[AI server setup] CUDA environment

be based on Ubuntu 18.04 System

One 、 install Nvidia drive

The basis of using graphics card , Use it directly ubuntu Warehouse installation

1. First, check the available drivers of the graphics card

ubuntu-drivers devices

2. Install recommended drive

sudo apt install nvidia-driver-470-server

3. Check to see if the installation was successful

nvidia-smi

Reference resources ：Ubuntu Lower installation NVIDIA Three methods of driving _Willard Leozi The blog of -CSDN Blog _ubuntunvidia Drive installation official GeForce The driver | NVIDIA

Two 、 CUDA Environmental installation

Most of the AI The framework passes cuda api To use NVIDIA graphics card for high-performance computing ,cuda Developer website .

1. Go to CUDA Toolkit Post a journal Check the driver support cuda edition

 2. Go to Historical release version Find what you need CUDA Version installed

Complete the installation according to the tutorial given on the official website

3. CUDA Environment variable configuration  

If you only configure editing for the current user ~/.bashrc Add at the end

If it is configured for all users ： stay /etc/profile.d/ New under the directory cuda11.4_env.sh And add

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

4. Test for successful installation

cuda There are some in the installation directory samples Program , You can try to compile and run .

cd /usr/local/cuda/samples/1_Utilities/deviceQuery
sudo make clean
sudo make -j8
./deviceQuery

 3. About unloading methods

CUDA Installation guide There is an unloading method in

  3、 ... and 、 install cuDNN

cuDNN Is based on GPU Accelerated AI Operator Library  .

1. First go cuDNN Website Create an account , Then go to Historical release Download corresponding in cuda Version of cuDNN library .

2. Unzip and download the installation package

Get one "cuda/" Catalog

tar xvf cudnn-11.4-linux-x64-v8.2.4.15.tgz

3. install

Namely the cuda The contents of the directory are installed into the CUDA Toolkit In the installation directory of .

#  Copy cudnn The header file 
sudo cp cuda/include/* /usr/local/cuda-11.4/include/
#  Copy cudnn The library of 
sudo cp cuda/lib64/* /usr/local/cuda-11.4/lib64/
#  Add... To all users cudnn.h Access rights of 
sudo chmod a+r /usr/local/cuda-11.4/include/cudnn.h

4. Check whether the installation is successful

cat /usr/local/cuda/include/cudnn.h | grep CUDNN_MAJOR -A 2

If the version definition information appears, the installation succeeds .

Four 、 other

1. Graphics information

#  View the graphics card device 
lspci |grep -i nvidia 
#  Check the blind model 
nvidia-smi
#  View driver information 
cat /proc/driver/nvidia/version

2. CUDA Information

# 1. see cuda edition 
nvcc -V

# 2.  see cuda edition 
cat /usr/local/cuda/version.txt

3. cuDNN Information

see cudnn.h The header file

cat /usr/local/cuda/include/cudnn.h | grep CUDNN_MAJOR -A 2