Guides · Updated 2026-10-17

Running local AI on AMD, Intel Arc and mini PCs

Which AMD Radeon and Intel Arc GPUs work with Ollama, ROCm vs Vulkan, and how unified-memory mini PCs like the Ryzen AI Max+ 395 compare.

NVIDIA is still the path of least resistance, but AMD and Intel GPUs, plus unified-memory mini PCs, are now solid options with Ollama.

AMD Radeon: ROCm or Vulkan

Ollama supports AMD two ways:

• ROCm (fastest, officially supported cards). Ollama requires ROCm v7.
  • Linux: RX 9070 XT/9070/9060 XT/9060, RX 7900 XTX/XT/GRE, 7800 XT, 7700 (XT), 7600 (XT), RX 6950 XT/6900 XT/6800 (XT), Radeon PRO W7000/W6000 series, Radeon AI PRO R9700, and Ryzen AI chips including the Ryzen AI Max+ 395.
  • Windows: RX 7900 XTX/XT/GRE, 7800 XT, 7700 XT, 7600 (XT) and Radeon PRO W7000 series.
• Vulkan (broader, on by default). Covers many other Radeon cards. On Windows, Vulkan ships with the normal driver; Ollama recommends it as the fallback for RX 6000-series cards that don’t get ROCm v7 on Windows.

Linux tips: install ROCm with AMD’s amdgpu-install, use the latest AMD driver rather than the stock kernel one, and add the ollama user to the render group if Vulkan isn’t detected. Unsupported cards can sometimes be forced with HSA_OVERRIDE_GFX_VERSION (for example 10.3.0 for RDNA2 chips close to gfx1030). This is experimental.

Intel Arc

Intel Arc GPUs (A-series and B-series, such as the 12 GB Arc B580) run in Ollama through Vulkan. On Windows, the standard Intel driver includes Vulkan. On Linux, install Intel’s client GPU drivers and Mesa Vulkan packages. Expect it to work, but with less community troubleshooting help than NVIDIA. LM Studio’s Vulkan engine is another good option on Arc.

Mixed integrated + discrete GPUs

If Ollama picks the integrated GPU or behaves unstably, set GGML_VK_VISIBLE_DEVICES to the index of your discrete card (for example 1). To disable Vulkan entirely: OLLAMA_VULKAN=0.

Mini PCs and unified memory

On these machines the GPU shares system memory, so “VRAM” can be 64-128 GB.

Best models for unified memory: mixture-of-experts models, because they read only a few billion parameters per token: qwen3.6:35b (3B active), gemma4:26b (4B active), gpt-oss:120b (5.1B active) and qwen3-coder:30b.

CPU-only mini PCs: try gemma4:e2b, qwen3.5:2b, llama3.2:3b or phi4-mini. Expect a handful of tokens per second.

See mini PC hardware picks, or check estimates in the GPU picker.

Sources: Ollama hardware support, Ollama Windows docs.