Install Ollama on Ubuntu with NVIDIA GPU Support (2026)

Reviewed August 15, 2026. Ollama’s current Linux documentation provides both an install script and a manual package path, plus a recommended systemd service pattern. Its GPU guide currently supports NVIDIA compute capability 5.0 or newer, requires driver version 550 or newer, and requires driver 570 or newer for compute capability 5.0 through 6.2. Those requirements replace the original article’s stale driver tables and unsupported speed claims.

Verify the exact GPU against Ollama’s current hardware page and the installed driver with nvidia-smi. Keep the API on the local or private network until authentication, TLS, authorization, and rate limits are provided by a controlled gateway.

Preflight Ubuntu and NVIDIA

Do not add an unofficial graphics-driver PPA as a routine first step. Use the distribution and NVIDIA-supported driver path for the actual hardware, and change drivers only with a tested recovery plan.

• Confirm architecture, supported Ubuntu release, free disk, RAM, GPU model, VRAM, power, and cooling.
• Verify nvidia-smi works before installing Ollama and record the driver and GPU.
• Check the current Ollama GPU compatibility list instead of assuming every NVIDIA card is supported.
• Resolve Secure Boot, kernel-module, container-runtime, or device-permission issues through official Ubuntu and NVIDIA procedures.
• Select an approved model and record its license, size, source, tag or digest, and expected memory footprint.

Install and run as a dedicated service

The current upstream quick path is curl -fsSL https://ollama.com/install.sh | sh. For controlled environments, inspect the script or use the documented manual package and service instructions, record the version and hash, and deploy through endpoint or configuration management.

Ollama documents a dedicated ollama user and systemd service. Avoid root execution and broad filesystem permissions. Keep model storage on a volume with appropriate capacity, ownership, backup policy, and encryption decisions.

Validate GPU use and failure paths

Ollama documents Linux logs through journalctl -u ollama. Its troubleshooting guide notes that GPU discovery can fail and recommends checking drivers, UVM, container visibility, and system logs. Diagnose the evidence instead of applying every workaround blindly.

1. Confirm ollama -v and the systemd service status.
2. Pull an approved test model and run a non-sensitive prompt.
3. Observe nvidia-smi and Ollama logs while the model is loaded.
4. Measure latency and memory with representative prompts; do not infer a universal CPU-to-GPU multiplier.
5. Test restart, insufficient VRAM, model load failure, suspend/resume, and rollback to the previous version.

Harden network and operations

Treat Open WebUI or another front end as a separate application with its own identity, storage, update, and security boundary. Do not expose Ollama directly to the internet to make a web interface convenient.

Implementation and review gate

Re-run the commands on a disposable Ubuntu host with the exact GPU before publication. Do not publish destructive driver-removal commands, private host details, real prompts, or performance claims that lack retained test evidence.

ITECS can help Dallas organizations plan and validate this work through AI consulting and strategy services. Product, legal, security, and compliance decisions remain subject to the organization’s current requirements and the named review gate below.

Primary sources

• Ollama Linux installation
• Ollama hardware support
• Ollama troubleshooting
• Ollama FAQ