A surrogate model or learned component changes a scientific workflow only when accuracy and validation hold. Training, inference and simulation have different resource profiles and should be scoped separately.
System considerations
Record numerical precision, model quality, memory capacity and interconnect traffic. Validate the complete workflow against the scientific baseline rather than comparing unrelated benchmark scores.
Evidence to bring
- Scientific baseline and validation metric
- Training versus inference resource needs
- Application profile and data-governance boundary
This is an illustrative research or industry context, not an available World Supercomputing service or a promised scientific outcome.
PRIMARY SOURCES · CHECKED 9 NOVEMBER 2026
HPE: open and heterogeneous AI infrastructure →2026-10-12 · Executive perspective; architecture direction, not independent performance validation.
NERSC: Roofline performance model →Living documentation; checked 2026-11-09 · Relates arithmetic intensity to compute and memory limits; measured kernels still matter.