3 ms·

Characterizing Warp Divergence from Pascal to Blackwell

• majke 2mo ago
  > Across all tested generations, divergent paths serialize linearly with the number of paths k, following T(k)≈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 So... Nvidia did a good job?