Mike Gore - 2020-08-05

Windows 10 CUDA cuDNN and Tensorflow GPU install

Overview

• These are instruction for installing Python 3.78, CUDA, cuDNN and tensorflow on Windows 10 with and without AVX support
• The first part is for Administrative setup and installs the NVIDIA support packages CUDA 10.1 and cuDNN 7.6.5
  • These version are required for TensorFlow version 2.0 thru 2.2
• The second part of for end users and outline how they can create a virtual environment and install tensorflow

Administrative instructions

Installing of Python 3.78, CUDA 10.1 and cuDNN 7.6.5 setup

• Assumptions - you are doing this on a clean system that has not had these packages installed before

Hardware Requirements

• nVidia GPU card

Software Requirements

• You need an NVIDA account - see: https://www.nvidia.com/en-us/account
• You need Visual Studio Run time for 2015 through 2019 * https://visualstudio.microsoft.com/downloads/ * Direct link: https://aka.ms/vs/16/release/vc_redist.x64.exe

CUDA 10.1 install

• Download 10.1 from the NVIDIA developer web site
• https://developer.nvidia.com/cuda-10.1-download-archive-base
  • Windows -> x86_64 -> 10 -> exe local > Download
  • Direct link: https://developer.nvidia.com/compute/cuda/10.1/Prod/local_installers/cuda_10.1.105_418.96_win10.exe

cuDNN 7.6.5 install

• You need an NVIDIA account for this step
  • https://developer.nvidia.com/cudnn
  • https://developer.nvidia.com/rdp/cudnn-archive
  • File name is: cudnn-10.1-windows10-x64-v7.6.5.32.zip
  • Extract cuDNN zip file into this folder: C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.1

Windows System Environment Variables

• Open Windows search: view advanced system setting - or open Control Panel -> System
• Select Advanced Tab -> Environment Variables -> System Variables -> Path -> Edit
• Use the New button to add each line - you can also click on and edit lines if you need to change settings
• Make sure you have the following lines added

        C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.1\bin
        C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.1\extras\CUPTI\lib64
        C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.1\include
        C:\tools\cuda\bin
        

Install python 3.78

• https://www.python.org/ftp/python/3.7.8/python-3.7.8-amd64.exe
  • Use full installer - not web based one
  • As you click through the install enable for all users and check all options.
  • Unlike 3.8 the 3.78 installer does not let you click on all users until after the first few windows

End user instructions for installing of TensorFlow GPU in a virtual environment

• You MUST always use a Python3 virtual environment otherwise you risk damaging the system wide Python installation
• Note: this installation can be done as a normal user without Administrator perms
• Open a Command Prompt -> Windows Search -> cmd
• Create a virtual environment called venv in your current directory
  • python3 -m venv --system-site-packages .\venv
  • .\venv\Scripts\activate
  • pip install --upgrade pip
  • Please remove \Users\userid\AppData\Local\Temp\pip - replace userid with your own userid pip install --upgrade pip

Check if your system has AVX instruction support

• Download and run CoreInfo
• https://docs.microsoft.com/en-us/sysinternals/downloads/coreinfo
  • Unzip this into a folder
  • Open a Command prompt -> Windows Search -> cmd
    • Coreinfo64.exe
    • Look for AVX - it should have a star in front of AVX rather then a dash

On systems WITHOUT AVX support

• install using this tensorflow 2.0 wheel as follows pip install https://github.com/fo40225/tensorflow-windows-wheel/raw/master/2.0.0/py37/GPU/cuda101cudnn76sse2/tensorflow_gpu-2.0.0-cp37-cp37m-win_amd64.whl

On newer systems with AVX support

• pip install tensorflow-gpu==2.2

Testing

• python -c "import tensorflow as tf;print(tf.reduce_sum(tf.random.normal([1000, 1000])))"