NVIDIA Xavier - JetPack 5.0.2 - Components - CuDNN
cuDNN is a library for deep learning acceleration. Most of the more popular deep learning frameworks use cuDNN as their default GPU implementation, or at least is an available option. cuDNN runs around all NVIDIA GPU platforms, which means that to run a cuDNN application developed on a host you just need to recompile on the new platform. cuDNN is optimized for each NVIDIA GPU architecture, so that performance is not affected by a change of the architecture.
Using cuDNN