User Guide

Jetson AGX™ Orin Developer Kit Reviewer's Guide
10
NVIDIA TAO (Train-Adapt-Optimize) is a framework that lets developers create custom, production-
ready models, in hours rather than months, without AI expertise or large training datasets. The NVIDIA
TAO Toolkit abstracts away the AI/deep learning framework complexity, letting you fine-tune on high-
quality NVIDIA pre-trained AI models with only a fraction of the data compared to training from scratch.
Customers can use the TAO Toolkit to fine-tune and optimize on a wide variety of use cases from
computer vision and automatic speech recognition to speech synthesis and natural language
understanding.
Data collection and annotation is an expensive and laborious process. Simulation can help bridge the
need for data. NVIDIA Omniverse Replicator uses simulation to generate synthetic data that is an
order of magnitude faster and cheaper to create than real data. With Omniverse Replicator you can
quickly create diverse, massive and accurate datasets for training AI models.
We have provided you with a demo that will enable you to use TAO to train a model in the cloud and
deploy the trained model on Jetson using DeepStream. Please refer to NVIDIA TAO section in
Appendix for instructions on running the demo.
NVIDIA Isaac ROS GEMs are hardware-accelerated packages that make it easier for ROS developers
to build high-performance solutions on NVIDIA hardware. NVIDIA Isaac Sim, powered by Omniverse,
is a scalable robotics simulation application. It includes Replicator - a tool to generate diverse synthetic
datasets for training perception models. Isaac Sim is also a tool that powers photorealistic, physically
accurate virtual environments to develop, test, and manage AI-based robots.
NVIDIA RIVA is an SDK for building GPU-accelerated conversational AI applications. RIVA includes
state of the art pre-trained models for Automatic Speech Recognition (ASR) and Text-To-Speech
(TTS). These pre-trained models are highly accurate and can be easily customized using the TAO
Toolkit to improve accuracy on desired domains, accents, languages and use cases. NVIDIA RIVA
speech models are optimized for TensorRT to deliver high inferencing performance and low
latencies on Jetson AGX Orin.
We have provided you with a RIVA ASR demo which is a dictation application that showcases
the performance of Jetson AGX Orin and the accuracy of the pre-trained speech recognition
neural networks. Please refer to NVIDIA RIVA section in Appendix for more details and
instructions to run these demos.
DeepStream is an SDK for rapidly developing and deploying Vision AI applications and services.
DeepStream offers hardware acceleration beyond inference as it offers hardware accelerated plugins
for end-to-end AI pipeline acceleration. It offers state-of-the-art throughput. Developers can also bring
their own TensorFlow, PyTorch, or ONNX models and deploy them using DeepStream.