Welcome to our AI working group
High-Performance Computing (HPC) is an ecosystem connecting many scientific communities, from mathematics and engineering to experimental sciences and computer science. It provides the tools and environments needed to address new hybrid and accelerated computing architectures that drive future research. The recent progress in AI is a game-changer for science, offering new ways to analyze data and improve simulations through deep AI/HPC integration. In this context, our research program NumPEx aims to bring AI and HPC together at scale — developing Exascale-ready software, AI-based data analysis methods, and large-model training capabilities. By doing so, NumPEx will accelerate the use of AI for science and strengthen collaborations across national and international research communities.
High-Performance Computing (HPC) is an ecosystem at the intersection of vast communities. It tackles the methodological, algorithmic, software, and portability challenges to efficiently leverage new hybrid, heterogeneous, and highly accelerated computing architectures, which will be the architectures of tomorrow in regional and national centers. HPC aims to provide a computing environment that accelerates the research in many communities from applied and industrial sciences, in fields such as astrophysics, high-energy physics, climate and environment, biology and digital health, neuroimaging, physical chemistry, humanities and social sciences, aeronautics, energy production/transition, to name but a few. In most of these domains, the recent development in AI is seen as a game-changer to better integrate and analyze experimental data as well as to dramatically improve numerical simulations with deep AI/HPC hybridization, pushing new frontiers to accelerate research. However, the use of AI in these domains often remains at the proof-of-concept level and thus needs a clear leap forward with implementation at scale and full-fledged scientific validation.
Recent advances in AI have shown that large available datasets and huge computing power are the core components required to develop powerful and flexible models for natural language processing and images. With the positioning of HPC at the crossroads of diverse scientific domains and the community expertise in managing large computing infrastructures and data, it offers the perfect ground to develop the novel usage of IA for science. Indeed, the HPC ecosystem gathers massive experimental and observational data, large computing resources, and expert practitioners, which are critical in the advent of powerful AI models to tackle complex scientific questions. Moreover, the HPC community’s expertise in efficiently leveraging large computing infrastructure also offers the opportunity to improve AI practices to train large models.
Our actions
In this context, the NumPEx initiative aims to push the joint use case of AI and HPC in several directions.
HPC and AI hybridization
With the end of Moore’s law, emerging research direction aims to develop hybrid approaches based on AI to accelerate parts of scientific computing applications. This hybridization can take
several forms, from the development of simulation codes including physic-informed AI models, to the optimization of runtimes based on log-informed models, or the reduction of observational data and digital simulations at work in reduced models and digital twins. To that end, NumPEx aims to push forward an Exascale-ready software stack embedding AI solutions that answer the needs of the application communities.
HPC Data Analysis
The massive data produced by scientific computing with HPC also raises particular challenges to apply AI models, as they often can’t be stored and collected on the computing system. This calls for specific AI methodologies, able to analyze or assimilate massive data at scale, whether they come from numerical solutions or observational data, with AI-based solutions for in-stream and in-situ data analytics, enhanced IO, or checkpointing to only name a few. Within NumPEx, AI-aided workflows will be investigated for distributed data management and distributed computing.
AI at scale at the Exascale and post-Exascale era
The last few years have seen the very fast (r)evolution of extremely large models, which include large language models or more generally foundational models such as diffusion models, multimodal generative models, etc. These models require extensive use of supercomputers
and huge data storage infrastructures to learn models with billions to trillions of parameters from petabytes ofdata. This raises challenges in data management and computing on Exascale-grade systems, with massive use of heterogeneous accelerated architectures. To that respect, NumPEx contributions to the Exascale software stack will further cover libraries of AI-based components, interface between traditional HPC librairies and standard AI and Machine Learning frameworks (e.g. PyTorch, Tensorflow, Scikit-Learn, etc.). NumPEx is at the cornerstone of the Exascale ecosystem, between the application communities, the HPC vendors, and the applied mathematics and computer science research community.
To that respect, NumPEx will contribute to:
Accelerate AI for Science
NumPEx aims to accelerate the diffusion and application of AI at all stages, with a very strong focus on AI for scientific and engineering applications. To that end, it will help scientific and engineering communities speed up the integration AI at scale with AI-centric co-design. Within the Exascale scientific and engineering user communities, AI uses span a every large spectrum from low-maturity close-toproduction applications to low-maturity proof-of-concept if not uses at all but planned integration. Consequently, NumPEx will further help application communities integrate scale up and deploy AI-based solutions on Exascale platforms. To that end, NumPEx will build upon an AI-centric co-design activity, focusing on shared flexible application use-cases. Comprised with curated data and well-defined application dependent metrics, these usecases will be composed the necessary testbed to test and benchmark AI-based solutions developped within NumPEx. In this context, NumPEx will strongly engage in open science (source and data) developments.
Bridging the gap between application and core-AI communities
The goal of NumPEx is to further build bridges between the traditional HPC and computer science community with core-AI communities from applied mathematics to computer vision or signal processing. For that purpose, NumPEx will participate and organise cross-disciplinary scientific events focusing on AI for Science (e.g. topical long-term programs, schools). The success of AI largely rely on massive open challenges 1, which exposes the needs of the application communities to a much wider audience, allowing for faster discovery. Building upon the co-design activity, NumPEx foster open challenges, focusing on key application use-cases.
Fostering collaborations and initiatives at the national and international level
AI is a complex and very fast evolving ecosystem, which mandates cross-displinary and international collaborations. NumPEx, through its international initiative InPEx, is perfectly connected with the international Exascale community, and several
of the pilot organisms are members of TPC. In this context, NumPEx will actively participate to AI-centric national and European, with a specific focus on enabling the development of large foundational models at scale for Science. NumPEx will further develop and strengthen international collaborations with through the
participation of transnational initiatives such the Trillion Parameter Consortium (TPC).
Contacts
For more information about the AI for science working group and its activities, please contact XXXX.