cuDNN
cuDNN is NVIDIA’s library for supported deep-learning operations, including convolution, attention and graph-based operation combinations.
How it works
A frontend describes a computation and its tensor attributes. Backend execution plans depend on the operation, data types, layout, device and available workspace. Plan selection and execution are distinct costs.
Work through an example
For a repeated training shape, cache an appropriate execution plan and measure steady execution separately from plan search. Recheck correctness and workspace when dimensions or precision change.
Where the term can mislead
cuDNN does not implement every possible model operation. Framework dispatch can mix its kernels with other libraries and generated code.
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.