Birds of a Feather Archive

Confidential Supercomputing: Essential Evolution or Elegant Illusion?


Authors: Sadaf Alam (University of Bristol, Bristol Centre for Supercomputing), Hai Ah Nam (Lawrence Berkeley National Laboratory (LBNL)), Albert Reuther (MIT Lincoln Laboratory), Nathaniel McCallum (Advanced Micro Devices (AMD)), Nigel Edwards (HPE), CJ Newburn (NVIDIA Corporation), Katerina Antypas (National Science Foundation (NSF))

Abstract: Sensitive workloads across health sciences, financial modeling, and AI/ML on proprietary data increasingly demand confidentiality guarantees traditionally out of scope for supercomputing infrastructure. Confidential computing—hardware-enforced trusted execution environments established in cloud and edge contexts—is proposed as a path to securing computation at HPC scale. But is this achievable, or does it conflict with the principles that make supercomputers fast? This BoF gathers the HPC community to debate whether confidential supercomputing is a necessary, achievable evolution or an aspirational posture satisfying compliance optics without meaningful security. Join us to map where it genuinely delivers—and where gaps remain.

Long Description: Supercomputing Digital Research Infrastructure (DRI) has historically optimised for throughput, openness, and shared access — principles that sit in tension with confidentiality. Yet a growing class of workloads arrives with hard confidentiality requirements: patient-level medical data governed by regulation, proprietary financial models, and AI training over commercially sensitive or personal datasets. These communities want the capability of HPC without surrendering control of their data, and increasingly ask facility operators a pointed question: can you run my workload confidentially, at scale?

Confidential computing offers one answer. Hardware-enforced trusted execution environments (TEEs), attestation, and encrypted memory are well established in cloud and edge settings, and vendors are now extending these primitives toward accelerators and multi-node configurations. But supercomputers derive performance from tight coupling, high-bandwidth interconnects, shared memory, and minimal per-operation overhead — precisely the surfaces that confidentiality mechanisms constrain. The central question this BoF poses is whether confidential supercomputing is a necessary and achievable evolution of DRI, or an aspirational posture that satisfies compliance optics without delivering meaningful security at scale.

This BoF builds directly on a well-attended panel at ISC 2026 in Hamburg, which brought together speakers from multiple international HPC sites and several hardware and technology vendors to debate confidential supercomputing before a strong and visibly divided audience. That the debate spanned competing institutions, vendors, and national contexts — rather than a single roadmap — underscored both the timeliness of the topic and the absence of community consensus. The questions left unresolved when time ran out demonstrated clear appetite for a deeper, participation-driven conversation, and a BoF is the natural venue to move from expert debate to community-wide deliberation.

The session convenes facility operators, security researchers, hardware and system-software specialists, and — critically — the research communities generating sensitive workloads. Short provocations from contributors holding deliberately contrasting positions will seed a structured, audience-led discussion. We aim to surface which threat models justify the engineering trade-offs; where current TEE and attestation approaches break down at multi-node and accelerator scale; what performance overheads facilities and users will tolerate; and whether "confidential" claims survive contact with real HPC architectures.

Expected outcomes: a shared, community-generated map of where confidential computing genuinely delivers at HPC scale versus where gaps remain wide open; identification of concrete open problems for researchers and vendors; and the beginnings of an international, cross-institutional community of practice around confidential DRI. Attendees will leave able to reason critically about confidential-computing claims and to assess whether the approach fits their own facility or workload. We anticipate strong interest from operators, security researchers, domain scientists in regulated fields, and vendors — a timely conversation as these capabilities move from roadmap to procurement.


Back to Birds of a Feather Archive Listing