Introduction

This post walks through installing the NVIDIA driver so that nvidia-smi works on Ubuntu, setting up the CUDA Toolkit, and validating the stack with a tiny kernel and deviceQuery. It targets Ubuntu 22.04/24.04 (20.04 is similar).

Prepare the System

  • Verify GPU visibility:
1

lspci | grep -i nvidia
  • Update packages and tools:
1
2

sudo apt update && sudo apt -y upgrade
sudo apt -y install build-essential dkms linux-headers-$(uname -r) wget git
  • Check Secure Boot status:
1

mokutil --sb-state

Install NVIDIA Driver (includes nvidia-smi)

1
2
3

ubuntu-drivers devices
sudo ubuntu-drivers autoinstall
sudo reboot

After reboot:

1
2

nvidia-smi
lsmod | grep nvidia

If nvidia-smi is missing, ensure the matching nvidia-utils-<version> package is installed.

Install CUDA Toolkit

Option A: Ubuntu repo (simple)

1
2

sudo apt install -y nvidia-cuda-toolkit
nvcc --version

Option B: NVIDIA repo (newer)

Install the CUDA keyring, enable the repo for your codename (e.g., jammy, noble), then:

1
2
3
4
5
6

sudo apt update
sudo apt install -y cuda-toolkit-12-4
sudo ln -sfn /usr/local/cuda-12.4 /usr/local/cuda
echo 'export PATH=/usr/local/cuda/bin:$PATH' | sudo tee /etc/profile.d/cuda.sh
echo 'export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH' | sudo tee -a /etc/profile.d/cuda.sh
source /etc/profile.d/cuda.sh

Test CUDA

Minimal sample:

1
2
3
4
5
6
7

#include <cstdio>
#include <cuda_runtime.h>
__global__ void noop() {}
int main(){ int c=0; cudaGetDeviceCount(&c); if(!c){puts("No CUDA"); return 1;}
cudaDeviceProp p; cudaGetDeviceProperties(&p,0);
printf("GPU: %s, SMs: %d\n", p.name, p.multiProcessorCount);
noop<<<1,1>>>(); cudaDeviceSynchronize(); puts("OK");}

Build & run:

1
2

nvcc -O2 hello.cu -o hello
./hello

Or compile NVIDIA’s deviceQuery from cuda-samples and ensure Result = PASS.

Troubleshooting

  • Secure Boot: Disable or enroll MOK; otherwise the kernel module may not load.
  • nouveau conflict: Blacklist and rebuild initramfs if necessary.
  • GCC mismatch: Use a supported GCC and pass -ccbin g++-<ver> to nvcc.
  • Library paths: Ensure LD_LIBRARY_PATH includes /usr/local/cuda/lib64 when using NVIDIA repo.

Conclusion

Get nvidia-smi working by installing the recommended driver, then add CUDA Toolkit via Ubuntu or NVIDIA repo. Validate with nvcc --version, a minimal kernel, and deviceQuery. Watch for Secure Boot, nouveau, and GCC compatibility to avoid common pitfalls.

Summary

  • Install the recommended NVIDIA driver; nvidia-smi should work after reboot.
  • Choose CUDA via Ubuntu repo (simple) or NVIDIA repo (newer).
  • Verify with nvcc, a tiny kernel, and deviceQuery.
  • Mind Secure Boot, nouveau, and GCC compatibility.