Get Started
There are two ways to run HIP EP on an AMD GPU on Windows: the binary package and the Python package.
Both paths assume a Windows system with a supported AMD GPU and a current AMD graphics driver. Each path takes you through installation and ends with a model running on the GPU. The main difference is how HIP EP is installed and what tools are available afterward. Choose the path that matches how you plan to use HIP EP; the two paths are independent.
Building HIP EP from source is not covered here. It is documented in the repository, in the Windows quick start.
1 · Binary package
Extract the release zip, add bin to
PATH, and run provided tools. The archive contains the required runtime libraries and binaries, so no separate Python or Visual Studio installation is required.
2 · Python package
Install the required wheels into a Python 3.14 environment. The installation order matters, and pip also installs the required ROCm runtime packages. The Python package provides benchmarks and example scripts that you can inspect and modify rather than a set of pre-built executables.
Which one
| If you want to | Read |
|---|---|
| Run the shipped models and benchmark them with minimal setup | Run models with the binary package |
| Use ONNX Runtime and OGA from your own Python code | Run models from Python |
Check that your GPU is covered
The Windows package includes kernels for three RDNA 3.5 integrated GPUs, so
there is no architecture-specific download to choose. Your GPU must be one of
the supported targets, however.
The benchmark data on this site was collected on Strix Halo
(gfx1151); Strix Point and Krackan Point are also included in the package, but
you should complete the GPU verification step on the corresponding installation
page before relying on the package for your system.
Get-CimInstance Win32_VideoController |
Select-Object Name, DriverVersion, AdapterCompatibility
| Adapter name contains | Architecture | Product family |
|---|---|---|
| Radeon 8060S / 8050S | gfx1151 |
Ryzen AI Max 300 (“Strix Halo”) |
| Radeon 890M / 880M | gfx1150 |
Ryzen AI 300 (“Strix Point”) |
| Radeon 860M / 840M | gfx1152 |
Ryzen AI 300 (“Krackan Point”) |
If your adapter is not listed — for example, a discrete Radeon GPU, an older
integrated GPU, or an Instinct GPU — do not assume that the Windows release
package supports it. The three architectures listed above are the targets
covered by the validation on this site. Other architectures have to be built
from source with --hip_arch, which the repository’s
Windows quick start
covers.
Also check that your AMD graphics driver is up to date. HIP EP uses the bundled HIP runtime to communicate with the kernel-mode driver, and an outdated driver can cause launch failures that are otherwise difficult to diagnose. If necessary, install the current AMD Adrenalin driver.
Before you start
The Windows binary package does not require a separate ROCm installation. The archive includes the HIP runtime, the code-object manager, hipBLASLt, rocBLAS and MIOpen required by the EP. A current AMD graphics driver and the release download are sufficient.
The Python package is different. Its EP wheel declares the ROCm runtime as a dependency instead of bundling it, so pip downloads the required ROCm packages from an AMD package index during installation. See that page for the index URL and installation details.
What you get
The three routes do not carry the same tools, and the differences are worth knowing before you pick one:
| Tool | What it is for | Where it comes from |
|---|---|---|
hip-onnx-runner |
Run an ONNX model through the EP; dump outputs; compare against CPU results | binary package only |
hip-compiler, hip-mlir-opt, hip-inspect |
Compile and inspect the MLIR pipeline directly | binary package only |
onnxruntime_perf_test |
Measure steady-state inference latency for a single graph | binary package only |
model_benchmark |
Run an end-to-end generative benchmark, including prefill and decode, through OGA | binary package only |
model_mm, benchmark_multimodal.py |
Run the corresponding benchmarks for vision-language models | binary package only |
run_onnx.py, benchmark_e2e.py, vlm_benchmark.py |
Run text and vision-language benchmarks as editable Python scripts | Python package only |
The binary package provides the most complete set of pre-built tools. The Python package provides the benchmark logic as Python scripts rather than as standalone executables.
The binary package does not install anything into a system directory or register global components. To remove it, delete the extracted directory.
The Python package is installed through pip, so it can be installed in a virtual environment and removed with the environment when no longer needed.