Characterizing Warp Divergence from Pascal to Blackwell

A new study finds that on NVIDIA GPUs from Pascal through Blackwell, warp divergence cost grows linearly with the number of paths, with no extra reconvergence penalty.

This means divergent execution remains predictable across generations, making the basic cost model stable over time.

At the same time, the compiler-emitted reconvergence machinery changed substantially, moving from Pascal’s SSY/SYNC stack to newer barrier-register-based mechanisms and Blackwell-specific control-flow features.

The results help programmers and compiler developers reason about performance portability and control-flow optimization.