Blackwell needs driver 560+ and CUDA 12.6+. On the RTX 5060 Ti 16GB at our hosting the box ships with a working driver, but if you ever need to rebuild, here’s the clean path.
Contents
Check Current State
nvidia-smi
lspci | grep -i nvidia
uname -r
If nvidia-smi fails or shows older driver, proceed. Stop all GPU processes first.
Install Driver (Ubuntu 22.04 / 24.04)
# Remove stale drivers
sudo apt-get remove --purge '^nvidia-.*' '^libnvidia-.*' '^cuda-.*' -y
sudo apt-get autoremove -y
# Add CUDA apt repo (keyring)
wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2404/x86_64/cuda-keyring_1.1-1_all.deb
sudo dpkg -i cuda-keyring_1.1-1_all.deb
sudo apt-get update
# Install recommended driver for Blackwell
sudo apt-get install -y nvidia-driver-560
sudo reboot
For 22.04 use ubuntu2204 in the repo URL.
Install CUDA Toolkit 12.6
sudo apt-get install -y cuda-toolkit-12-6
# Add to path
echo 'export PATH=/usr/local/cuda-12.6/bin:$PATH' >> ~/.bashrc
echo 'export LD_LIBRARY_PATH=/usr/local/cuda-12.6/lib64:$LD_LIBRARY_PATH' >> ~/.bashrc
source ~/.bashrc
Verify
nvidia-smi # Should show RTX 5060 Ti 16GB, driver 560+
nvcc --version # Should show CUDA 12.6+
python -c "import torch; print(torch.cuda.is_available(), torch.cuda.get_device_name(0))"
Expected output: True RTX 5060 Ti 16GB (or similar).
Troubleshooting
- Secure Boot blocks unsigned kernel module: disable Secure Boot in BIOS, or enroll the MOK key following the prompt
- DKMS build failure on newer kernel:
sudo apt install linux-headers-$(uname -r) - “GPU has fallen off the bus”: usually a hardware/PCIe issue – contact us if on our hosting
- Old PyTorch: install the CUDA-12.6 wheel with
pip install torch --index-url https://download.pytorch.org/whl/cu126
See also: Docker CUDA, first day checklist, sanity test, vLLM setup.