Installing GPU drivers and software packages

You need to install the following software before you can use a GPU profile with your virtual server.

• NVIDIA drivers - allows your operating system to communicate with the GPU.
• CUDA Toolkit - development environment for high-performance GPU-accelerated applications.
• cuDNN - (CUDA Deep Neural Network) library that is used for neural network applications.
• NCCL - cluster communications library for multi-GPU and or multi-system communications.
• BLAS (Basic Linear Algebra Subprograms) - routines that provide the building blocks for performing basic linear algebraic operations such as one of the following libraries:
  • Atlas®
  • cuBLAS
  • MKL
  • OpenBLAS

If you are installing a machine learning framework, you need to install framework software such as one of these platforms in addition to the preceding software packages.

• Caffe (Convolutional Architecture for Fast Feature embedding) - licensed, open source deep learning framework.
• Tensorflow (Tensorflow 1.2+) - provides an open source software library for numerical computation.