Administrative information

The purpose of this lecture is to show students what hardware and software architectures help train and deploy deep learning solutions. We must acknowledge that these hardware and software components are brilliant technical solutions that enable us to scale training and inference. Apart from the hardware, NVIDIA GPUs and Google TPUs are the best choices today because they implement optimized deep learning algorithms with high-quality and fast drivers.

For deep learning, we use much more software than deep learning frameworks alone. We use configuration, scheduling, orchestration, and many other tools. The short introduction in this lecture only scratches the surface.

At the last part of the lecture, we show how multi-GPU training can be realized with Horovod. The aim is not to have a deep dive into multi-GPU trainings, but to show, that it is not so hard to implement a basic solution.

Opening the web sites of the hardware manufacturers and of the software providers might help the students to have some hands-on experience.

Outline

• hardware solutions - from desktop to server grade
• deep learning software frameworks
• additional softwares for deep learning solutions

Time schedule

Duration (Min)	Description
5	The need of parallel computing in deep learning
15	Hardware solutions
5	How to compare different hardware for deep learning
10	Deep learning software architecture
10	Deep learning frameworks
10	Additional software components