AI Café: How can HPC technologies help AI
On March 17th, FocusCoE participated in a live AI for Media Web Café alongside the three Centres of Excellence: RAISE, CoEC, and HiDALGO. The virtual session brought together the CoEs working in AI sectors to explain how HPC technologies can help AI. In all, over 40 participants from industry and research joined the hour and a half café.
Starting off the presentations, Xavier Salazar introduced FocusCoE and the resources available at the “one stop shop” of our website such as technological offerings. Here, anyone from industry or research who wants to learn more can also read up on use cases, search available codes and software packages, and link directly to the CoEs of interest.
Next, the CoEs presented several case studies on how they are using AI in combination with HPC technologies to solve real-life problems. Although each CoE’s application of AI differed, some common themes emerged in answer to the question, “How can HPC help AI?” Firstly, AI is now benefitting from the increasing availability of large and even “big” data sets but often can’t use them in their entirety due to excessive processing time. This is by far the clearest example of how HPC can help. In a use case described by Andreas Lintermann on behalf of CoE RAISE, a dataset that was estimated to take over 300 hours to process using 4 GPUs was modified to run on HPC systems theoretically as large as 2000 GPUs in as little as 45 minutes! With the ability to more quickly train AI models using more data, it is also possible to increase the accuracy of the resulting models or surrogates. In turn, building more accurate surrogates speeds up the ability to run accurate simulations since one no longer needs to build the simulation models by hand.
Using AI to build data model surrogates also has benefits for data privacy, as discussed by Christoph Schweimer from HiDALGO. When modelling how messages spread across social media, researchers initially had to build social network graphs manually from data harvested from real social media users, whose privacy had to be strictly protected. However, with HPC computing resources, HiDALGO researchers were able to use those real graphs to train AI to build simulated social network graphs instead. These simulated graphs share the same characteristics of real graphs but require far less time to create and don’t rely on any real-user data: thus holding no privacy risks to users.
The experience gained through these use cases has naturally brought several opportunities and challenges to light, which were also discussed over the course of the program. For instance, Temistocle Grenga from CoEC highlighted the existing bottleneck of moving data between different types of processors (CPU and GPU, as examples).
Lastly, CoEs summarized the numerous resources in terms of services and training opportunities they provide to help AI experts learn to exploit the benefits of HPC. As an immediate example, CoEC will participate this week in South-East Europe Combustion Spring School 2022. For ongoing information on training like this, make sure to bookmark our training calendar, which shows events from all the EU HPC CoEs.
For the full recording of this event, check out the video below!