Tales of Exascales – AI and HPC Supercomputing Platforms Synergies for Large Language Models (LLMs) and Scientific Workflows

The BOF session aims at engaging ISC24 attendees and promoting information exchange on contemporary approaches for architecting, implementing, service delivery, cybersecurity and quality-of-service for Exascale AI and HPC platforms for open research. A target outcome is to identify synergies while exploring the current state-of-the-art and the state-of-the-practice. AI research and development (R&D), particularly Exascale LLMs, has been largely led by proprietary solutions leveraging Cloud technologies. These include virtualisation for continuous integration and delivery of software pipelines and XaaS (infra-, platform-, and software-as-a-Service) delivery models. HPC, particularly Exascale scientific simulations, maintains specialised stacks for a predefined service or workflow such as numerical weather prediction. Changes within a classic HPC environment are typically infrequent due to the vertical integration of user and system software such as math libraries, drivers for high performance network and accelerators to name a few. Other distinctions include resource management for batch vs interactive and elastic scheduling, and cloud native vs Unix style identity and access management. Funding agencies around the world have ongoing Exascale R&D initiatives for the development of applications, platforms, and infrastructure. Exascale HPC platforms have been formally launched in the US (Top500) while proprietary, trillion parameters LLMs exploiting Exascale capabilities have been introduced with reduced-precision AI Flops. Recently national AI research resources for open research have been introduced int the UK (AI RR) and USA (NAIRR) to support open research. Attendees will have opportunities to engage with panellists representing different perspectives for AI for science, national Exascale programs and vendors.