AI Lab

With Ziguang AI Lab, enterprises or organizations can easily experience end-to-end accelerated computing solutions tailored for AI and data science no matter where they are by using the online accelerated computing platform. These platforms not only provide flexible access, but also focus on the specific needs of AI and data science, from data processing to model training, to reasoning and deployment, to achieve full process acceleration and greatly improve work efficiency.

A select set of experiences designed to get developers, designers, and IT professionals up and running to quickly start building and deploying data-intensive applications. These projects support rapid prototyping and validation, helping to make better and more reliable software and infrastructure choices in the early stages of software development, accelerating the entire development cycle.

Model parallelism-building and deploying large neural networks

Data parallelism-training neural networks with multiple GPUs

Building Transformer-Based Natural Language Processing Applications

Very large deep neural networks (DNNs), whether for natural language processing (such as GPT-3), computer vision (such as large-scale visual Transformer), or speech AI (such as Wave2Vec 2), have special properties that distinguish them from their smaller-scale counterparts. The increasing size of DNNs trained on massive data sets allows them to adapt to new task requirements with only a few more examples, thus accelerating the move towards general artificial intelligence. Using huge data sets, training models with tens of billions to hundreds of billions of parameters is no easy task. This requires a unique method to combine artificial intelligence, high-performance computing (HPC) and system knowledge. The goal of this course is to learn how to train very large neural networks and deploy them into production.

Modern "deep learning" is facing the challenge of increasing data set size and model complexity. Therefore, in order to effectively and efficiently train models, strong computing power is required. Learning distributes data across multiple GPUs during deep learning model training, enabling more deep learning-based application development.

In addition, efficient use of systems with multiple GPUs reduces training time, which speeds up application development and further reduces iteration cycles. Teams that can use multiple GPUs to perform training will have a greater advantage. They can build models that are trained on more data in less time, greatly improving the productivity of engineers.

The use of natural language processing (NLP) has exploded over the past decade. With the emergence of AI assistants in abundance, and enterprises incorporating more human-computer interaction experiences into their business, it is critical to understand how to use NLP technology to manipulate, analyze, and generate text-based data. Modern technology can capture the nuances, contexts, and complexities of language as humans do. If properly designed, developers can use these technologies to build powerful NLP applications that enable natural and smooth human-computer interaction in chatbots, AI voice agents, and many other programs.


Assemble a simple robot using Isaac Sim

Customized development in Omniverse


Step-by-step assembly of a two-wheeled mobile robot in Isaac Sim's real-time GPU environment to complete the "assembly simple robot"

Connect on-premises streaming clients to Omniverse Isaac Sim servers in the cloud

Loading the USD simulation robot into the Isaac Sim environment

Add joint drives and joint attributes to the robot's body

Add joint connections to the robot


Do you want to change the functionality and user interface (UI) of the Omniverse to your liking? Then use Python code to customize the experience of the Omniverse by extending the functionality (Extensions). Extended functions (Extensions) can be used for various modifications, from generating objects by pressing a button to applying custom physical laws on selected objects. Optimize your workflow by copying frequently repetitive actions into extended functionality, or add a new way to manipulate objects in the UI.