Image source: generated by Unbounded AI

Is there anyone more suitable to be the opening keynote speaker of CES in 2025 than Nvidia founder Jensen Huang?

As an event marking the official unveiling of CES2025, Huang Jen-Hsun’s keynote speech has ignited everyone’s expectations for the opening event of the year from the beginning.

On January 6, local time, Huang Jen-Hsun changed into a shiny snakeskin jacket—not the iconic jacket he often wears at conferences. "How about my jacket?" Lao Huang used this leather jacket to joke with people at the beginning.

After all, this is Las Vegas, and anything less than "exaggerated" would be disrespectful to the place where CES is held.

The shiny leather jacket worn by Jen-Hsun Huang not only represents Nvidia’s exaggerated market value performance this year, but also means that in the opening speech that night, he will release a series of explosive products.

If you think that new “nuclear bombs” like RTX 5090 are the focus of Huang’s speech, then you obviously underestimated the conference that night. NVIDIA has long become the "arsenal" of the AI ​​industry. At the press conference, Huang Renxun not only announced Omniverse+Cosmos, an AI model based on the physical world and a complete set of production processes, but also exposed an unnamed desktop-level AI supercomputer. Project Digits".

To sum up, the reins of AI development are still firmly in the hands of NVIDIA.

01, RTX5090 series: Big discount for increasing computing power

RTX5090 series｜Image source: NVIDIA

First up is the highlight of this year’s flagship graphics card: the RTX50 series.

Although it is under pressure from AMD, after months of rumors, Nvidia finally officially released the next-generation RTX Blackwell GPU series at CES 2025, including a total of four models:

RTX 5090, priced at $1,999 RTX 5080, priced at $999 RTX 5070 Ti, priced at $749 RTX 5070, priced at $549

Among them, the RTX 5090 and RTX 5080 will be released on January 30, and the RTX 5070 Ti and RTX 5070 will be released in February.

The RTX 50 series features a new Founders Edition design, equipped with bi-directional flow fans and GDDR7 video memory, and all models support PCIe Gen 5 and DisplayPort 2.1b connectionsport, can drive display output up to 8K/165Hz.

Thanks to new architecture Blackwell and DLSS 4, the RTX 5090 is said to perform up to two times faster than the RTX 4090. Power consumption is also increased by 125 watts, but Nvidia says it's more efficient, reaching 575 watts only at full load.

The performance of the RTX 5080 has also been greatly improved, and it is said to be twice as fast as the RTX 4080; Huang Renxun also said that the performance of the RTX 5070 is equivalent to the RTX 4090, but the price is only $549.

Nvidia claims that the RTX 5070 Ti is twice as fast as the RTX 4070 Ti, and the RTX 5070 has the same performance as the RTX 4090, priced at just $549.

The performance of the RTX50 series has been "doubled" compared to the 40 series | Image source: NVIDIA

But in this "doubled" performance improvement, the new generation supports multi-frame The generated DLSS 4 technology also plays an important role, it can generate up to 3 additional frames based on traditional rendering, and the frame rate can be increased by up to 8 times. It also leverages generative AI to improve material compression and game character facial quality.

Huang Renxun also bluntly said, "Without generative AI, we would not be where we are today."

In the demo, Nvidia showed that the RTX 5090 achieved 238fps with DLSS 4 enabled in Cyberpunk 2077, while the RTX 4090 could only achieve 106fps with DLSS 3.5. Both GPUs have full ray tracing enabled.

DLSS 4 also implements a real-time Transformer model that improves image quality, reduces smearing and provides greater detail in dynamic footage. Some of the upgrades to DLSS 4 will also be compatible with existing RTX GPUs.

In addition to the desktop graphics card series, NVIDIA will also launch the RTX50 series of GPUs on laptops, which will be launched by multiple PC manufacturers starting in March.

Of course, we cannot forget the RTX5090D version specially provided for the Chinese market: Judging from the information currently disclosed on NVIDIA’s official website, the RTX5090D has been mainly reduced in AI computing power, and the performance of other parts has not changed.

Performance difference between RTX5090D and RTX5090｜Image source: NVIDIA official website

02. Omniverse+Cosmos: World Model and Digital Twin< p class="ne-empty-p">

In addition to the RTX5090 series graphics cards, supporting the BritishVida's stock price has skyrocketed due to Nvidia's progress in the world model field.

In the keynote speech, Huang Renxun introduced a series of basic models of the world called Cosmos: these models can predict and generate "physics-aware" videos, and their purpose is to solve the current problem of AI. The problem of "knowing what is happening but not knowing why" is the problem of the model.

Currently, the Cosmos series models released by NVIDIA are divided into three major categories:

Nano: for low-latency and real-time applications. Super: for high-performance baseline models. Ultra: For maximum quality and high-fidelity output.

The model parameter sizes range from 4 billion to 14 billion, with Nano being the smallest model and Ultra being the largest. While more parameters generally lead to better performance, developers can still fine-tune it for specific applications, and it’s available through NVIDIA’s API and NGC catalog, GitHub, and the Hugging Face AI development platform.

According to Huang Renxun’s speech, these models are trained based on 900 trillion tokens and 20 million hours of real-world human-computer interaction, environment, industry, robotics and driving data, but they are not The specific source of these learning data was not disclosed: it has been previously reported that the source of these data is YouTube videos.

In this regard, NVIDIA simply stated that "the data used comes from a variety of public and private sources, and these are in compliance with legal requirements."

Back to the world model itself, with massive real-world video data as support, the Cosmos WFM model can generate "controllable high-quality" synthetic data based on text or video frames for use in robots and autonomous driving. Model training and development in automotive and other fields.

Cosmos can output reasonable images for training robots based on the laws of physics｜Image source: NVIDIA

Although there are some deviations from the "open source model" in the traditional sense, NVIDIA still allows it Regardless of company size, researchers and developers are free to use Cosmos models under NVIDIA's Open Model License, which allows commercial use.

According to the cases shown by NVIDIA, Cosmos has been used to simulate real environments, such as factory floors or driving scenes. Or, like Sora, use inputs from multimodal content such as text, images, video, and robot sensor data to generate videos based on the laws of physics.

Currently, autonomous driving companies including Wayve and Uber are using Cosmos to accelerate the advancement of autonomous driving technology. Huang Renxun also expressed the hope that Cosmos "can bring changes to the fields of robotics and industrial AI like Llama has brought to enterprise AI."

In additionWorld Model, and also announced new results in the field of digital twins from NVIDIA: Mega Omniverse Blueprint, a framework for creating industrial digital twins.

Nvidia said that this new framework brings a new platform for industrial AI and robot simulation to factories and warehouses through software-defined testing and optimization.

According to Jen-Hsun Huang, the world’s current 10 million factories, nearly 200,000 warehouses and 40 million miles of highways constitute the “computing” network of our physical world. However, the network of production facilities and distribution centers within this vast network still requires manual design, operation, and optimization.

In warehousing and distribution, operators face highly complex decision optimization problems – including variables and interdependencies between human workers, robots, agent systems and equipment that remain complex, current digital twins It is still difficult to achieve analysis and processing in such a complex environment.

In response to this demand in the physical world, NVIDIA has released an Omniverse Blueprint framework called "Mega" for large-scale development, testing and optimization of physical AI and robot queues in a digital twin environment, and then Deployed to real-life facilities.

Currently, advanced warehouses and factories already use more than hundreds of autonomous mobile robots, robotic arms, and humanoid robots to work collaboratively with humans. The implementation of increasingly complex sensor and robotic autonomous systems requires coordinated training in simulations to optimize operations, ensure safety, and avoid disruptions.

For the robot's autonomous system, simulate various possibilities of path planning to find the most efficient version｜Image source: NVIDIA

In response to this demand, Mega provides a Reference architecture based on Nvidia accelerated computing, AI, Nvidia Isaac and Nvidia Omniverse technologies for developing and testing digital twins. These digital twins can be used to test AI brains that drive robots, video analysis AI agents, devices, and more to handle massive complexity and scale.

Through this digital twin, companies can continuously update the robot brains in their facilities to achieve intelligent path planning and task allocation, thereby improving operational efficiency.

In a simulated environment, these robots can complete tasks through perception and reasoning, plan their next actions and execute them in the digital twin. This cycle continues, with Mega accurately tracking the status and location of all assets in the digital twin.

Unlike the IT industry, the physical industrial market is still waiting for its own software-defined moment. "In the future, every factory will have a digital twin," Huang said.

03. Project DIGITS: The supercomputing center moves home

As the large-scale model "alchemy" has rapidly become a national trend in recent years, even ordinary users have emerged with many people interested in high computing power. Demand usage scenarios - but obviously not everyone has the conditions to build a GPU cluster to train their own large models.

In response to this gap between demand and reality, NVIDIA has come up with a unique solution this time: Project Digits, a computing unit equivalent to the size of a Mac mini.

Project Digits that can be placed on the desktop｜Image source: NVIDIA

The core of Project Digits is the new generation GB10 Grace Blackwell chip. This desktop-level system can handle up to 200 billion parameters AI model, while using a standard household power socket to achieve power supply - this is usually unimaginable in the past when the same computing power required larger and more power-consuming hardware.

In the CPU part, Project Digits uses NVIDIA's own Grace CPU, using a customized 20-core ARM architecture. Each system is equipped with 128GB of unified memory (ordinary laptops may only have 16GB or 32GB of RAM) and up to 4TB of NVMe storage.

NVIDIA also offers a variety of AI software to Project Digits users, including development kits, orchestration tools, and pre-trained models through the NVIDIA NGC catalog. The operating system runs the Linux-based NVIDIA DGX OS and supports mainstream frameworks such as PyTorch, Python, and Jupyter.

Developers can leverage the NVIDIA NeMo framework to fine-tune models and accelerate data science workflows using the NVIDIA RAPIDS library.

After developing and testing AI models locally, users can deploy these models to cloud services or data center infrastructure, maintaining seamless integration with the Grace Blackwell architecture and NVIDIA AI enterprise software platform.

Not only is the appearance similar to the Mac mini, Project Digits is even very close to the Mac mini in usage: for applications that require more powerful computing power, two Project Digits systems can be connected together to obtain enough processing power to process up to Performance of a 405 billion parameter model.

If you don’t have an accurate idea of ​​the volume of this model, here is a reference: Meta’s current top model Llama 3.1 has 405 billion parameters

In other words, with two suchWith "Mac mini", you can build a "supercomputing center" at home that can run the most advanced large models.

Project Digits will be officially launched this May, starting at $3,000. Although this is still not a cheap price, it is still a solution provided by NVIDIA for individual developers to have access to super computing power. It also brings greater flexibility and flexibility to the development and deployment of AI applications. efficiency.

Although consumer electronics is the mainstream at CES, Huang Renxun’s opening once again proved that in the still-developing wave of AI, the rapid evolution of infrastructure is the top priority.

Computing power, models, and application implementation, NVIDIA’s “three-piece AI infrastructure set” has become more and more mature. This is why Lao Huang did not say the classic saying “Buy more, "Save a lot more" - he no longer needs to say it. Nvidia's focus is no longer simply "selling cards", they have greater goals and ambitions. At the press conference that night, Huang Renxun had already "spoiled the company's plans in advance."