Last December, the Center for National High Performance Computing at Friedrich-Alexander-Universität Erlangen-Nürnberg (NHR@FAU) placed an order with MEGWARE Computer Vertrieb und Service GmbH from Chemnitz-Röhrsdorf for the installation of two new supercomputers. With the new high-tech equipment, the NHR site in northern Bavaria will continue to grow and will be available to researchers from all over Germany with even more power. The systems are financed by funds from the National Supercomputing Program as well as FAU, the Bavarian State Ministry of Science and the Arts, the Catholic University of Eichstätt-Ingolstadt and Hof University of Applied Sciences.
We spoke to Prof. Dr. René Peinl, Professor at the Faculty of Computer Science and Head of the Institute for Information Systems (iisys) at Hof University of Applied Sciences, about the project.
Prof. Peinl, what is the aim of the project?
The aim is to gather the necessary computing resources for research into deep neural networks, e.g. for voice assistants, chatbots and image-based quality assurance, at a central location instead of setting up smaller computing clusters at each individual university. On the one hand, this allows purchasing prices to be reduced because the negotiating position is much better due to the size of the order. On the other hand, the operating team in Erlangen is better equipped with representatives for vacation and illness. Capacity utilization also increases, making it more efficient overall. I think the strategy with a few large data centers is a very good one. Strong centralization usually paralyzes and means too great a concentration of power and dependencies. On the other hand, small-scale, everyone doing their own thing is often inefficient and usually less professionally organized. I would like to see the same approach adopted for other central university IT systems such as e-mail, student administration and collaboration software.
How did the cooperation come about?
Prof. Dr. Gerhard Wellein, Head of NHR@FAU, has been a member of our University Council since October last year and is in close contact with the university management in this role. We came into contact with him when we were looking to expand our computing technology. Personally, I was a little skeptical at first. There were challenges to clarify in terms of data protection and data storage. If we had to spend hours copying data from Hof to Erlangen every time before we could get started, it would be difficult. After all, we are talking about large amounts of data in the terabyte range. However, we were able to solve these issues and now I am very happy that the cooperation has come about.
What does the cooperation with the FAU look like in concrete terms?
There is an online portal where we can activate new users ourselves. A user then logs in there remotely, copies their data and the training program to the computing cluster in Erlangen, configures a work order to the computing cluster and sends it off. The scheduler, which is an automated management system for computing capacity, sorts the work order when the requested capacity is available. Jobs can run for a maximum of 24 hours. Then they must save their (interim) result and, if necessary, trigger a follow-up job that continues to work on the interim result. Training runs can quickly take 2-3 weeks.
What are the supercomputers used for?
In Erlangen itself, we work a lot with simulations in the natural sciences. Here in Hof, it’s mainly about deep learning, the technology behind the current AI hype. We build on existing open source AI models and train them further with our own data so that they are better suited to certain applications than before. To do this, we cooperate with companies, especially SMEs from the region. At the moment, this is mainly happening in connection with our ERDF project “Multimodal human-machine interface with AI”, or M4-SKI for short. Multimodal means that the model not only understands language or can analyze images, but that a model can use inputs in the form of images, text, audio and other modalities and/or also produce results in the form of several modalities. For example, an image could help the user to understand a textual response more easily.
How does Hof University of Applied Sciences or iisys benefit from this project?
As already mentioned, we get lower purchase prices, do not have to use our own resources to support the computing cluster and, last but not least, do not have to pay for the electricity. Training such models requires considerable amounts of energy. The entire computing cluster in Erlangen requires up to 672 kW at peak times. Of course, we Hofer only use a small proportion of this. Even if the impact of AI on climate protection is often exaggerated, because, for example, a single overseas flight produces significantly more CO2 equivalents than training many models, this should not be ignored. I hope the FAU mainly uses green electricity.