The problem

What does a throughput target mean when the device changes frequency under power or thermal pressure?

What current research shows

GPU performance depends on execution configuration, memory behavior, and the device's available operating point. Power and thermal controls can change clocks and therefore change the time required for the same work.

Where the evidence stops

The magnitude and direction of the effect are hardware- and workload-specific. A short run can miss thermal steady state.

What Valen Systems is testing

Report work, clock behavior, temperature, power, and time over a sustained run rather than treating peak throughput as a constant.

Sources

NVIDIA CUDA C++ Programming GuideNVIDIA CUDA C++ Best Practices Guide