Impact of limited FP64 performance on observed GPU results

Determine the extent to which limited double-precision (FP64) support on the GPU contributed to the observed performance outcomes of AoS-to-SoA annotated SPH kernels.

Background

The authors note the GPU’s inferior FP64 performance relative to other precisions and speculate that this may have affected their results. They indicate that converting kernels to FP32 might qualitatively change the outcome but would require care due to potential numerical instabilities.

They explicitly state that it is not clear what role FP64 limitations played, motivating targeted experiments to isolate the effect of precision on performance.