hipDNN documentation

2026-03-31

2 min read time

Applies to Linux and Windows

hipDNN (Deep Neural Network) is a graph-based library that improves performance for deep learning workloads on AMD GPUs.

hipDNN uses operation graphs as an intermediate representation for describing computations, allowing different backend engines to optimize and execute them efficiently.

hipDNN provides an interface that follows established deep learning conventions and a plugin-based architecture, which allows advanced users to extend hipDNN without modifying the core library.

Tip

hipDNN provides instructions to Migrate NVIDIA CUDA cuDNN projects to hipDNN.

The public repository for hipDNN is located at ROCm/rocm-libraries.

Note

hipDNN is in beta. Running production workloads is not recommended. hipDNN is only supported on ROCm 7.12.

Install

• hipDNN installation

• Add hipDNN to your project

Conceptual

• High-level architecture

• Engine configuration knobs

• Backend architecture

How to

• Build and execute operation graphs

• Get and set engine knob configurations

• Migrate a cuDNN project to hipDNN

• Develop plugins

Reference

• Operation support

• Environment variables

• Glossary

• API reference

You can find licensing information on the Licensing page.