CPU cores, vector execution and memory locality work together. A job that spends most of its time waiting for memory does not become compute-bound simply by adding cores.
Profile representative inputs. Record memory per rank, NUMA placement, vectorization and the ratio of arithmetic to data movement. Compare supported software builds with the same accuracy checks.
Evidence to bring
- Problem size and peak resident memory
- Memory bandwidth and access locality
- Compiler, libraries and rank/thread placement
Record the test conditions and the person responsible for validating the result. Resolve the largest uncertainty before treating this as a deployable design.
PRIMARY SOURCES · CHECKED 9 NOVEMBER 2026
NERSC: parallelism →Living documentation; checked 2026-11-09 · Strong scaling fixes total problem size; weak scaling fixes work per processor.
NERSC: Roofline performance model →Living documentation; checked 2026-11-09 · Relates arithmetic intensity to compute and memory limits; measured kernels still matter.