One Year of BayernKI: Significant Demand for AI Computing Power

Are these resources sufficient, given the current AI hype?
Wellein: High utilisation rates, rapidly growing user numbers and new application scenarios suggest that capacity limits are on the horizon.

BayernKI was designed for applications such as machine learning, pattern recognition, computer vision, natural language processing and generative AI: who is making use of it?
Hammer: Access to various AI systems is attracting users from all Bavarian universities to the two data centres. It is noteworthy and encouraging that, as we have observed, BayernKI is appealing to researchers from fields that have previously been under-represented – such as universities of applied sciences. Consequently, BayernKI is helping us attract new users and open up new application scenarios for HPC and supercomputers. These systems now also run AI methods, alongside traditional calculations and simulations, which are currently laying the foundations for the development of proprietary AI models.
Wellein: Originally, BayernKI’s focus was indeed on the use cases described above: the development and training of models that benefit particularly from the use of GPUs. However, AI is evidently being used very widely in science. This is where inference now comes into play: the use of trained models to analyse or generate text or data, or as an approximation for complex models. This development affects almost all fields of research, from the natural sciences to the humanities, ranging from physics and climatology to medicine. 

Even before BayernKI, there was extremely successful research into AI in Bavaria – will the initiative succeed in bringing further disciplines, beyond computer science, data science and the natural sciences, into the BayernKI fold and thus into the realm of this new technology? And if so, how?
Hammer: We are indeed seeing the first user groups emerging from the humanities and social sciences, but also from engineering, economics and management. I find proposals particularly exciting from teams that use AI models to evaluate materials or components and estimate their service life, or to analyse buildings and improve their energy efficiency or ventilation – these everyday applications demonstrate the social and economic relevance of research and its technical potential.
Wellein: We’re working to raise awareness of BayernKI even more widely within the research community; to this end, we’re networking with potential user groups and presenting BayernKI at information events organised by universities and higher education institutions, so that we can exchange ideas and gain an understanding of researchers’ technical needs. The training programmes have proven to be extremely helpful in raising awareness of BayernKI, particularly those aimed at students and researchers with little prior experience.

You offer courses for beginners and for advanced learners. Which ones are more in demand? 
Hammer: Around two-thirds of participants book courses for beginners, whilst roughly one-third attend the workshops for advanced learners. This shows that BayernKI is fulfilling its educational and training mandate. The courses introduce students and researchers to AI technology and its potential applications. Practical experience then usually sparks further ambitions. We therefore see many participants returning later to the advanced courses.

As well as training, the two data centres offer consulting to researchers – where do they most often need support?
Hammer: As with HPC and supercomputing, the challenges include identifying errors or problems with the software, optimisation, and boosting performance for training. That is why, right from the start, we planned BayernKI to have a skilled support team spread across both data centres. The advisers accompany research groups throughout the entire project duration, answering questions, helping with the implementation of AI applications on parallelised systems, or supporting the scaling up to the next level of systems.
Wellein: In some cases, the consultancy begins much earlier, for example when researchers want to know which projects they can actually carry out using BayernKI’s resources, or which systems they need for their applications. Many questions also concern background storage and the necessary infrastructure; both are all too often neglected when using AI.

Have you encountered any surprises that, for example, led to changes in your plans?
Wellein: The availability of hardware and price fluctuations in the markets have affected our plans – we have therefore accelerated application procedures and procurement processes in order to build up additional computing power. We will certainly have to continue to deal with this situation; there remains a huge global demand for AI hardware.
Hammer: We were pleasantly surprised by the high level of interest shown by universities of applied sciences. Not least, we saw a high quality of AI applications and a wide variety of application areas among the projects submitted – Bavaria and Germany certainly have nothing to be ashamed of when it comes to ideas and the development of innovative AI methods.