GPUs and other accelerators need suitable algorithms, supported libraries and enough local memory. Porting only a small part of a program leaves its remaining serial work in the critical path.
Measure host-to-device movement and the complete job. Numerical precision, convergence and reproducibility belong in the comparison. A theoretical FLOP figure is not a forecast for an application.
Evidence to bring
- Accelerated fraction of the runtime
- Device memory and transfer overhead
- Supported precision and validation suite
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: Roofline performance model →Living documentation; checked 2026-11-09 · Relates arithmetic intensity to compute and memory limits; measured kernels still matter.
NERSC: performance portability →Living documentation; checked 2026-11-09 · Architecture diversity makes application portability a design concern.