CUDA cuDNN

CUDA cuDNN

Installation of CUDA, cuDNN

(updated 2024.4)

For DLIP 2024 course,

CUDA=11.8, PyTorch 2.1

This covers the installation of CUDA, cuDNN on Windows 10/11.

Here, we assume that you have a CUDA-compatible GPU already installed on your PC.

See here for more detailed instruction by NVIDIA

The order of installation is

1. NVIDIA Graphic Card Driver

2. CUDA toolkit and cuDNN

---

1. Installation of NVIDIA Graphic Card Driver

Prerequisite: Visual Studio Community

Visual Studio is a Prerequisite for CUDA Toolkit

Visual Studio Community is required for the installation of Nvidia CUDA Toolkit. If you attempt to download and install CUDA Toolkit for Windows without having first installed Visual Studio, you will get a message for the installation.

Follow: How to install Visual Studio Community

Step 1: Find Your GPU Model

You can check your GPU Card in Window Device Manager

항목	Graphics Card Info	Window Version
Short Key	Win Key → Device Manager (장치관리자)	Win key→ System (또는 내PC 우클릭 → 속성)

Step 2: Check If Graphic Driver is Installed

Go to anaconda prompt Admin Mode.

If Anaconda is not installed, see here Anaconda Installation

You can also use r Window Command Prompt

Start > All Programs > Accessories > Command Prompt

nvidia-smi

image

If you have the summary of CUDA version, you already have installed the driver.

Here, CUDA version 11.6 means, the driver can support up to CUDA ~11.6.

If you don't see any information, go to Step 3. Otherwise Skip Step 3

Step 3: Download NVIDIA Graphic Driver

Download Site: https://www.nvidia.co.kr/Download/index.aspx?lang=kr

Select your Graphic Card Model and Window Version. Then, download.

image

It does not matter whether you select (a) GRD(game-ready driver) (b) NVIDIA Studio Driver.

image

Now, Install. There is NO need to select GeForce Experience. Choose the default Settings.

image

Check if you have installed, check the driver version.

Go to anaconda prompt Admin Mode. Then, go to the environment you want to work with.

conda activate py39
nvidia-smi

2. Install CUDA & CuDNN using Conda

Step 1. Find supported CUDA version

Depending on your graphic card, there is a minimum version for CUDA SDK support.

First, find the range of CUDA SDK versions for your Graphic Card (Compatible version)

Example:

• GTX 1080 is PASCAL(v6.1) and it is supported by CUDA SDK from 8.0 to 12.4

• RTX 4xxx is AdaLovelace(v8.9) and it is supported by CUDA SDK from 11.8 to 12.4

For most GPUs, CUDA SDK >11.x will work fine

BUT, if you use RTX4xxx, you may need to install CUDA SDK > 11.8

See here for detail

image

Step 2. Install CUDA and cuDNN via CONDA

It is recommended to install specific CUDA version in the selected Python environment.

CUDA=11.8, for DLIP 2024-1

CUDA=11.2, for DLIP 2023-1

CUDA=10.2.89, for DLIP 2023-1

Run Anaconda Prompt(admistration).

Activate conda virtual environment. Then, Install specific CUDA version

[$ENV_NAME] is your environment name. e.g. conda activate py39

DLIP 2024-1

#conda activate [$ENV_NAME]
conda activate py39
    
# CUDA 11.8 & CuDNN
conda install -c anaconda cudatoolkit=11.8 cudnn 

DLIP 2022-1

#conda activate [$ENV_NAME]
conda activate py39
    
# CUDA 10.2 & CuDNN
conda install -c anaconda cudatoolkit==10.2.89 cudnn 

Important Note

Depending on your CUDA version, the minimum version of PyTorch is determined.

For example

• CUDA 11.6 supports PyTorch 1.13 or higher

• CUDA 11.8 supports PyTorch 2.0 or higher

See here for CUDA and PyTorch Version matching