Mardi 17 octobre 2023 – Siège CNRS/16ème Paris
[FR] Avec des machines de classe « Exascale » aux USA et bientôt en Europe, nous entrons dans une nouvelle ère du calcul extrême. Le 50ème Forum ORAP avait commencé à en explorer les applications, ce Forum 51 revient sur les tendances en termes de technologies et d’architectures. Cet état des lieux permettra également de se projeter et de discuter les enjeux associés aux machines de classe « post Exascale » attendues avant la fin de la décennie : environnement, souveraineté, programmation, hétérogénéité.
[EN] With Exascale class machines in the USA and soon in Europe, we are entering a new era of extreme computing. The 50th ORAP Forum had begun to explore future exascale application ; this 51st Forum will review Exascale trends in terms of technologies and architectures. This inventory will also make it possible to foresee and discuss the issues associated with « post Exascale » class machines, expected before the end of the decade: environmental stakes, sovereignty, ease of programming, heterogeneity.
Introduction
- Introduction (15″) : Prénom Nom (Institution, Ville)
9:00 – 9:30 Titre
Présentation / Abstract
Bio
9:30 – 10:00 Point France : Projet Jules Verne (S. Requena, Genci)
Présentation / Abstract
10:00 – 10:30 Leonardo Supercomputer and the Big Data Technopole in Bologna (D. Cesarini, Cineca)
Abstract
The supercomputer Leonardo is one of the three pre-exascale supercomputers that will form the European high-performance computing network of EuroHPC Joint Undertaking. Leonardo is co-funded by the EuroHPC and the Italian Ministry of Universities and Research and it’s hosted in the new CINECA’s data centre at the Bologna Big Data Technopole. The Technopole is considered the Italian Data Valley Hub where is located over 80% of total computing capacity in Italy. It’s a complex innovation ecosystem in the Emilia-Romagna region where a broadly ecosystem based of research, businesses and high-tech expertise are focused to extract value from data, carry-out direct decision-making, highlight correlations and protect the environment, citizens and their assets. The Technopole also host the supercomputer of ECMWF and it has been recently announced that the United Nations University (UNU) will establish a research and training institute to catalyse global collaboration and innovation in the utilization of big data and artificial intelligence to manage human habitat change.
Bio
Dr. Cesarini is a Project Manager & HPC Technology Specialist at the HPC department of CINECA where his works is focused on the design of the next-generation HPC architectures to define the strategic roadmap of Italian and European supercomputers. He is Vice-chair for Research of ETP4HP and he is part of the scientific advisory board of EuroHPC JU (RIAG). He manages for CINECA several European founded projects (EPI-SGA1, EPI-SGA2, EUPEX, REGALE, etc.) and national projects of the Italian Recovery and Resilience Plan (PNRR). He graduated in Computer Engineering from the University of Bologna (Italy) in 2014, where he also earned his Ph.D. in Electronics, Telecommunications, and Information Technologies Engineering in 2019.
10:30 – 11:00 Adastra (G. Hautreux, Cines)
Présentation
Adastra, le dernier supercalculateur acheté par GENCI et exploité au CINES, est basé sur les dernières technologies exascales fournies par HPE et AMD. Ce supercalculateur était classé numéro 1 dans le Top500 et numéro 3 dans le Green500 en novembre 22. Il est maintenant classé 12e dans le Top500 mais conserve sa 3e position dans le Green500.
La même architecture a été déployée à Frontier (n°1 au Top500) pour franchir le cap de l’exascale pour la première fois dans le Top500. La machine n°3, achetée par EuroHPC et exploitée à LUMI Finlande, est également basée sur la même architecture. Cette architecture sera présentée et un premier retour d’expérience sera fourni.
Les perspectives de la prochaine étape pour l’exascale au sein de cette architecture seront discutées.Abstract
Adastra, the latest supercomputer bought by GENCI and operated at CINES, is based on the latest exascale technologies provided by HPE and AMD. This supercomputer was ranked # in Top500 and #3 in Green500 in November 22. It is now ranked #12 in Top500 but keeps its 3rd position in Green500. The same architecture was deployed at Frontier (#1 Top500) to break the exascale for the first time in the Top500. The #3 machine, bought by EuroHPC and operated in LUMI Finland is also based on the same architecture. This architecture will be presented, and a first return on experience will be provided. On top of that, perspectives on what will be the next step for exascale in such architecture will be discussed.
Bio
Gabriel Hautreux est responsable du département HPC au CINES et a géré le déploiement du cluster de pointe Adastra, #11 au TOP500 et #3 au GREEN500 en novembre 2022. Ses activités actuelles visent à permettre à la communauté des chercheurs d’exploiter les applications utilisant les technologies exascales, ainsi qu’à augmenter l’efficacité énergétique des applications et à réduire l’empreinte énergétique et carbone des centres HPC.Il est spécialiste des architectures exascales, des nouveaux paradigmes de développement et de l’efficacité énergétique pour les architectures HPC et AI.
Gabriel Hautreux is in charge of the HPC department at CINES and managed the deployement of the leading edge cluster Adastra, #11 at TOP500 and #3 at GREEN500 in November 22. His current activities aim to enable the research community to leverage application using exascale technologies, as well as increasing the energy efficiency of applications and reducing the global energy and carbon footprint of HPC center. He is a specialist in exascale architectures, new development paradigms and energy efficiency for HPC and AI architectures..
11:00 - 11:30 Pause
11:30 - 12:00 Exascale for Science (Frédéric Parienté, Nvidia)
Abstract
The modern supercomputer will leverage the new workloads of AI, Digital Twins, Quantum Computing, to solve the grand scientific challenges of the 21st century.
Bio
Frédéric Parienté (he/him) is a director in the solutions architecture and engineering group at NVIDIA. He is working with European customers in the Higher-Education & Research and Supercomputing sectors notably, on the adoption of HPC and AI. He is also managing the NVIDIA AI Technology Center (NVAITC) in EMEA, whose mission is to enable and accelerate AI research, education and adoption across the network of NVAITC Joint Labs.
12:00 - 12:30 Building, delivering, and sustaining the path to Exascale and beyond with AMD (T. Robson, AMD)
Abstract
While Frontier is the 1st machine to break the Exaflop barrier, it is also one of the most power energy efficient systems in the world, sustaining real scientific production today. This prowess will continue in the coming months and years: AMD is driven by advancing both HPC and AI capabilities, providing TCO advantages while supporting the scientific community to adopt massively accelerated resources. Thus, AMD Instinct has a well-defined path to support users’ needs beyond Exascale.
Bio
Tim Robson has been in the semiconductor industry for more than 25 years with roles encompassing chip design through sales marketing for large multinationals through startups. Currently, he is running WW business development for the Data Center GPU business unit, working with end customers to introduce Instinct for HPC and AI.
12:45 - 14:00 Repas
14:15 - 14:45 TBD (C. Prunty, SiPearl)
14:45 - 15:15 RISC V (A. Bartolini, Université de Bologne)
15:15 - 15:45 Calculateur efficient dédié à l’IA / Efficient AI compute platform (G. Soubrane, Graphcore)
Résumé
Graphcore occupe une position unique dans l’écosystème européen de l’IA, composé d’entreprises capables de mettre sur le marché du matériel dédié ainsi qu’une pile logicielle complète. Nous présenterons notre approche unique permettant d’obtenir des performances sans précédent en tirant parti d’une énorme capacité de calcul parallèle indépendante colocalisée avec une énorme mémoire sur puce. Nous explorerons comment une telle densité de calcul est obtenue grâce à notre architecture MIMD s’appuyant sur les derniers procédés de fonderie et à un traitement BSP en basse précision, réunissant la meilleure efficacité pour les calculs d’IA en termes de performances, de coût et de consommation d’énergie.
Abstract
Graphcore holds a unique position in the European AI ecosystem of companies capable to bring to market dedicated hardware along with a full software stack. We will present our singular approach allowing to drive unprecedented performances leveraging massive independent parallel computing capacity collocated with huge on-chip memory. We will explore how to such a compute density is achieved thanks to our MIMD architecture on latest process nodes and a BSP low-precision processing, together bringing best-in-class AI compute efficiency for performance, cost and energy consumption.
Bio
Gautier Soubrane dirige le développement des activité commerciales pour Graphcore sur al zone EMEA et APAC. Il cumule plus de 20 ans d’expérience dans l’industrie du semiconducteur, des télécommunications, finance, matériel médical. Il est titulaire d’un MBA de l’IMD et d’un Master en Électronique et Télécom de l’ESEO. Avant de rejoindre Graphcore, Gautier a travaillé chez DDN en dirigeant la division commerciale IA en Europe et au début des années 2000, il a contribué fièrement au lancement des tout premiers smartphones et solutions RF+DSP/ARM monopuce depuis la France avec Texas Instruments.
Gautier Soubrane is leading business development at Graphcore in EMEA and APAC regions. He cumulates 20+ years of experience in the semiconductor, telecom, finance and medical devices industries. He holds an MBA from IMD and a Master’s degree in Electronics and Telecom from ESEO. Prior to joining Graphcore, Gautier worked at DDN managing the AI business division in Europe and back in early 2000’s, he proudly contributed to launch the very first smartphones and single-chip RF+DSP/ARM solutions with Texas Instruments.
15:45 - 16:15 Pause
16:15 - 17:15 Quel horizon pour le Zettascale ? (Animation F. Bodin, Université de Rennes)
By 2025, and the arrival of the Exascale Jules Verne machine, the Zettascale race will be at full speed. Core issues such as scalability, complexity management, related human resources, sustainability and environment footprint, as well as energy management and data management during their operation, to name a few, will just be more acute challenges compared to the today’s ones. Furthermore, the vision of Zettascale as one big centralized system may be disputed for eco-systemic, technical and operational reasons. Indeed, more data sources for feeding scientific applications, more constraints on energy consumption, and a shifting HPC market toward cloud technology will very likely promote a vision of the Zettascale as a collaborative system-of-systems.
This round-table will present this forum speakers’ visions.
17:20 Cloture