Characterizing Warp Divergence from Pascal to Blackwell
This paper tests the assumption of how NVIDIA GPUs handle warp divergence and finds that the cost remains stable and predictable despite architectural changes.
Provides empirical evidence on the stability and predictability of warp divergence cost on NVIDIA GPUs despite architectural changes.
Before reading this…
Applications
- →GPU architecture research
- →Compiler optimization
To understand this paper, make sure you know these concepts first:
- Understanding of GPU architecturefind papers →
- Basic knowledge of compiler optimizationfind papers →
Abstract
More Like ThisSince Volta introduced Independent Thread Scheduling (ITS), NVIDIA GPUs have been widely assumed to handle warp divergence in a fixed manner. We test this assumption across Ampere, Hopper, and datacenter and consumer Blackwell GPUs, using pre-ITS Pascal as a baseline. Combining cycle-accurate microbenchmarks, hardware counters, and static analysis of compiler-generated SASS, we separate stable behavior from architectural change. Across all tested generations, divergent paths serialize linearly with the number of paths $k$, following $T(k) \approx sk$ with no super-linear reconvergence penalty. Warp execution efficiency falls as $32/k$, the penalty is independent of occupancy, and predication removes the serialization cost. The same behavior appears on Pascal, showing that this programmer-visible cost model predates ITS. The compiler-emitted reconvergence machinery, however, has changed substantially. Pascal uses a per-warp SSY/SYNC instruction stack, whereas later generations use barrier-register instructions. Deferred reconvergence beyond the immediate post-dominator falls from 29 cases on Ampere to 2 on Blackwell. Blackwell also introduces a two-tier convergence-barrier classification, uniform-branch instructions, and explicit partial-mask warp synchronization, none of which appear on Ampere or Hopper. Controlled bit-flip experiments indicate that the new barrier class is a static compiler classification with no observable runtime effect in our tests. Thus, divergence retains a stable and predictable performance cost even as NVIDIA's control-flow ISA and reconvergence mechanisms continue to evolve.