Characterizing Warp Divergence from Pascal to Blackwell
Summary
The arXiv paper Characterizing Warp Divergence from Pascal to Blackwell investigates warp divergence behavior across NVIDIA GPUs from pre-ITS Pascal through Ampere, Hopper, and Blackwell. It shows warp divergence cost scales linearly with the number of divergent paths and remains stable across generations, while the compiler's reconvergence machinery evolves (Pascal using per-warp SSY/SYNC stacks, Blackwell moving to barrier-register instructions). The study also notes that predication can remove serialization cost and that divergence costs are largely predictable despite architectural changes.