Deep-Learning-Workstation-Setup

This work is licensed under a Creative Commons Attribution 4.0 International License.

I am sharing my experiences setting up my Deep Learning Workstation. The main reason I am doing this documentation is to be able to redo all of the installation and configuring if needed. Furthermore, I hope this documentation helps others who are getting started with this topic. Feel free to contribute.

---

Overview

Advantages of using Docker for a Deep Learning Workstation are that you will only need to install the following on your host system.

• Linux OS (here: Ubuntu)
• Docker CE
• NVIDIA Driver

Everything else will be installed in the Docker container. Using containers you can make usage of different versions of CUDA at the same time in different containers. Therefore I believe using them is the best way in developing Deep Learning models.

(source: https://github.com/NVIDIA/nvidia-docker)

---

Installation instructions

The InstallationInstructions.md file provides information on:

• Installing Ubuntu Server
• Enabling remote access via SSH
• Installing the NVIDIA driver and prior blacklisting of the Nouveau driver
• Installing Docker CE
  After having installed Docker, checkout the following sections below
  • Docker for beginners
  • docker compose examples with GPU support
    Here you will find also examples for PyTorch and TensorFlow

Docker for beginners

This DockerForBeginners.md files provides information on some of my commonly used docker commands, e.g.:

• List running docker containers
• Retrieve the token of a running dockerized Jupyter notebook instance

docker compose examples with GPU support

The folder ./examples contains multiple examples for running Docker containers with GPU support.

1. nvidia-smi: examples/nvidia-cuda/README.md
   Start with this example which uses a pre-built image with docker compose indicating how a GPU can get made accessible within a container using docker compose. Afterwards try out the PyTorch or TensorFlow examples

2. TensorFlow, PyTorch
   After having tried out the example mentioned in 1.) try out this one which customized an image based on nvcr.io/nvidia/pytorch and nvcr.io/nvidia/tensorflow images.

   1. PyTorch examples/pytorch/README.md
   2. TensorFlow examples/tensorflow/README.md

ℹ️ I personally prefer using docker-compose.yml files as they offer a clean way to build images and start containers without the need of long and tedious commands on the cli or the need for hard to maintain bash scripts.

---

My hardware setup

In 2018 I got a good deal on a used Lenovo ThinkStation P520 (30BE006X**) equiped as follows:

• Xeon W-2133 6C/3.6GHz/8.25MB/140W/DDR4-2666
• 900W Platinum Power Supply
• 1x 32GB RDIMM DDR4-2666 ECC
• 256GB SSD + 1 TB HDD, both SATA
• 1x NVIDIA Quadro P4000 (Data sheet)
  • 1792 CUDA cores
  • 8 GB GDDR5 GPU Memory
  • CUDA compute capability 6.1
  • 5.3 TFLOPS FP32 performance

Modifications over time:

• Replacements
  • GPU: NVIDIA TITAN RTX replacing the NVIDIA Quadro P4000
  • SSD: 1TB SSD (SATA)replacing the 1TB HDD
  • RAM: 4x 32GB RAM (same type but different vendor) replacing the 1x 32GB RAM
• Added hardware
  • SSD: 512GB SSD (NVMe)