Ollama Running 100% on CPU in Windows? Update to Driver 531+

If Ollama is running 100% on CPU in Windows while your NVIDIA card sits idle, the cause is almost always a graphics driver older than version 531. Ollama needs that driver floor before it will hand inference to the GPU, and once you update past it, the model loads onto the card without touching Ollama itself.

Quick Answer

Update your NVIDIA driver to version 531 or later, restart, and run ollama ps again. A driver below 531 is the single most common reason Ollama falls back to CPU on Windows. No reinstall of the models is needed, though reinstalling Ollama after the driver update guarantees the CUDA runtime is bundled correctly.

Confirm the Problem First

Before changing anything, verify what is actually happening. Open PowerShell and run ollama ps while a model is loaded. If the processor column reads 100% CPU, the GPU is not being used.

A second quick tell is raw speed. On a 7B model, a generation rate comfortably above 10 tokens per second points to GPU acceleration, while anything under about 5 tokens per second means you are on the processor. If you suspect CPU mode, run nvidia-smi in the same PowerShell window. If that command fails or reports a driver below 531, you have found your answer.

The Fix, Step by Step

Updating the driver is the whole job in most cases.

1. Download the latest Game Ready or Studio driver for your card from NVIDIA and install it. Anything from the recent release branches sits well above the 531 floor.
2. Restart Windows so the new driver loads cleanly.
3. Open PowerShell and run nvidia-smi to confirm the card and driver version are now detected.
4. Run your model again and check ollama ps. The processor column should now show the GPU carrying the load.

If Ollama was installed before the driver was working, its installer may have set up without GPU support compiled in. In that case, reinstall Ollama after nvidia-smi confirms a healthy driver. The installer then detects the card and bundles the CUDA runtime it needs.

One trap to avoid: do not install the CUDA toolkit separately to fix this. Ollama ships with the CUDA runtime it requires, so the only thing you need to manage is the NVIDIA driver. Adding a standalone CUDA install adds confusion, not capability.

For local AI work, VRAM is the spec that decides which models fit on the card, so check it on the AI PC range at Evetech before buying. The PC best sellers list is also a quick way to see which GPU-equipped machines South African builders are running for this kind of workload.

Frequently Asked Questions

Why does Ollama default to CPU at all?

Ollama checks for a usable GPU and a compatible driver at load time. If it cannot confirm both, it safely runs on the CPU rather than failing, which is why the symptom is slow output rather than an error message.

Do I need to reinstall my models after updating?

No. Models are stored separately from the runtime. Updating the driver, or even reinstalling Ollama, leaves your downloaded models untouched.

What if nvidia-smi works but Ollama still uses the CPU?

Reinstall Ollama after confirming the driver. This forces a fresh setup that detects your now-working GPU and bundles the correct CUDA runtime, which resolves the case where Ollama was installed before the driver was ready.

Does this apply to WSL2 as well?

The same 531 floor applies, but inside WSL2 you rely on the Windows host driver. Do not install a separate Linux NVIDIA driver inside WSL2, as that breaks the host driver projection and sends inference back to the CPU.