Skip to content

Performance Guides

Performance material comes in two kinds: numbers already measured, and tools for locating a problem yourself.

  • Benchmarks — a nightly run on an H200, reporting device time per op per workload against the fastest other implementation of the same op. How the numbers are taken and how to read the ratio is set out in "How these numbers are taken" in that section.
  • In-Kernel Timeline Trace — annotate a kernel body with markers and read back a per-CTA timeline: gaps, stalls, and how far the producer and consumer overlap. None of that is visible to a per-kernel profiler. The API is Trace.

The work one call costs comes from op.eval_roofline(), derived from the manifest's roofline field — the ceiling a measurement is read against. The model and the field specification are in Roofline.