Ollama running painfully slow, your CPU at 100% and the GPU doing nothing? That means Ollama is not using your graphics card, or is only using part of it.
It is one of the most common Ollama problems, and it almost always comes down to one of a few causes: an old driver, an unsupported GPU, a model that does not fit in VRAM, or a Docker or Linux setup issue.
This guide shows you how to check what Ollama is really using, then walks through every fix for NVIDIA, AMD, Intel and Apple, on Windows, Linux, Mac and Docker. Requirements and commands are taken from Ollama’s official hardware and troubleshooting docs as of November 2026.
- Check first: run a model, then
ollama ps. The PROCESSOR column should say 100% GPU. - Shows a CPU/GPU split? The model plus context does not fit in VRAM: use a smaller model or quant, or lower
num_ctx. - NVIDIA: needs compute capability 5.0+ and driver 550+ (570+ for compute 5.0 to 6.2 cards). Update the driver and restart Ollama.
- AMD: needs a ROCm v7 driver; Windows support covers RX 7000-series and Radeon PRO W7000 cards. Vulkan adds support for more GPUs.
- Docker: install the NVIDIA Container Toolkit and run with
--gpus all. - Read the log:
%LOCALAPPDATA%\Ollama\server.log(Windows),~/.ollama/logs/server.log(Mac),journalctl -u ollama(Linux).
Step 1: Check If Ollama Is Using Your GPU
- Run a model in one terminal:
ollama run llama3.2 "hello" - In a second terminal, run:
ollama ps - Read the PROCESSOR column:
- 100% GPU: perfect, the whole model is on the GPU.
- A CPU/GPU split (for example 40%/60%): part of the model spilled into system RAM. Go to Fix 1.
- 100% CPU: Ollama did not find or cannot use your GPU. Go to the fix for your GPU brand.
- Watch GPU use while it answers:
nvidia-smi(NVIDIA), Task Manager > Performance > GPU (Windows), Activity Monitor > GPU History (Mac), orrocm-smi/amd-smi(AMD on Linux).
Common Causes at a Glance
| What you see | Most likely cause | Fix |
|---|---|---|
CPU/GPU split in ollama ps |
Model or context too big for VRAM | Smaller model/quant, lower num_ctx |
| 100% CPU on NVIDIA | Old driver or unsupported card | Driver 550+ (570+ for older cards), restart |
| 100% CPU on AMD | Card not on ROCm list or old driver | ROCm v7 driver, Vulkan, or override (Linux) |
| Worked, then CPU after sleep (Linux) | NVIDIA UVM driver bug | Reload nvidia_uvm |
| CPU only in Docker | No GPU passed to container | NVIDIA Container Toolkit + --gpus all |
| Uses weak integrated GPU | Wrong GPU picked | Select GPU with environment variables |
| Slow on Intel Mac | No Metal acceleration on Intel Macs for most models | Use smaller models or an Apple Silicon Mac |
Fix 1: Model Does Not Fit in VRAM (Partial GPU)
This is the number one cause of “Ollama is slow”. Ollama loads as many layers as fit into VRAM and runs the rest on the CPU, which can be 5 to 20 times slower.
Remember that VRAM is used by the model plus the context (KV cache). Ollama’s default context is 4,096 tokens, and every increase costs more memory. Running several requests in parallel multiplies the context memory too.
- Pick a smaller quant: try
model:q4_K_Minstead of q8 or fp16. - Pick a smaller model: an 8B instead of a 14B on an 8 GB card.
- Lower the context: in a chat type
/set parameter num_ctx 4096, or set it in a Modelfile, or lowerOLLAMA_CONTEXT_LENGTHon the server. - Unload other models:
ollama stop model-name, and setOLLAMA_MAX_LOADED_MODELS=1if two models fight for VRAM. - Close other GPU apps such as games, ComfyUI or video editors.
- Check again with
ollama ps.
| VRAM | Fits fully on GPU (Q4, ~4K context) |
|---|---|
| 6 GB | Up to about 7B to 8B (tight) |
| 8 GB | 7B to 8B comfortably |
| 12 GB | Up to about 14B |
| 16 GB | 14B with more context, some 20B-class |
| 24 GB | 27B to 32B |
Exact numbers vary by model; our VRAM guide has full tables.
Fix 2: NVIDIA GPU Not Detected
| Requirement | Value |
|---|---|
| Compute capability | 5.0 or newer (GTX 750 Ti and later) |
| Driver | 550 or newer |
| Older cards (compute 5.0 to 6.2, e.g. GTX 900 and 10-series) | Driver 570 or newer |
| RTX 50-series (compute 12.0) | Supported; use a current driver |
- Check your card is supported on NVIDIA’s CUDA GPUs page (compute capability 5.0+).
- Update the driver from NVIDIA’s site or the NVIDIA app. Run
nvidia-smiand confirm the driver version. - Fully restart Ollama: quit it from the system tray (Windows/Mac) or
sudo systemctl restart ollama(Linux). - Update Ollama to the latest version, which bundles its own CUDA libraries.
- Linux after sleep: if the GPU vanished after suspend/resume, run
sudo rmmod nvidia_uvm && sudo modprobe nvidia_uvm, or reboot. - Linux UVM driver missing: run
sudo nvidia-modprobe -u. - Check the log for error codes like 3, 46, 100 or 999 (see below).
Fresh install? Follow our Ollama install guide first.
Fix 3: AMD Radeon Not Used
Ollama supports AMD GPUs through ROCm, and newer releases need a ROCm v7 driver stack.
| Platform | Supported cards (ROCm) |
|---|---|
| Windows | RX 7900 XTX, 7900 XT, 7900 GRE, 7800 XT, 7700 XT, 7600 XT, 7600; Radeon PRO W7900 to W7500 |
| Linux | RX 9070 XT/9070/9060 series, RX 7000 series, RX 6800 to 6950 XT, Radeon AI PRO R9700, Ryzen AI Max and Ryzen AI 9 APUs, Instinct cards |
- Update the AMD driver. On Linux, install ROCm v7 with AMD’s
amdgpu-installtool, then reboot. - Linux log shows “failed to finish discovery before timeout”? Your driver is too old for the bundled ROCm 7 libraries; upgrade it.
- Linux permissions: add your user (and the
ollamauser) to thevideoandrendergroups. - Unsupported card on Linux: try
HSA_OVERRIDE_GFX_VERSIONwith the closest supported target, for exampleHSA_OVERRIDE_GFX_VERSION="10.3.0"for RX 6000-class cards not on the list. This is experimental. - Still nothing? Use Vulkan (Fix 4), which covers many cards ROCm does not, including RX 9000 series on Windows.
Thinking about an AMD card for AI? Read our RX 9070 XT for AI review.
Fix 4: Intel and Other GPUs via Vulkan
Ollama also supports GPUs through Vulkan on Windows and Linux, which helps Intel Arc and AMD cards outside the ROCm list.
- Windows: most GPU drivers already include Vulkan; update your driver.
- Linux: install Vulkan drivers (Mesa or vendor packages). Intel and AMD have install guides.
- Better VRAM detection on Linux: Vulkan needs extra rights to report free VRAM. Run
sudo setcap cap_perfmon+ep /usr/local/bin/ollama, or Ollama will estimate model sizes. - Choose a Vulkan GPU: set
GGML_VK_VISIBLE_DEVICESto the GPU number. On laptops with an unstable integrated GPU, point it at the dedicated GPU. - Turn Vulkan off if it causes problems:
OLLAMA_VULKAN=0.
Fix 5: Mac (Apple Silicon and Intel)
- Apple Silicon (M1 and newer): Ollama uses the GPU through Metal automatically. If it feels slow, the model is too big for your unified memory; use a smaller model or quant.
- Check memory pressure in Activity Monitor. Yellow or red means the Mac is swapping.
- Do not run under Rosetta. Install the native Apple Silicon version of Ollama.
- Intel Macs have very limited GPU acceleration, so expect CPU speeds.
Fix 6: Ollama in Docker Uses CPU
- Install the NVIDIA Container Toolkit on the host (Linux, or WSL2 on Windows).
- Test it:
docker run --gpus all ubuntu nvidia-smi. If this fails, Ollama cannot see the GPU either. - Start Ollama with the GPU:
docker run -d --gpus=all -v ollama:/root/.ollama -p 11434:11434 --name ollama ollama/ollama - AMD: use the
ollama/ollama:rocmimage and pass--device /dev/kfd --device /dev/dri, plus--group-addwith the numeric group IDs of those devices. - Works, then drops to CPU later? Add
"exec-opts": ["native.cgroupdriver=cgroupfs"]to/etc/docker/daemon.jsonand restart Docker. - SELinux blocking AMD devices:
sudo setsebool container_use_devices=1.
Full Compose files are in our Ollama Docker GPU guide.
Fix 7: Laptops and Multiple GPUs
- Windows laptops: in Settings > System > Display > Graphics, set Ollama to “High performance” so it uses the NVIDIA or AMD GPU, and plug in the charger.
- Pick NVIDIA GPUs: set
CUDA_VISIBLE_DEVICESto GPU UUIDs fromnvidia-smi -L(more reliable than numbers). - Pick AMD GPUs: set
ROCR_VISIBLE_DEVICES(list them withrocminfo). - Force CPU on purpose: set the variable to an invalid ID such as
-1. - Gibberish across several AMD GPUs on Linux: a known ROCm multi-GPU issue; try one GPU first.
How to set environment variables: on Windows, quit Ollama, add the variable under “Edit environment variables for your account”, then start Ollama again. On Linux, run sudo systemctl edit ollama and add Environment="NAME=value" under [Service], then restart.
Read the Ollama Logs
| System | Where to find the log |
|---|---|
| Windows | Win + R > explorer %LOCALAPPDATA%\Ollama > server.log |
| Mac | cat ~/.ollama/logs/server.log |
| Linux (systemd) | journalctl -u ollama --no-pager --follow --pager-end |
| Docker | docker logs ollama |
For more detail, set OLLAMA_DEBUG=1 and restart. On Windows: quit Ollama from the tray, then in PowerShell run $env:OLLAMA_DEBUG="1" followed by & "ollama app.exe". Look for lines about GPU discovery, the chosen library and how many layers were offloaded.
Error Messages and What They Mean
| Log message or code | Meaning | Fix |
|---|---|---|
| Code 3 (not initialized) | NVIDIA driver not ready | Reload nvidia_uvm, reboot, update driver |
| Code 46 (device unavailable) | GPU busy or in a bad state | Close other GPU apps, reboot |
| Code 100 (no device) | No usable NVIDIA GPU found | Check driver, container GPU access |
| Code 999 (unknown) | Driver error, often after sleep | Reload nvidia_uvm or reboot |
| “failed to finish discovery before timeout” | AMD driver too old for ROCm 7 | Upgrade to ROCm v7 driver |
| Permission denied on /dev/kfd | User not in video/render group | Add groups, log out and back in |
| Offloaded 0/NN layers to GPU | GPU found but no room, or not supported | Free VRAM, smaller model, check support lists |
My Take
Nine times out of ten, “Ollama not using GPU” is really “model does not fit”. Run ollama ps, and if you see a CPU/GPU split, drop to a Q4 quant or a smaller model before touching drivers.
If it says 100% CPU, update the GPU driver, restart Ollama completely and read the server log. That log tells you exactly why the GPU was skipped.
FAQ
How do I know if Ollama is using my GPU?
Run ollama ps while a model is loaded. The PROCESSOR column shows 100% GPU, a CPU/GPU split or 100% CPU.
Why is Ollama using CPU instead of GPU?
Usually the model does not fit in VRAM, the driver is too old, the GPU is not supported, or Docker has no GPU access.
Which NVIDIA driver does Ollama need?
Driver 550 or newer, and 570 or newer for older cards with compute capability 5.0 to 6.2.
Does Ollama work with AMD GPUs on Windows?
Yes, with a ROCm v7-capable driver on supported RX 7000-series and Radeon PRO W7000 cards. Vulkan support covers more GPUs.
Can Ollama use Intel Arc GPUs?
Yes, through Vulkan on Windows and Linux, with current drivers.
Can Ollama use two GPUs?
Yes. It can split a model across several GPUs. Use CUDA_VISIBLE_DEVICES or ROCR_VISIBLE_DEVICES to choose which.
How do I force Ollama to use CPU only?
Set CUDA_VISIBLE_DEVICES (or ROCR_VISIBLE_DEVICES) to -1 before starting the server.




