Ollama not using your GPU on Windows? Start here.

Before reinstalling drivers or changing environment variables, establish what Ollama and Windows can actually see. These checks help narrow down the layer causing the problem.

This guide is diagnostic, not a guarantee that every GPU is supported. GPU support depends on the hardware, driver, Ollama version and model configuration.

1. Check what Ollama reports

Start a model you already have, then open a second PowerShell window and run:

ollama ps

Inspect the PROCESSOR column if your version shows it. It can indicate CPU, GPU, or a split. A mixed result is not automatically a fault: model layers can be split when available memory is limited.

2. Check that Windows sees your GPU

For NVIDIA

nvidia-smi

If this command fails, first check the NVIDIA driver installation and whether the GPU is visible to Windows. Do not download individual CUDA DLLs from unofficial sites to patch the error.

For AMD or Intel

Get-CimInstance Win32_VideoController | Select-Object Name, DriverVersion

This lists display adapters and driver versions. It does not prove that Ollama can use a particular adapter; use the official hardware and platform documentation to check support.

3. Restart after a driver update

Quit Ollama from the Windows system tray, reopen it, run a model and repeat ollama ps. If the problem remains, record the Ollama version and exact error before making more changes:

ollama --version

4. Consider memory pressure

A model can exceed available GPU memory, causing partial offload or other performance changes. Try a smaller model you already have installed and compare. Don't assume that every model labelled “7B” or “8B” has the same memory requirement; quantization, context length and architecture matter.

5. Check official sources before changing settings

Ollama GPU documentation
Ollama for Windows
Ollama issue tracker — search for your exact error and version.

What to include when asking for help

Share the Ollama version, GPU model, driver version, the output of ollama ps, and the exact error. Review output first and remove private paths, usernames, tokens, prompts or other sensitive information.