View page source

Execution and Performance#

Warp performance depends on more than device kernel time. Python submits work, Warp may compile or load native modules, arrays may allocate or move memory, and CUDA devices usually execute their queued operations asynchronously. Separate those costs before deciding what to optimize.

Guides in this section#

  • Concurrency: Streams, events, synchronization, graph capture, and multi-device execution.

  • Memory Allocation and Access: Allocation strategies, memory pools, residency, and cross-device access.

  • Profiling: Measuring host and device work with Warp timers, NVTX, and NVIDIA profiling tools.

  • Deterministic Execution: Reproducibility guarantees, limitations, and performance tradeoffs.