Co-Designing Next Generation Supercomputing Systems

Following the transition to exascale, high-performance computing (HPC) is entering a new phase in which future supercomputing systems are defined not only by peak capability, but by their role within a heterogeneous, end-to-end computational ecosystem. Scientific discovery increasingly depends on workflows that span edge devices for data acquisition, intermediate “fog” layers for data movement and preprocessing, disaggregated and fabric-attached resources, and centralized systems for large-scale simulation—augmented throughout by artificial intelligence (AI). At the same time, global semiconductor initiatives are creating new opportunities to rethink how supercomputing systems are architected, customized, and realized.
This BoF focuses on end-to-end co-design as the key enabler for this next generation of systems, bringing together domain scientists, application developers, software and system researchers, hardware architects, integrated circuit (IC) designers, and device technologists. Without such coordinated co-design—and without access to open, interoperable toolchains—innovation risks remaining trapped in the “valley of death” between architectural concepts and deployable systems, limiting scientific customization to what can be justified by broad commercial markets.
A central theme of the session is the emerging role of chip prototyping as a critical mechanism for bridging this gap. In a post–Moore’s Law era, full-scale production designs at leading-edge technology nodes are increasingly slow and costly, making early-stage prototyping, minimal viable products (MVPs), and chiplet-based systems essential for architectural innovation. Rapid design–fabrication–test cycles enable proof-of-concept systems, provide feedback that cannot be captured through simulation or emulation alone (e.g., energy efficiency, security, and system-level behavior), and attract early software engagement—enabling true hardware–software co-design.
The BoF will also explore how future supercomputing systems are both designed for AI and increasingly designed by AI. On one hand, emerging workloads demand architectures that efficiently support large-scale AI training, inference, and hybrid simulation–AI workflows. On the other, massively parallel computing systems are becoming foundational tools for the chip design process itself, enabling generative and agentic AI approaches to architecture exploration, electronic design automation (EDA) optimization, multi-physics chiplet analysis, and system tuning.
Finally, the discussion will address how novel paradigms—quantum computing, analog computing, neuromorphic computing, and hybrid systems—can be integrated into coherent hardware and software stacks through co-design and prototyping. The goal of this BoF is to identify concrete pathways by which the research community can act as “lead users,” accelerating innovation in next-generation supercomputing systems for AI-enabled scientific discovery.

---

Organizers:

• Antonino Tumeo
• James Ang