Setting up an NVIDIA graphics card

If no Beckhoff hardware with the associated Beckhoff image is used for the GPU-accelerated AI model execution, the customer is independently responsible for creating the necessary framework conditions on his system for operating the graphics card.

If you are using a Beckhoff IPC with GPU and associated Beckhoff image, you can skip this section.

System requirements

System requirements are specific to NVIDIA components. The software versions listed below should not be newer than the specified versions (e.g. do not install CUDA 12.6), otherwise incompatibilities may occur.

NVIDIA drivers

• Normal NVIDIA GPU: Version 560.67
• Quadro/RTX GPU: Version 522.86

CUDA (Compute Unified Device Architecture)

• CUDA Version12.5.1

cuDNN

• cuDNN Version 9.2.1.18 (as zip file, do not use as Windows Installer)
• Installation instructions can be found at NVIDIA: Tarball Installation.
• It is necessary to add the path to the cuDNN libraries to the PATH variable of the system, not that of the user.

In general, the above measures meet the following system requirements:

• Availability of the cuda.dll library
• Availability of the nvml.dll library

Optimized runtime behavior

When using Beckhoff hardware, e.g. C6043 Industrial PC with GPU and associated Beckhoff image, the runtime behavior of the NVIDIA GPU is optimally designed for interaction with the TwinCAT real-time. When using third-party GPUs, considerable runtime fluctuations may occur during model execution, depending on the specific GPU and its settings.

The TwinCAT Machine Learning Server informs about possible challenges of the used graphics card in the log file immediately after starting the server.