Windows ML Runtime NVIDIA TensorRT-RTX Execution Provider 2.30.49.0

It is an execution provider module for Windows Machine Learning that enables inference acceleration using NVIDIA TensorRT on supported NVIDIA RTX GPUs. It allows Windows ML to run compatible models using TensorRT for improved performance and lower latency.

You need Windows 10 or later, an NVIDIA RTX GPU that supports TensorRT (Turing/RTX or newer), the appropriate NVIDIA driver for your GPU, and the NVIDIA TensorRT runtime libraries. A recent Windows ML and machine learning SDK or app that supports execution providers is also required.

Install it from the Microsoft Store listing for the package. Follow the Store prompts to download and install. You may also need to install or update NVIDIA drivers and TensorRT libraries separately from NVIDIA's website if your app requires them.

It supports models compatible with Windows ML that can be translated to TensorRT—commonly ONNX models that are supported by TensorRT. Specific operator and version compatibility depends on the TensorRT and Windows ML implementations.

Applications that support Windows ML execution providers will typically detect available providers automatically. Developers can select the TensorRT-RTX provider programmatically when creating a session or using Windows ML APIs that enumerate and choose execution providers.

No—your model files remain unchanged. The provider performs runtime optimization and execution using TensorRT. However, certain unsupported operators or model patterns may require conversion or fallback to CPU/GPU default providers.

Performance gains vary by model, batch size, and GPU. TensorRT often gives faster inference and lower latency compared with generic GPU execution, especially for optimized layers and FP16/INT8 precision. Measure performance in your workload to determine gains.

Verify system requirements: correct NVIDIA GPU, drivers, and TensorRT runtime installed. Ensure the Store package is installed and your app is up to date. Check application logs for provider selection messages and fallbacks. If operators are unsupported by TensorRT, the app may fall back to another provider. Reinstalling drivers or the package can resolve some issues.

The execution provider itself runs locally and is used for on-device model inference. Privacy depends primarily on the application using Windows ML; review that app’s privacy policy. The Microsoft Store listing also provides publisher and privacy information.

Use the Microsoft Store to update the package when updates are available. To uninstall, open Apps & features in Windows Settings or the Microsoft Store app and remove the package. Also ensure any dependent NVIDIA libraries or drivers are managed via NVIDIA’s tools if needed.