CUDA Graphs
A CUDA graph describes a collection of operations and their dependencies so a prepared executable graph can be launched repeatedly with less host submission work.
How it works
An application builds or captures the graph, instantiates it, then replays it. Reusing the structure can reduce gaps between short operations. Capture rules, updates, buffer lifetimes and graph-associated memory still need explicit handling.
Work through an example
If a hypothetical step spends 30 microseconds submitting work and 70 executing it, eliminating all submission cost would cap the speedup at 100/70, about 1.43×. A real graph retains some overhead.
Where the term can mislead
Graph replay does not make the arithmetic inside an unchanged kernel cheaper. Measure capture cost, memory use and steady execution separately.
Connect the idea
Take it into practice
Continue with the existing concept, apply it to an interview problem, or follow a course lesson. Access requirements are shown before you open a page.
New to GPUs? Start with the free execution-model concept.
Sources and scope
Examples here are teaching calculations, not measured GPU performance. Consult the linked documentation for the exact device, instruction and tool version. Source review: 25 October 2026.