Tuesday, 21 November 2023
 

Time:
5:00 - 6:30 PM
Place: This Tutorial is a hybrid event. You can join the event on-site and online.
Material Engineering Center Saarland (MECS), Saarland University
Campus, Building D3.3, Room 2.15 (1. Floor, seminar room)
Moderator:
Dr. Tim Dahmen, German Research Center for Artificial Intelligence (DFKI)

Deep Learning (DL) has tremendous impact in various scientific fields including Materials Science and Engineering (MSE) and has very successfully been used for various tasks in the field. However, the method is both data-hungry and requires large amounts of computational resources. While most experimental groups are used to purchasing expensive equipment or sharing it in facilities, the novel need for compute power raises questions: rent or buy, individual servers or virtualization, cluster or interactive use? 

A brief introduction to the topic focusses on arguments why deep learning hardware is scientific machinery, not office equipment. We present several possibilities to source compute power for deep learning projects. We start by the option of renting hardware from cloud-providers like google colab or amazon EC2 and discuss the economic and administrative challenges involved. We then present the option to purchase hardware and map it to various users and projects using GPU virtualization. Here we put a focus on the technical realization. We finally show how centralized compute resources like national supercomputers can be accessed, focusing on the process of applying for access via a grant. We end by a discussion how a hybrid strategy can match complex research needs.