Is Ollama Using My GPU? Supported Cards and How to Check
Note: Some future links on this site may be affiliate links. We never claim hands-on testing; guides are researched from the public sources listed in each article.
How we wrote this: A summary of Ollama's official Hardware support page and FAQ (linked below), checked October 29, 2026. Supported cards and driver versions change with Ollama releases, so check the live page before you buy or upgrade anything.
Local models feel fast or painfully slow mostly depending on one thing: whether they run on your graphics card or fall back to the CPU. Here is what Ollama's own documentation says about which GPUs work and how to tell what is actually happening.
Step 1: Check whether your model is on the GPU
Ollama's FAQ gives a one-line check. With a model loaded (for example, right after you chat with it), run:
ollama ps
Look at the PROCESSOR column:
| What it shows | What it means (per the FAQ) |
|---|---|
100% GPU |
The model was loaded entirely into the GPU |
100% CPU |
The model was loaded entirely in system memory |
48%/52% CPU/GPU (for example) |
The model was split between system memory and the GPU |
A split usually means the model is bigger than your card's memory. A smaller model or a smaller quantization (see what a GGUF file is) can bring it back to 100% GPU. A long context length also uses memory.
NVIDIA
The hardware page says Ollama supports NVIDIA GPUs with compute capability 5.0 or higher and driver version 550 or newer. Cards with compute capability 5.0 through 6.2 need driver 570 or newer. NVIDIA's own CUDA GPUs list (linked from Ollama's page) tells you your card's compute capability.
Ollama's table runs from GeForce GTX 750-era cards (5.0) through the GTX 10-series (6.1), RTX 20-series (7.5), RTX 30-series (8.6), RTX 40-series (8.9) and RTX 50-series (12.0), plus professional and data-center cards.
AMD
AMD cards run through AMD's ROCm library:
- Linux: Ollama requires the AMD ROCm v7 driver. The page lists Radeon RX 9000, 7000 and several 6000-series cards, Radeon PRO and AI PRO cards, some Ryzen AI processors, and Instinct accelerators.
- Windows: Ollama requires a ROCm v7 / HIP7-capable driver stack. The Windows list is shorter: Radeon RX 7900 XTX, 7900 XT, 7900 GRE, 7800 XT, 7700 XT, 7600 XT and 7600, and Radeon PRO W7000-series cards.
On Linux, the page describes an override (HSA_OVERRIDE_GFX_VERSION) for trying a close supported target on an unsupported card. It's an experiment, not a guarantee.
Apple Silicon
Ollama supports GPU acceleration on Apple devices through Apple's Metal API. Its macOS page adds that Apple M-series Macs get CPU and GPU support, while Intel (x86) Macs run on the CPU only. See installing Ollama on a Mac.
Vulkan: extra GPU support on Windows and Linux
Ollama also supports more GPUs through Vulkan, which the page says is enabled by default when that backend is installed. On Windows, most GPU drivers already include Vulkan; most Linux distributions need extra packages. If a Vulkan GPU causes problems, the docs say you can turn Vulkan off with OLLAMA_VULKAN=0.
Choose which GPU Ollama uses
If you have more than one GPU, environment variables limit Ollama to the ones you choose:
| GPU type | Variable |
|---|---|
| NVIDIA | CUDA_VISIBLE_DEVICES (UUIDs from nvidia-smi -L are more reliable than numbers) |
| AMD (ROCm) | ROCR_VISIBLE_DEVICES (list devices with rocminfo) |
| Vulkan | GGML_VK_VISIBLE_DEVICES |
To force CPU-only, the docs suggest giving an invalid ID such as -1. How to set these variables depends on your system; the FAQ covers macOS, Linux and Windows, and Ollama on Windows walks through the Windows steps.
A Linux gotcha
The page notes that on Linux, after suspend and resume, Ollama sometimes fails to find an NVIDIA GPU and falls back to the CPU. Its workaround is reloading the NVIDIA UVM driver: sudo rmmod nvidia_uvm && sudo modprobe nvidia_uvm.
Before you buy hardware
Check the card against Ollama's current list and its memory size against the models you want. Our laptop checklist for local AI covers the other specs that matter.