NVIDIA during the CES 2025 conference today expanded the scope of its platforms for building artificial intelligence (AI) applications that are embedded into physical systems, including robots and autonomous vehicles.
Additionally, NVIDIA revealed it has launched Project DIGITS, an initiative to build a desktop system based on a system-on-chip (SoC), dubbed GB 10 Superchip, that includes Arm processors and Blackwell graphical processor units (GPUs) due out later this year. Priced starting at $3,000, this system is aimed at developers and data scientists that need local compute resources to build AI applications.
At the same time, NVIDIA said it will make additional models from Black Forest Labs, Meta, Mistral and Stability AI as microservices that are embedded within its RTX AI PCs, in addition to providing blueprints that make creating, for example, images and podcasts simpler.
NVIDIA today also unveiled a Llama Nemotron family of open models based on large language models (LLMs) developed by Meta, which will also be added to make building agentic AI applications simpler and is previewing Project R2X, a vision-enabled avatar for RTX AI PCs that makes use of a generative AI algorithm to augment traditional rasterization with entirely generated pixels. The face is then animated using a new diffusion-based NVIDIA Audio2FaceTM-3D model that improves lip and tongue movement.
In general, NVIDIA is trying to expand its reach further into efforts to develop physical systems such as robots that are infused with AI capabilities by providing access to a virtual NVIDIA Cosmos development environment made up of open source foundational AI models, tools for converting video and images into tokens, data processing pipelines, safety guardrails and video processing accelerators.
As part of that effort, NVIDIA is also extending NVIDIA Omniverse, an existing framework for accessing 3D renderings, that has been extended to add a synthetic data multiplication engine to create content for training models along with additional blueprints that make it simpler to build digital twins. An NVIDIA Edify SimReady model has also been added to automatically label existing 3D assets to enable developers to process 1,000 3D objects in minutes. Organizations that have committed to creating 3D include Accenture, Altair, Ansys, Cadence, Foretellix, Microsoft, Neural Concept and Siemens.
Developers using the latest Blackwell class of GPUs will be able to process, curate and label 20 million hours of videos in 14 days, compared to 3.4 years that would be required using a CPU-only pipeline. NVIDIA also claims its tokenizer delivers eight times more total compression and 12 times faster processing than existing tokenizers.
The overall goal is to rescue the cost of training these types of physical systems by reducing the need to build multiple physical prototypes that then need to be trained to perform a specific set of tasks, says Rev Labaredian, vice president of Omninverse and simulation technology for NVIDIA.
Instead, organizations will be able to train the custom models using simulations that are created using massive amounts of photorealistic synthetic data created as inputs or generated by a sensor, including data that, for example, accounts for object permanence.
That approach will make it simpler for developers to access the data needed to cost-effectively model physical environments using a virtual environment, notes Labaredian. “The biggest challenge with physical AI is the data strategy,” he says.
Developers can preview the first of these models on the NVIDIA Application Programming Interface (API) catalog or download the family of models and fine-tuning framework from the NVIDIA NGCTM catalog or Hugging Face cloud service. Cosmos models will also soon be available as NVIDIA NIM microservices that can be integrated into application development workflows using containers.
Companies such as 1X, Agile Robots, Agility, Figure AI, Foretellix, Fourier, Galbot, Hillbot, IntBot, Neura Robotics, Skild AI, Virtual Incision, Waabi, XPENG and Uber, are among the first to adopt Cosmos.
NVIDIA also revealed that Toyota, Aurora and Continental have joined the list of companies building vehicles that incorporate NVIDIA AI technologies.
While it may be a while before the technologies required to build AI applications are pervasively available, the one thing that is clear is that barrier to entry for building them continues to rapidly fall.
