RTX 3060 CUDA Setup on Ubuntu 22.04 (CUDA 11.6)

Setting up a reliable deep learning environment requires tight alignment between GPU hardware, kernel modules, and NVIDIA’s software stack. This guide documents a known-stable configuration for RTX 3060 (LHR) on Ubuntu 22.04 LTS using CUDA 11.6 and cuDNN 8.8.

🧩 System Baseline & Prerequisites

#

Confirm the GPU model and ensure the system sees the device correctly.

lspci -vnn | grep VGA
# Expected: NVIDIA Corporation GA106 [GeForce RTX 3060 Lite Hash Rate]

This guide assumes:

• Ubuntu 22.04 LTS (standard kernel)
• Clean system or previously purged NVIDIA drivers
• Secure Boot disabled (recommended for simplicity)

🚫 Disable the Nouveau Driver

#

The open-source nouveau driver conflicts with NVIDIA’s proprietary driver and does not support CUDA.

Create a blacklist file:

sudo vim /etc/modprobe.d/blacklist-nouveau.conf

Add the following:

blacklist nouveau
options nouveau modeset=0

Update initramfs and reboot:

sudo update-initramfs -u
sudo reboot

Verify Nouveau is disabled after reboot:

lsmod | grep nouveau
# No output means success

🧠 Install NVIDIA Driver (510 Series)

#

CUDA 11.6 pairs best with the NVIDIA 510 driver series, offering long-term stability.

Check the recommended driver:

ubuntu-drivers devices

Install the driver:

sudo apt install nvidia-driver-510 -y
sudo reboot

Verify installation:

nvidia-smi

You should see:

• GPU: RTX 3060
• Driver Version: 510.xx
• CUDA Version: 11.6

⚙️ Install CUDA Toolkit 11.6

#

The CUDA Toolkit provides nvcc, runtime libraries, and developer tools.

Download the local runfile from NVIDIA’s CUDA archive to avoid apt conflicts.

Run the installer:

sudo ./cuda_11.6.0_510.39.01_linux.run

During installation:

• Uncheck NVIDIA Driver (already installed)
• Install Toolkit only

Environment Variables

#

Append to ~/.bashrc:

export PATH=/usr/local/cuda-11.6/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/cuda-11.6/lib64:$LD_LIBRARY_PATH

Apply changes:

source ~/.bashrc
nvcc --version

🧠 Install cuDNN 8.8

#

cuDNN accelerates deep learning primitives for frameworks like PyTorch and TensorFlow.

Install via NVIDIA’s local repository package for Ubuntu 22.04:

sudo dpkg -i cudnn-local-repo-ubuntu2204-8.8.1.3_1.0-1_amd64.deb
sudo cp /var/cudnn-local-repo-*/cudnn-local-*-keyring.gpg /usr/share/keyrings/
sudo apt-get update
sudo apt-get install libcudnn8 libcudnn8-dev libcudnn8-samples

Verify headers and libraries:

ls /usr/include/cudnn*.h
ls /usr/lib/x86_64-linux-gnu/libcudnn*

🚀 Validation & Performance Test

#

Use a minimal Numba CUDA kernel to confirm GPU execution.

import numpy as np
from numba import cuda

@cuda.jit
def increment_kernel(arr):
    idx = cuda.grid(1)
    if idx < arr.size:
        arr[idx] += 1

data = np.zeros(10_000_000, dtype=np.int32)
d_data = cuda.to_device(data)

increment_kernel[256, 256](d_data)
cuda.synchronize()

While running, check GPU activity:

nvidia-smi

You should see Python consuming GPU memory and compute.

🛠️ Troubleshooting FAQ

#

🧩 Final Notes

#

This RTX 3060 + CUDA 11.6 + cuDNN 8.8 stack is a conservative, production-proven configuration. While newer CUDA releases exist, this pairing prioritizes driver stability, framework compatibility, and reproducibility, making it well-suited for long-running training workloads on Ubuntu 22.04.