Install cuDNN and NCCL

Install cuDNN

NVIDIA CUDA® Deep Neural Network library (cuDNN) is a GPU-accelerated library of primitives for deep neural networks. cuDNN provides highly tuned implementations for standard routines such as forward and backward convolution, pooling, normalization, and activation layers.

This topic describes how to configure cuDNN on Ubuntu 16.0.4.

1. Log in to the GEPC instance. Access cuDNN, and then click Download cuDNN.

   

2. Register for an account and log in as prompted.

   

3. Select the version for the installed CUDA. In this example, the version for CUDA 9.1 is selected.

   

4. Select the version for your operating system. In this example, cuDNN v7.0.5 Library for Linux is selected and the cudnn-9.1-linux-x64-v7.tgz file is downloaded.

   

5. Go to the directory where the downloaded file is saved and execute
   the following command to decompress the file:

   sudo tar –xvf cudnn-9.1-linux-x64-v7.tgz –C /usr/local         ---- Installation is completed.
   

Install NCCL

NVIDIA Collective Communications Library (NCCL) implements multi-GPU collective communication primitives that are performance optimized for NVIDIA GPUs. NCCL provides routines that are optimized to achieve PCIe, NVLink, and InfiniBand high-speed interconnection.

To install NCCL:

1. Log in to the GEPC instance. Access NCCL and click Download NCCL.

   

2. Register for an account and log in as prompted. Then, select the correct NCCL version. In this example, Download NCCL v2.1.4, for CUDA 9.1, Jan 18, 2018 is selected.

   

3. Select a version for the operating system. In this example, the operating system is Ubuntu 16.0.4, and NCCL 2.1.4 for Ubuntu 16.04 and CUDA 9 is selected. The nccl-repo-ubuntu1604-2.1.4-ga-cuda9.1_1-1_amd64.deb file is downloaded.

   

4. Go to the directory where the downloaded file is saved and execute
   the following command:

   sudo dpkg -i nccl-repo-ubuntu1604-2.1.4-ga-cuda9.1_1-1_amd64.deb  
   

The command decompresses the file and saves the files in the include and lib folders to the corresponding folders in the /usr/local/include/usr/local/lib directory.