AMD Radeon 760M iGPU Acceleration in Ollama

A concise guide for enabling hardware-accelerated local LLM inference on AMD Phoenix APUs (Radeon 760M / RDNA3) under Ubuntu using ROCm overrides and systemd service tuning.

1. Driver & System Dependencies

Ensure standard Mesa drivers, Vulkan components, and direct rendering permissions are configured.

# Install Vulkan drivers and diagnostics
sudo apt update
sudo apt install -y mesa-vulkan-drivers vulkan-tools libvulkan1

# Grant render/video group permissions
sudo usermod -aG render,video $USER
sudo usermod -aG render,video ollama

2. Configure Ollama Systemd Override

By default, Ollama ignores integrated graphics to preserve system VRAM and lacks native targets for gfx1103. We override the ROCm architecture and explicitly enable the iGPU.

Open the systemd service override editor:

sudo systemctl edit ollama

Paste the following environment variables into the unit configuration:

[Service]
# Map gfx1103 (760M) to supported gfx1100 (RX 7900) target
Environment="HSA_OVERRIDE_GFX_VERSION=11.0.0"

# Force Ollama to utilize integrated GPUs
Environment="OLLAMA_IGPU_ENABLE=1"

# Memory & attention optimizations for APUs
Environment="OLLAMA_FLASH_ATTENTION=1"
Environment="OLLAMA_KV_CACHE_TYPE=q8_0"

Apply the changes and restart the daemon:

sudo systemctl daemon-reload
sudo systemctl restart ollama

3. Verification

Confirm that Ollama successfully initializes the GPU without dropping the target device:

journalctl -u ollama -b | grep -iE 'gpu|rocm|vulkan'

Performance Note: Token generation speed on integrated graphics is strictly bound by system DDR5 memory bandwidth (~50–80 GB/s) rather than GPU compute core count.

For optimal VRAM throughput, allocate 4GB to 8GB of UMA Frame Buffer memory in your system BIOS settings.

📺 Video Walkthrough: For a live demonstration and benchmarks of this setup, watch the YouTube video: Offloading Local LLMs to AMD iGPU on a Mini PC.