NVIDIA

TensorFlow GPU support

Note

This chapter is about using TensorFlow with GPU support on a Windows workstation. Since TensorFlow no longer supports GPU on native Windows, you will need to install TensorFlow on WSL2. Therefore you need to install NVIDIA Drivers on your Windows host, CUDA Toolbox and cuDNN Library local on your WSL2 host with Linux or the tensorflow[and-cuda] package including CUDA and cuDNN in additional pip packages on WSL2.

Install options

graph TB; classDef red fill:#f96,stroke:#AAA,stroke-width:0.5px; classDef green fill:#90EE90,stroke:#AAA,stroke-width:0.5px; subgraph " Host " A[Host: Windows] A --> B[NVIDIA Driver]; end subgraph "Option 1 " B --> C[Guest: WSL2]; C --> D[CUDA Toolkit + cuDNN Libraries]; D --> E["TensorFlow package"]; end subgraph "Option 2 " B --> F[Guest: WSL2]; F --> G["TensorFlow[and-cuda] package"]; end class B,D red; class A,C,F green;

Installation Tasks

Installation Tasks	Target Platform
Install NVIDIA Drivers	Windows
Install CUDA / cuDNN	Linux
Info TensorFlow with CUDA package	Python
Config .bashrc	Linux

• NVIDIA @ Windows
  • NVIDIA Installer
  • Verify NVIDIA Driver
  • Verify CUDA Toolkit
• CUDA @ WSL2 Ubuntu
  • Installing CUDA Toolkit 11.8
    • 1. Download the CUDA repository package:
    • 2. Move the downloaded file to /etc/apt/preferences.d/ directory
    • 3. Download CUDA 11.8 installation package:
    • 4. Install the downloaded CUDA package:
    • 5. Copy the keyring file to the appropriate directory:
    • 6. Update your repository information:
  • Adding CUDA to System Paths
  • Verify CUDA version
  • Removing the CUDA Toolkit
  • NVIDIA.com Links
• cuDNN @ WSL2 Ubuntu
  • Installing cuDNN 8.6.0
    • Download cuDNN 8.6.0
    • 1. Install the zlib1g package:
    • 2. Extract the downloaded tar.xz file:
    • 3. Copy the extracted files to the CUDA include directory:
    • 4. Copy the files to the CUDA lib64 directory:
    • 5. Modify the permission:
  • Updating CUDA Library Links
    • 1. Open Powershell as admin
    • 2. Change directory
    • 3. Delete existing files
    • 3. Create Symlinks
    • 4. Restart WS
    • 5. Adding cuDNN to System Paths
  • Uninstall / Update cuDNN
    • 1. Identify the Files to be Removed
    • 2. Remove CUDAnn Library Files
  • NVIDIA.com Links
• .bashrc
  • Editing .bashrc
    • Step 1: Open .bashrc
    • Step 2: Save and Exit Nano
  • .bashrc Configurations
    • Adding CUDA to System Paths
    • Adding cuDNN to System Paths