Tutorial

Run Local LLMs with Ollama on the Radeon 780M

Use the Radeon 780M iGPU in a Ryzen 7 7840HS / 8845HS mini PC to run quantised 7B–14B models with Ollama on Linux, using ROCm and a larger UMA frame buffer.

Intermediate ⏱ 30 minutes 5 steps

What you need

  • Ryzen 7 7840HS / 8845HS (Radeon 780M) mini PC
  • 32 GB+ dual-channel DDR5 (64 GB recommended)
  • Ubuntu 24.04 LTS or newer
  • ~20 GB free disk space for models

1Raise the UMA frame buffer in BIOS

The iGPU uses part of system RAM as VRAM. Look for "UMA Frame Buffer Size" (often under Advanced → AMD CBS → NBIO → GFX Configuration) and set it as high as your board allows — 8 GB or 16 GB with 32 GB+ installed. Leave enough for the OS.

2Install Ollama

The official install script sets up the ollama binary and a systemd service. It bundles the ROCm libraries it needs for AMD GPUs.

Shell
curl -fsSL https://ollama.com/install.sh | sh
sudo usermod -aG render,video $USER

3Tell ROCm to treat the 780M as a supported GPU

The 780M reports itself as gfx1103, which ROCm does not officially list. Overriding the GFX version to 11.0.0 makes ROCm use its gfx1100 kernels, which work on the 780M. Add the override to the Ollama service.

Shell
sudo systemctl edit ollama
Override contents
[Service]
Environment="HSA_OVERRIDE_GFX_VERSION=11.0.0"
Restart
sudo systemctl daemon-reload
sudo systemctl restart ollama

4Pull a model and confirm the GPU is used

Start with an 8B model at Q4 quantisation. While it runs, ollama ps shows how much of the model sits on the GPU; the service log mentions the detected AMD GPU.

Shell
ollama run llama3.1:8b "Explain OCuLink in two sentences."
ollama ps
journalctl -u ollama --no-pager | grep -i -E "amdgpu|rocm|gfx"
NoteIf ollama ps shows 100% CPU, the override did not apply — check systemctl show ollama | grep HSA and that your user is in the render group.

5Size models to your memory

Token generation on an iGPU is limited by memory bandwidth, not compute. As a rule of thumb, a Q4 model needs roughly 0.6 GB per billion parameters plus context. 7B–8B models are comfortable; 14B works with 32 GB+; beyond that, consider an OCuLink eGPU.