Authors: Torsten Hoefler (ETH Zürich, Microsoft Corporation), Satoshi Matsuoka (RIKEN Center for Computational Science (R-CCS)), Rob Schreiber (Cerebras Systems), Jeff Hammond (NVIDIA Helsinki Oy, NVIDIA Corporation), Tal Ben-Nun (Lawrence Livermore National Laboratory (LLNL)), Steve Scott (Advanced Micro Devices (AMD))
Abstract: The rapid evolution of AI is driving unprecedented innovation in computing architectures, from GPUs and AI accelerators to novel memory, networking, and system designs. This BoF brings together leaders from industry, academia, and national laboratories to discuss emerging directions in AI acceleration for future zettascale systems. Rather than focusing on specific products, participants will explore common challenges in architecture, programmability, scalability, energy efficiency, software ecosystems, and deployment at scale. The session will emphasize audience interaction, community discussion, and roadmap development, providing attendees an opportunity to help shape the future of AI acceleration and AI-enabled supercomputing.
Long Description: This BoF will bring together leading architects and practitioners from industry, academia, and national laboratories to discuss the future of AI acceleration and its implications for HPC in the zettascale era. The explosive growth of large-scale AI workloads has triggered unprecedented innovation across the computing stack, including new accelerator architectures, memory systems, interconnects, software frameworks, and system-level designs. At the same time, the HPC community is increasingly confronted with fundamental questions regarding programmability, efficiency, scalability, sustainability, and the long-term convergence or divergence of AI and traditional scientific computing architectures.
The primary goal of this BoF is to facilitate a community-wide discussion on the architectural directions that are emerging beyond today's GPU-centric paradigm. Rather than focusing on specific products or vendor roadmaps, the session will examine common trends, challenges, and opportunities affecting the entire AI and HPC ecosystem. Participating speakers from industry, academia, and national laboratories will provide brief opening perspectives to stimulate discussion, after which the majority of the session will be dedicated to audience interaction and community dialogue.
Topics to be discussed include emerging accelerator architectures, the evolution of memory and storage hierarchies, scale-up and scale-out interconnect technologies, energy efficiency at extreme scale, programmability and software ecosystems, and the role of open standards and interoperability. Additional discussion topics may include the balance between general-purpose and domain-specific accelerators and the implications of rapidly evolving AI workloads for future supercomputing systems.
The relevance of this topic to the SC community is significant and timely. AI workloads are increasingly driving system procurement decisions, influencing datacenter architectures, and shaping future roadmaps for both research and production systems. National laboratories are deploying AI-capable supercomputers at unprecedented scale, while academic institutions and industry are investing heavily in dedicated AI infrastructure. As AI and scientific computing continue to converge, the HPC community faces important choices regarding programming models and ecosystem development. A forum that enables open discussion among accelerator developers, system builders, researchers, and end users is therefore highly relevant to a broad cross-section of SC attendees.
This BoF is designed to be interactive and community-focused. The objective is not to present a series of vendor talks, but rather to create an opportunity for participants to identify common challenges, share experiences across organizations, and discuss future directions for the field. The expected outcome is a richer understanding of the opportunities and obstacles facing next-generation AI systems and a community discussion on priorities for future research, standards, software ecosystems, and architectural innovation.
To the best of our knowledge, this specific BoF has not been held previously at SC. While prior SC events have explored topics related to AI hardware, networking, and large-scale systems, this proposed session uniquely brings together representatives from multiple accelerator developers, academia, and national laboratories to discuss shared architectural challenges and future directions in a single community-driven forum. The emphasis on cross-organizational dialogue, audience participation, and the evolving relationship between AI acceleration and HPC distinguishes this BoF from past discussions and reflects the growing importance of AI-driven computing within the supercomputing community.