In Notch Blocks

Updated: 26 Jun 2026

The ONNX Model workflow can also be used within a Notch Block and deployed to a media server for production use. This provides the flexibility of a standard Notch Block while being empowered by ONNX nodes, enabling a wide range of AI-driven effects and workflows.

Before deploying a Block that contains an ONNX model, ensure that the environment on the target machine is also configured using the same steps outlined for the ONNX workflow on your workstation. These steps can be found in the Working with AI Models guide .

Please note that we provide an integration to load ONNX AI models in Notch. However, we do not provide support for creating, training, modifying, or troubleshooting ONNX models. We do not guarantee that all ONNX models will be compatible with Notch, specific GPU architectures or driver versions. The AI model space is highly dynamic and you should undertake significant testing before attempting to deploy a specific model in your production environment. See Overview and Setup of the manuals for a guide to the setup of, and requirements of, ONNX models.

Using ONNX models inside of Notch Blocks #

When creating a block do not try to embed the ONNX model, instead the playback and build machines should share the same file path for the .ONNX and .TRT files . The model does not require exposed parameters to work correctly.

.TRT files are specific to the GPU, NVIDIA driver version, and CUDA version installed on the machine where they were generated. As a result, each machine will normally need to generate its own .TRT files from the ONNX model.

Only share .TRT files between systems if they are running the same GPU model, driver version, and CUDA version.

Preloading TRT files #

It is also recommended that you run the Notch Block on the target media server before it is required in production. The first time a Block containing an ONNX model is loaded, the model must be unpacked and the corresponding .TRT files generated on the playback machine, which can take some time.

This process can be completed in advance by simply loading the Block on the media server prior to use. Doing so helps ensure that the Block is ready for playback when needed and avoids any delays during production.

Moving Between Machines #

When moving a Block containing an ONNX model from a workstation to a production machine, the ONNX model must be stored in the same file path on both systems. This ensures that the Block can locate and load the model correctly.

If the production ecosystem is made up of the same model of media server, the generated .TRT files can be shared between machines. This can be achieved using a file synchronization tool such as FreeFileSync.

By synchronizing the .TRT files to the same location on each server, you can avoid the need for every machine to generate its own copies.

But be aware that if the ecosystem is not exactly the same on any machine, that machine will have to create its own .TRT files independently, so syncing from another machine will cause issues.

Some manufacturers do not support the installation of newer NVIDIA drivers required for CUDA 12.9. Before planning to use the ONNX model workflow, please verify with the manufacturer that the machine supports the required NVIDIA driver and CUDA version.

Failure to meet these requirements will prevent ONNX models from running correctly on the system.