Aulafy
Courses/Claude Code + Local AI/Ollama Is Not Using the GPU on Windows

Ollama Is Not Using the GPU on Windows

If Ollama responds slowly and your processor hits 100%, the model is probably running on the CPU. This guide gives you a structured diagnosis for NVIDIA, AMD, WSL2, and Docker without randomly tweaking settings.

  • Check whether Ollama is using the GPU or CPU.
  • Review drivers, compatibility, and VRAM on Windows.
  • Decide when to use native Windows, WSL2, LM Studio, or a smaller model.

Quick diagnosis

Open PowerShell and try this while generating a long response in Ollama:

TerminalCode
ollama ps
ollama run qwen3:4b "Escribe una explicación larga sobre IA local"

# En otra terminal, si tienes NVIDIA:
nvidia-smi -l 1

Read the logs before changing anything

The logs usually say whether Ollama found a GPU, fell back to CPU, or if a driver failed during detection.

TerminalCode
# PowerShell
Get-ChildItem "$env:LOCALAPPDATA\Ollama" -Recurse -Filter "*.log"

# Abre el log más reciente:
notepad "$env:LOCALAPPDATA\Ollama\server.log"

Look for words like cuda , rocm , vulkan , gpu , fallback , memory or no compatible GPUs . If nothing about the GPU appears, Windows isn't even presenting it properly to Ollama.

NVIDIA checklist

  • Update the NVIDIA driver. Ollama documents support for NVIDIA GPUs with compatible compute capability and recent drivers.
  • Restart Windows after installing the driver.
  • Verify that nvidia-smi works in PowerShell.
  • If you have a hybrid laptop, force the dedicated GPU for Ollama from Windows Graphics settings or the NVIDIA Control Panel.
  • Try a small model to rule out insufficient VRAM.
TerminalCode
nvidia-smi
ollama pull qwen3:4b
ollama run qwen3:4b "Responde con 20 frases para probar rendimiento"

NVIDIA + Intel hybrid laptops

This is the trickiest case: Windows may launch Ollama with the Intel iGPU even if you have a dedicated NVIDIA GPU.

  • Open Settings → System → Display → Graphics.
  • Add the Ollama app if it doesn't appear.
  • Set High performance to use the dedicated GPU.
  • In the NVIDIA Control Panel, use High-performance NVIDIA processor for Ollama if your system allows it.
  • Quit Ollama from the system tray and reopen it.
TerminalCode
# Comprueba antes y después:
nvidia-smi -l 1
ollama run qwen3:4b "Haz una prueba larga de rendimiento"

AMD Radeon checklist

Ollama for Windows includes AMD Radeon support, but practical compatibility depends heavily on the GPU, driver, and available backend.

  • Update AMD Adrenalin and restart.
  • Try native Ollama on Windows first, not Docker.
  • If your iGPU or APU doesn't accelerate well, try LM Studio with Vulkan on that machine.
  • On Linux, check the ROCm version and drivers; if they're outdated, Ollama may fall back to CPU.

Vulkan as plan B for AMD, iGPU, and unusual setups

If your GPU doesn't work through CUDA or ROCm, Vulkan can be a useful path on some machines. Don't treat it as a universal guarantee: test it and measure.

TerminalCode
# PowerShell: variables persistentes para tu usuario
setx OLLAMA_VULKAN 1
setx OLLAMA_IGPU_ENABLE 1

# Cierra Ollama completamente, abre una terminal nueva y prueba:
ollama run qwen3:4b "Prueba de Vulkan en Ollama"

Windows Defender can slow down models

Models are huge files. On some machines, Defender may scan every download or read and make it feel like Ollama is broken.

TerminalCode
# Ruta habitual de modelos:
%USERPROFILE%\.ollama

# PowerShell:
explorer "$env:USERPROFILE\.ollama"

Add that folder to Windows Security exclusions only if you understand the risk and download models from trusted sources. Don't exclude generic folders like Downloads or your entire user profile.

WSL2 or native Windows

For most people, native Windows is simpler. WSL2 makes sense if you already work in Linux, Docker, or backend development.

TerminalCode
wsl --status
wsl --shutdown

# Dentro de Ubuntu/WSL, si tienes NVIDIA:
nvidia-smi

Docker on Windows

If you run Ollama in Docker, you need to pass the GPU to the container. Before blaming Ollama, verify that Docker can see the GPU.

TerminalCode
docker run --rm --gpus all nvidia/cuda:12.6.0-base-ubuntu22.04 nvidia-smi

If it still uses the CPU

  • Restart Ollama from the tray icon or restart Windows.
  • Update Ollama to the latest version.
  • Try a smaller model or a lighter quantization.
  • Check free VRAM before launching the model.
  • On a hybrid laptop, plug in the charger and enable high-performance mode.
  • Compare with LM Studio if you have an iGPU or AMD and need easy Vulkan offload.