A first-time participant reviews ISC26

My work here at EPCC, the UK National Supercomputing Centre, largely takes place in the areas of data science and machine learning (ML). As such, when I first started my job, “supercomputing” mostly felt like something my colleagues worked on rather than part of my own role. 

In the past few years however, that separation has been shrinking: supercomputers enable the training of large-scale AI models, and researchers are increasingly integrating machine learning into their workflows, combining it with a diverse range of HPC methods. Hence, I was excited to get the opportunity to attend ISC High Performance for the first time, presenting my ongoing work on the CONTINENTS project.  

Community exchange 

Meeting like-minded people and exchanging ideas was at the heart of ISC this year, with a newly added Community Stage and various informal formats. Things kicked off on Tuesday with Women in HPC's session 'Finding your people in HPC'. This was a great starting point for a newcomer like me, from providing helpful ISC tips to getting to know other attendees. Later in the evening, the 'Tech Talks by Women in HPC' lightning talks showcased a diverse range of voices, prompting interesting conversations over the next few days during poster sessions and lunch breaks. 

Poster: A Weather and Climate Case Study in Sustainable AI 

On the Wednesday afternoon, I presented a poster about my ongoing work on the CONTINENTS project in the 'WHPC Poster Session'. CONTINENTS is a four-year long collaboration between EPCC, the National Centre for Atmospheric Science (NCAS) and the National Centre for Atmospheric Research (NCAR). It involves people from a range of backgrounds, all working on different avenues towards a step change in sustainable climate and weather modelling.  

Currently, we’re looking at Sustainable AI and how this set of techniques might be used in weather and climate science. After looking at how machine learning is used in the weather and climate community, two domain-specific ML toolkits stood out, namely Anemoi (developed at ECMWF), and CREDIT (developed by MILES at NCAR). These tools allow domain scientists to integrate ML into their research with relatively low barriers.  

It is often argued that in weather and climate research in particular, ML alternatives can be significantly more energy efficient than their traditional dynamical counterparts. To investigate this in detail, we designed a case study that allows us to quantify and evaluate the energy use along the entire ML lifecycle when using these domain-specific frameworks. We start by fine-tuning the AIFS model in Anemoi for extreme weather events to establish baseline energy use. That baseline then allows us to explore sustainable AI techniques along four aspects (data, model, training and inference).  

This work is ongoing, and it was fantastic to be able to talk to the wider HPC community about it and hear their thoughts and ideas. 

A summary of the poster 'A case study in sustainable AI: Applying carbon-reducing techniques in weather forecasting frameworks' is available on the ISC26 website.

Other programme highlights 

Weather and climate simulation

Peter Dueben (ECMWF) chaired a session of invited talks celebrating recent advances in running weather and climate simulation at 1km resolution. All speakers highlighted the integration of machine learning into a hybrid approach that combines dynamical modelling with physics-constrained AI components. The resulting models also create rich datasets to support the wider research community.  

HPC sustainability and efficiency

A range of sessions were dedicated to sustainability and efficiency in HPC, such as Nikela Papadopoulou’s talk discussing two adaptive software frameworks, Parallea and Odin, and the '2nd International Workshop on Energy Efficiency with Sustainable Performance', which covered a variety of topics from posit arithmetic to efficiency evaluation. 

Outreach and education

Sessions such as 'Super(computing)heroes' and 'Quality Assurance in HPC/AI Training' discussed challenges and ideas around communication and education relevant to our Outreach activities and MSc programme. 

AI Factory Antennas

EPCC’s Director Mark Parsons spoke in a BoF session titled 'A European HPC Ecosystem for AI: Challenges and Opportunities', where he presented Pilot-UKAIFA, the UK AI Factory Antenna. It was interesting to also hear from other AI Factories and AI Factory Antennas about their plans for a diverse range of systems and user requirements. This session highlighted the need for community engagement and talent development. 

Final thoughts

Attending a conference is always fun and stimulating, but it can also be a bit draining. With ISC being a week-long event with long days, I was expecting to need a good rest after coming back. So I was pleasantly surprised when I returned home energised and inspired! I suspect this was largely because of the company (and the gorgeous park next to the venue). Everyone was welcoming and keen to exchange ideas, and I hope to attend again in the future.