Authors: Sagar Dolas (SURF), Olivera Kotevska (ORNL), Tim Kok (SURF), Chester Chen (NVIDIA Corporation)
Abstract: Federated learning (FL) lets institutions train shared AI models without exchanging raw data — a capability increasingly essential for science where data is sensitive, proprietary, or too large to move. Yet FL on HPC remains uncharted: job schedulers, site firewalls, identity federation, and heterogeneous accelerators reshape the problem beyond what cloud-based FL assumes. This BoF brings together framework developers, HPC centers, and scientific users to discuss running trustworthy FL across the world's supercomputers. We will surface open algorithmic, systems, and governance challenges and work toward a community roadmap for cross-facility FL benchmarks, interoperability, and shared practice.
Long Description: Federated learning (FL) is moving from a cloud-and-edge curiosity to a foundational training paradigm for science. Multi-institutional medical imaging, climate and Earth-system modelling, materials discovery with industrial partners, high-energy physics across distributed detectors, and cross-pharma drug-discovery foundation models all share a defining constraint: the data cannot leave its home institution, but the model must learn from all of it. HPC facilities — the very systems with enough compute to train the models science needs — have until recently been absent from the FL conversation, which has overwhelmingly focused on mobile devices or commercial clouds. Recent deployments (APPFL across four DOE leadership-class supercomputers, OmniFed at ORNL, xFFL across European HPCs, SURF's federated infrastructure, NVFlare and Flower in academic-medical settings) show cross-facility FL is now practically achievable — but they also expose problems unique to HPC that the mainstream FL community has not addressed.
This BoF convenes the emerging community of FL-on-HPC practitioners — framework developers, HPC facility operators, scientific application teams, and FL algorithms researchers — to (1) surface the HPC-specific challenges that mainstream FL venues do not cover; (2) compare practical deployment experience across DOE, EuroHPC, and other national systems; (3) identify near-term opportunities for interoperability, benchmarks, and shared tooling; and (4) seed a continuing community of practice.
This session is intentionally complementary to prior SC BoFs on continuum computing, trusted research environments, and data commons. Those sessions treat FL as one ingredient inside broader infrastructure or governance discussions. This BoF treats FL itself as the subject — its algorithms, frameworks, and deployment practice across HPC facilities.
Topics for discussion. Job-scheduler asynchrony and reservation policies for federated rounds; cross-site identity federation (Globus Auth, EOSC AAI) and firewall realities; differential privacy, secure aggregation, and homomorphic encryption at HPC scale; Byzantine robustness and defenses against model poisoning; non-IID data and fairness across federated sites; communication and compression for FL updates over high-bandwidth WANs; energy and sustainability when each client is a multi-GPU HPC job; reproducibility of federated runs; and data-residency and export-control constraints across jurisdictions.
FL on HPC sits at the intersection of three communities SC serves directly: AI/ML practitioners scaling training across facilities, HPC centre staff designing the next generation of cross-facility services, and computational scientists with sensitive or proprietary data. With FL frameworks (NVFlare, Flower, APPFL, OmniFed) now maturing and cross-facility deployments documented in SC25 papers, the topic is timely.
Expected outcomes are : (1) A shared inventory of HPC-specific FL challenges to anchor future work; (2) initial agreement on what a cross-facility FL benchmark suite should measure; (3) a working group of interested participants spanning DOE, European, and international HPC centers; (4) a follow-up community whitepaper and a recurring SC presence — workshop or BoF — to track progress.