PASC conference

The Platform for Advanced Scientific Computing (PASC) conference is a popular annual event for sharing the latest developments and findings in the field of computational sciences. Held each year in Switzerland, this year PASC will return to Zurich under the theme Synthesizing Applications Through Learning and Computing. With a general focus this year on the grand challenge of combining physics-based simulations with novel machine learning and AI based methods to address interdisciplinary problems in science, PASC always attracts a very wide ranging audience and is an excellent way in which traditionally disparate fields can connect.

A major component of PASC are the mini-symposia, which is a two hour session focussed on a specific topic. Combining technical talks with panels, these enable world experts in a specific topic to connect with the wider PASC audience to discuss how these fields can be leveraged by the wider scientific computing community most effectively.

Phil Hasnip from the PAX-HPC ExCALIBUR project will be talking about how machine learning techniques can compliment existing numerical methods for predicting material properties. However, the very large parameter space means that this is a tricky endeavour and-so coupling a traditional computational model with ML, which learns on-the-fly, is likely to deliver better results. This talk will describe the use of Gaussian Process-based ML models driven by the popular CASTEP simulation package, reproducing and predicting atomic forces substantially faster and with controllable uncertainty.

The Modern PDE discretisation methods and solvers in a non-smooth world minisymposium will explore the tension between high-order discretisation methods for PDEs and the fact that many physical phenomena are non-smooth. High order discretisations in space and time can make optimal use of FLOP-bound exascale hardware and have the potential to unlock additional parallelism. However, it is an open question how these methods can be applied to time-dependent PDEs with elliptic constraints. This session will consist of talks by Marc Marot-Lassauzaie (TU Munich), Jemma Shipton (Exeter and part of the ExCALIBUR parallel in time and RSE training projects) and Olindo Zanotti (Trento); a discussion session will explore how the advantages of sophisticated PDE solvers and Machine Learning can be combined productively.

The Riding the Cambrian explosion in hardware for scientific computing minisymposium has been organised by the RISC-V, CGRA, and PETSc for CS-2 H&ES testbeds. This session will focus on exploring the potential of the plethora of new hardware that is coming available (much of it driven primarily by AI workloads) for accelerating HPC applications. It will consist of talks by Teresa Cervero Garcia (Barcelona Supercomputing Centre) who is an expert on RISC-V and will be discussing the role of this technology in HPC, Justs Zarins (EPCC) who has been working heavily with the Cerebras CS-2 as part of the CGRA and PETSc on CS-2 ExCALIBUR projects, and lastly Mario Ruiz (AMD) who has worked heavily with AMD’s AI engines which are vectorised accelerators that are packaged with AMD CPUs and FPGAs.

Kirsty Pringle from the Universe-HPC ExCALIBUR project will be giving a presentation in the Ethical and Societal Considerations for Scientific Computing mini-symposium about embedding ethical and societal considerations into HPC Education and Training, making RSEs more aware from an early stage around these concerns as they are working with, and developing, HPC and AI workloads. This talk will also address the question around what should be done to make HPC more accessible and whether we can make it a viable career option for underrepresented groups, the ultimate aim to be to start a discussion and gather community feedback to start outlining best practices.

George Bisbas from the ExCALIBUR xDSL project will be giving a presentation in the Motif-Based Automated Performance Engineering for HPC mini-symposium about work done in the xDSL project around enabling Devito and PSyclone to leverage the MLIR ecosystem. He will describe how structured-grid based abstractions in Python are lowered to highly efficient parallel code, George will highlight the benefits that MLIR provides when it comes to sharing common compiler infrastructure and introduce the xDSL framework that lowers the barrier to entry in leveraging this compiler technology.