your CPU spikes to 100%, the configuration in Step 6 is incorrect.

This setup allows you to run AI models locally on AMD consumer GPUs, avoiding expensive cloud GPUs while retaining performance and control. For businesses using dedicated servers or colocation, this configuration is ideal for private inference, experimentation, and GPU‑intensive workloads.

If you are deploying this on dedicated hardware or in a data center environment, platforms like CTCservers provide the stability, bandwidth, and power reliability needed for continuous AI workloads.

Explore GPU Servers

CTCservers Recommended Tutorials

Discover CTCservers Dedicated Server Locations

CTCservers servers are available around the world, providing diverse options for hosting websites. Each region offers unique advantages, making it easier to choose a location that best suits your specific hosting needs.

North America

South America

Europe

Asia

Australia

Africa