By VR Top Wan Li

This morning, NVIDIA CEO Jensen Huang delivered a keynote speech at CES 2025.

Besides Huang's signature leather jacket, the event unveiled multiple new products and solutions, including the RTX 50 series graphics cards, AI PCs, autonomous driving systems, and the groundbreaking world foundation model, Cosmos.

It's evident from this conference that NVIDIA's AI ambitions extend far beyond being just a chip manufacturer.

RTX 50 Series Graphics Cards and Ultra-High-Performance AI PC

GTX 50 Series Graphics Cards

The 50 series graphics cards all feature the Blackwell architecture, upgraded to PCIe 5.0 interface, and support new features like DLSS 4. The RTX 5090 graphics card, specifically introduced, is more expensive but also more powerful than the 4090: Price: $1999, with a suggested retail price in China of 16499 yuan.

In terms of specific parameters, the RTX 5090 boasts 92 billion transistors, 21760 CUDA cores, equipped with 32GB DDR7 memory, and AI computing power of up to 2375 TOPS. Thanks to architectural improvements, the GPU performance of this new graphics card is essentially double that of its predecessor.

Furthermore, the conference mentioned that the RTX 5070 graphics card offers performance comparable to the 4090 but at a more affordable price of $549, making it a promising budget option.

It is understood that for desktop graphics cards, the RTX 5090 and RTX 5080 GPUs (8299 yuan) will be available at the end of January; RTX 5070 Ti and RTX 5070 will be available in February. For laptop products, GeForce RTX 5090, RTX 5080, and RTX 5070 Ti laptops will be available in March, with RTX 5070 series laptops hitting the market in April this year.

RTX 50 Series Specifications, Image Source: Network

NVLink72: NVIDIA's Server Interconnect Technology

The shield-like object held by Huang at the beginning of the presentation was to demonstrate the company's latest AI server interconnect technology, NVLink72.

Based on NVLink72, a total of 72 Blackwell GPUs and 36 Grace CPUs can be concentrated on a single silicon chip. If a similar product were built using traditional architecture, it would be as large as a wardrobe.

Key parameters and features of NVLink72 include: Based on the Blackwell framework; it provides a total bidirectional bandwidth of 1.8TB/s, which is twice that of the previous generation; the total bandwidth of NVLink72 is more than 14 times higher than PCIe Gen 5; multiple NVLink72s can be connected via NVLink Switch.

Ultra-High-Performance AI PC

Project Digits, NVIDIA's first AI supercomputing host, was introduced as the 'One more thing' at the conference.

Project Digits has a compact design similar to the Mac mini, but inside it houses a GB10 Grace Blackwell superchip with an AI computing power of up to 1 petaflop, enabling a single device to run AI models containing 200 billion parameters. MediaTek also participated in the chip design of GB10.

Additionally, Project Digits is equipped with 128 GB RAM and 4TB ROM. The host is powered by the Grace CPU, which consists of 20 Arm cores and includes components such as Connectivity and ConnectX NIC.

Project Digits runs on Nvidia DGX OS, a Linux-based system that integrates NVIDIA's series of AI software libraries. The product is priced at $3000 and is scheduled to be available in May this year. During the conference, Huang stated that the product is suitable for "every data scientist, AI researcher, and student, enabling them to participate in and shape the era of AI."

From AI Models to Autonomous Driving and Embodied Intelligence: NVIDIA's Latest AI Blueprint

Large Language and World Models

Two new AI models were introduced at the conference. NVIDIA Llama Nemotron is a large language model built on the Llama foundation model, using NVIDIA's latest technology and high-quality datasets for pruning and training to enhance agent capabilities. They excel in following instructions, chatting, function calls, coding, and mathematics, while being optimized in size to run on various NVIDIA accelerated computing resources.

Llama Nemotron comes in three versions: Nano, Super, and Ultra. The Nano version is optimized for low-latency real-time applications and is ideal for deployment on PCs and edge devices. This model will soon be available for free download or can be accessed via NVIDIA NIM.

In addition to the large language model, NVIDIA also unveiled NVIDIA Cosmos for the first time, a platform comprising advanced generative world foundation models (WFMs), tokenizers, guardrails, and accelerated video processing pipelines, aimed at accelerating the development of physical AI systems such as autonomous vehicles (AVs) and robots.

A world foundation model is one that understands the world's language, physical properties, spatial locations, and causal relationships. It is crucial for accelerating the popularization of AI terminals such as smart cars and embodied intelligence. Compared to the booming development of LLMs, world models are still in a relatively early stage. The creation of world models also involves massive data capture and training based on real-world data, making the development cost significant.

Now, the Cosmos model is available under an open model license and can be downloaded and used via NVIDIA API, NVIDIA NGC catalog, or Hugging Face, which is undoubtedly good news for developers in fields such as robotics.

NVIDIA Cosmos WFM has been trained on 180 quadrillion tokens, including 20 million hours of real-world autonomous driving, robotics, drone footage, and synthetic data. Cosmos WFM provides developers with a simple way to generate a large amount of physically based photorealistic synthetic data for training and evaluating their existing models. Developers can also build custom models by fine-tuning Cosmos WFM.

Huang said, "The ChatGPT moment for robotics is coming. Like large language models, world foundation models are the foundation driving the development of robots and autonomous vehicles, but not all developers have the expertise and resources required to train their own models. We created Cosmos to democratize physical AI, making general-purpose robotics technology accessible to every developer."

NVIDIA Cosmos's partners include robotics and intelligent driving enterprises such as 1X, Agile Robots, Agility, Figure AI, Foretellix, XPeng Motors, and Uber.

Based on NVIDIA Cosmos, Huang also showcased more eye-catching use cases at the conference, such as combining it with its digital twin platform Omniverse to create a synthetic data multiplication engine, allowing developers to easily generate large amounts of controllable, photorealistic synthetic data. Developers can author 3D scenes in Omniverse and render images or videos as output. These scenes can then be combined with text prompts to adjust the Cosmos model, generating countless synthetic virtual environments for physical AI training.

Autonomous Driving and Embodied Intelligence

Regarding autonomous driving and robotics, Huang stated that future enterprises will require three sets of computing systems: DGX for AI model training, Omniverse for simulation, synthesis, and data enhancement, and DGX AI chips for terminals. (This also provides a good summary of NVIDIA's current and future core business focus)

Mentioning autonomous driving, Huang pointed out that "this could be the first trillion-dollar robotics industry." One of the major announcements he made today was that NVIDIA will collaborate with Toyota to develop autonomous vehicles.

Furthermore, the conference also introduced the new generation of intelligent driving chip DRIVE AGX Thor, which is based on the Blackwell architecture and offers 20 times the computing power of the previous generation. Today, Zeekr announced that it will be the first to launch this chip.

NVIDIA DRIVE Hyperion is an autonomous driving system for vehicles. Today, Huang announced that the system has passed industry safety assessments by TÜV SÜD and TÜV Rheinland, making it the first and only end-to-end autonomous driving platform in the industry, with current partners including Mercedes-Benz, Land Rover, and Volvo.

In addition to vehicles, Huang also unveiled the NVIDIA Isaac GR00T synthetic motion generation blueprint for humanoid robots. Currently, an important area in humanoid robotics is imitation learning, where skills are acquired by observing and imitating human behavior. This also involves a significant amount of tedious data collection work behind the scenes.

NVIDIA Isaac GR00T is divided into three major sections: The GR00T-Teleop workflow captures human motion data in digital twins based on Vision Pro; the GR00T-Mimic workflow uses the captured human demonstrations to build a larger synthetic motion dataset. Finally, the GR00T-Gen workflow, built on the NVIDIA Omniverse and NVIDIA Cosmos platforms, exponentially expands this dataset through domain randomization and 3D upgrades.

Capturing operational data based on Vision Pro, image source: NVIDIA

Conclusion

Founded in 1993, NVIDIA launched CUDA in 2006, which can be seen as an important cornerstone of the company's current AI chip empire.

Today, beyond its graphics card and AI chip businesses, NVIDIA's reach has clearly extended further, such as its ongoing efforts in autonomous driving systems and the highly imaginative Cosmos.

Moreover, it's evident from Huang's enthusiastic antics at the event that today's NVIDIA remains vibrant and full of energy.