Performance bottleneck
A performance bottleneck is a resource or dependency that limits progress for the workload and interval being measured.
How it works
A kernel may wait for memory, instruction dependencies or a particular arithmetic pipeline. An application may instead wait between kernels for CPU work or communication. Improve the part on the relevant critical path, then measure again because the limit can move.
Work through an example
If a hypothetical request spends 80 ms outside a 20 ms kernel, making that kernel twice as fast reduces total time from 100 to 90 ms. The request improves by about 1.11×.
Where the term can mislead
A busy resource need not be the resource controlling completion. Correlate counters with the timeline and test a targeted change.
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.