Dice Question Streamline Icon: https://streamlinehq.com

Cross-CTA Warp Scheduling Behavior under CGGTY

Determine the issue scheduling behavior of NVIDIA Ampere GPUs when selecting among warps belonging to different CTAs, and establish whether the Compiler Guided Greedy Then Youngest (CGGTY) policy holds across CTAs or is modified by CTA-level constraints.

Information Square Streamline Icon: https://streamlinehq.com

Background

The paper presents a Compiler Guided Greedy Then Youngest (CGGTY) issue policy validated by microbenchmarks but notes the validation only covers warps within the same CTA.

A reliable methodology to analyze interactions among warps from different CTAs is not yet devised, leaving cross-CTA scheduling behavior unresolved.

References

However, we have only confirmed this behavior for warps within the same CTA, as we have not yet devised a reliable methodology to analyze interactions among warps from different CTAs.

— Analyzing Modern NVIDIA GPU cores (2503.20481 - Huerta et al., 26 Mar 2025) in Section 5.1.2 (Issue Scheduler: Scheduling Policy)