TALK / LOCAL AI
Ollama's documentation lists GPU back ends. It does not list NPUs, so do not buy a mini PC for Ollama because of an NPU badge.
Short answer
Based on Ollama’s own hardware page, read on 2026-10-24, Ollama accelerates models on GPUs. The page covers NVIDIA (CUDA), AMD Radeon (ROCm), Apple (Metal) and a Vulkan path. It does not list an NPU as a supported back end. That is an absence in the documentation, not proof that no NPU path will ever exist, so treat an NPU as unusable for Ollama until the docs say otherwise.
What the integrated GPU can and cannot do
An integrated GPU can be used if a supported back end recognises it. On Linux, Ollama lists AMD Ryzen AI parts such as the Ryzen AI 9 HX 370 and the Ryzen AI Max+ 395 under ROCm. Additional support on Windows and Linux comes through Vulkan, which Ollama says is enabled by default once the back end is installed. Linux often needs extra Vulkan driver packages, and for AMD you may need to add the ollama user to the render group.
The model also has to fit. Ollama’s FAQ explains that ollama ps shows whether a model sits 100% in system memory (CPU) or is split between GPU and system memory. So the practical check is not the marketing badge but the Processor column on your own machine.
How to check your own mini PC
1. Find your exact CPU or GPU name in Ollama’s hardware page. 2. Install the driver stack the page names for your operating system. 3. Load a small model and run ollama ps. 4. If the Processor column says 100% CPU, the model is not using the GPU; check drivers, the Vulkan setting and available memory before blaming the hardware.
Ollama also documents switches such as OLLAMA_VULKAN=0 to disable Vulkan GPUs if an integrated GPU proves unstable, and an environment variable to force CPU use. Those exist because integrated GPUs can misbehave on some driver combinations.
A cautious buying rule
If local models are the goal, rank the checks in this order: memory capacity, then a GPU path that Ollama lists, then everything else. Write down the exact chip name from the listing and search for it on Ollama’s page. For Intel graphics on Linux, the page points to Intel’s driver instructions for the Vulkan path rather than listing individual chips. If you cannot find your chip, plan for CPU-only operation and size your expectations to match.
Boundaries
This is a software-support answer. We have not benchmarked any mini PC, so we make no speed claim. Other runtimes may use an NPU differently; that is a separate question from Ollama. Memory capacity usually limits which models you can load before the processor type does, so compare RAM first. See our Ollama mini PC guide and NPU overview for the wider picture; both are being revised.
Other perspectives
Adding to “Short answer”
Do not pay extra for an NPU if Ollama is the goal. Spend on memory capacity and a CPU or integrated GPU that the Ollama page names. Ask the seller for the exact processor model, not just the family, because Ollama lists specific chips. Treat any listing that quotes a TOPS figure as a statement about the chip, not about Ollama.
Adding to “What the integrated GPU can and cannot do”
Match three things: your chip appears on Ollama’s list (or Vulkan covers it), your operating system has the needed drivers, and memory is large enough for the model you want. On Linux the AMD list is more specific than on Windows, where Ollama lists only certain Radeon RX and PRO cards under ROCm and relies on Vulkan elsewhere. Check your operating system column before shortlisting.
Adding to “How to check your own mini PC”
Re-check after driver or Ollama updates, because support lists change. Keep a note of the version you tested and the output of ollama ps. If the model falls back to CPU after an update, roll back the driver first. For always-on use, plan cooling and power; our heating guide covers the basics. Also keep your model files on a drive with room to spare, because large models take a lot of storage.
Quick questions
- Will Ollama use an NPU in the future?
- We cannot know. The current documentation does not list one, so we do not recommend buying for it.
- Is an integrated GPU faster than the CPU for Ollama?
- Not tested here. The documentation describes support, not speed. Use ollama ps to see where your model runs, then compare on your own workload.
Hardware mentioned
Plain links to manufacturer or project pages. They are not affiliate links, and we show no prices here because prices change; see our price and data policy.
- GMKtec EVO-X2 (Ryzen AI Max+ 395, listed by Ollama under ROCm on Linux) (manufacturer or project page, read 2026-10-24)
- Minisforum MS-01 (Intel integrated graphics; check Vulkan path) (manufacturer or project page, read 2026-10-24)
What we checked
We separate documented facts from our own advice and from what we could not confirm. We did not run tests for this answer. See how we test and our editorial policy.
Facts from sources
- Ollama hardware page lists NVIDIA, AMD ROCm, Metal and Vulkan; no NPU entry (read 2026-10-24).
- Linux ROCm list includes Ryzen AI 9 HX 370 and Ryzen AI Max+ 395.
- Vulkan is enabled by default when the back end is installed; OLLAMA_VULKAN=0 disables it.
- ollama ps shows CPU/GPU split (Ollama FAQ).
Our advice
- Prioritise RAM and a listed GPU over NPU badges.
- Verify with ollama ps on your own unit.
Not confirmed
- Absence from docs is not proof of no NPU path.
- No speed measurements.
Sources (read 2026-10-24)
- Ollama documentation — Hardware support — read 2026-10-24
- Ollama documentation — FAQ — read 2026-10-24