Nvidia unveils Workbench for managing AI workloads, updates Omniverse

Nvidia has recently focused more on its support for AI applications, but it still had plenty of news from CEO Jensen Huang in a keynote address during the annual computer graphics conference, SIGGRAPH.

Huang had a few AI announcements to make, including the release of AI Workbench, a new PC application enterprises can use to help create AI models and deploy them to their data centers or to the cloud. There was also an update to Nvidia AI Enterprise, version 4.0, adding support for the company’s cloud-native NeMo framework to build large language models (LLMs), as well as a new tool to manage multiple instances of Triton inference server to scale AI systems more easily.

Plus, Nvidia is working with Hugging Face, provider of a platform for training and tuning generative AI models, to accelerate model training. Hugging Face will add Nvidia DGX Cloud as one of the cloud-based destinations to which enterprises can send their training workloads.

As befits a conference about computer graphics, though, most of the news revolved around Nvidia Omniverse, the company’s platform for real-time 3D graphics collaboration. This is used across enterprises, from engineering, simulating prototypes, or arranging production lines, to marketing, using CAD files to render photorealistic images, and video for packaging or advertising.

Making USD the common currency

Omniverse is built around Universal Scene Description, a framework for exchanging 3D graphics data developed by Pixar and later released as an open-source project, OpenUSD. Despite its name, it’s not quite universal, with different vendors implementing it in different ways. That prompted Nvidia to team up with Adobe, Apple, Autodesk, and Pixar to form the Alliance for OpenUSD (AOUSD) at the start of August to standardize the specification and promote USD interoperability.

Nvidia hopes this will make USD—and by extension, its USD-related products—more widely adopted.

Those products will soon include ChatUSD and RunUSD, two APIs through which Nvidia will deliver generative AI services to create and manipulate USD graphics.

ChatUSD uses an LLM to answer developers’ queries about how to create and deploy 3D graphical objects, and can respond with text or Python code. RunUSD renders images from OpenUSD files on Nvidia’s cloud-based servers, and functions as a compatibility checker for OpenUSD-format files.

The AOUSD founders also see the organization as a way to coordinate extensions to the standard. Nvidia is already working on specifications to incorporate geospatial data and material, and physical properties in OpenUSD. This will enable the creation of true-to-life digital twins of machines, buildings, cities, and even the whole planet, taking into account details such as the curvature of the earth or the way different objects grip or bounce off one another.

Nvidia also wants to standardize the way different OpenUSD systems handle measurement units; how applications handle dimensions have caused problems for engineers in the past. The way two versions of the same software measured things lead to a problem with the dimensions of wiring looms for the Airbus 380 aircraft, for instance, while the use of two entirely different systems of units in software at each end of a communication link resulted in the 1999 loss of Nasa’s Mars Climate Orbiter.

Kit of parts

Nvidia is expanding the platform in other ways, too. Omniverse Kit, its engine for developing extensions to Omniverse, is getting a refresh, with rendering optimizations, new application templates for OpenUSD, and the addition of an extension registry to simplify adding functionality.

The company is also leaning on OpenUSD to extend its collaboration with existing partners and open up Omniverse to new ones.

It plans to make Adobe’s Firefly family of generative AI models available as APIs; use OpenUSD as a gateway to Wonder Dynamics’ Wonder Studio, a tool for animating and lighting computer-generated characters in live-action scenes; and import 3D models of objects captured from video in USDZ format by Luma.ai.