Attribution of GPU underperformance: accelerator-specific tuning versus kernel granularity

Determine to what extent the limited performance improvements from GPU offloading of the annotated SPH kernels on the Grace–Hopper system are due to insufficient accelerator-specific tuning (e.g., warp-size tailoring) or launching too many small kernels, as opposed to inherent data-transfer and runtime overheads.

Background

The GPU experiments show comparatively low throughput and limited improvements despite the Grace–Hopper architecture’s tight memory coupling. The authors discuss possible reasons, including a lack of accelerator-specific tuning and excessive small kernel launches, and note that naive mapping of individual small compute kernels may not yield satisfactory performance.

They explicitly state uncertainty regarding how much of the observed underperformance is attributable to these factors, indicating the need for a systematic attribution paper.