If one or more GPUs are needed for serving an ML Skill or running a training pipeline, the following requirements must be met:
- The node must have the latest NVIDIA drivers installed. You can download them from here.
- NVIDIA Container Runtime must be installed in order to install and use a GPU. You can download the latest version and check the installation instructions here.
- A Daemonset is needed on the nodes to identify GPUs based on taints. You can get the script from here.
- A mechanism to attach GPUs to pods needs to be set in place. The type of mechanism depends on the kubernetes platform.
Note: Trainable Document Understanding ML Packages provided by UiPath will work on both CPU or GPU for datasets up to 500 images in size. GPU is strongly recommended to achieve faster training times and better model performance. Validation Station retraining loop is not supported on deployments without a GPU, due to the fact that dataset sizes increase too fast and may hit a compute wall with CPU very quickly. If you run UiPath OCR (non-edge version) on AI Center to process more than 2 million pages of documents a year, GPU is strongly recommended for a better product experience.
Updated 3 months ago