NVIDIA’s INSANE New Tech - Mini Supercomputer, Agents, RTX 5090!! (CES 2025)

CES 2025 Keynote Highlights by Jensen Huang

Introduction to the Event

• Jensen Huang, CEO of Nvidia, kicked off CES 2025 with multiple announcements focusing on futuristic technologies powered by Nvidia.

• The speaker mentions their attendance at the event and notes that while Nvidia covered part of their trip expenses, they were not compensated for creating this video.

Evolution of AI Technologies

• A graph presented highlights the evolution from AlexNet in 2012 to current advancements in perception AI and generative AI.

• Emphasis is placed on "agentic AI" as a key focus for 2025, which includes applications like coding assistants and patient care.

• Agentic AI combines large language models with real-world functionalities such as memory tools and web browsing capabilities.

Transition to Aid-driven Computing

• Discussion on the shift from classical computing to aid-driven computing; predictions suggest a more integrated role for AI than traditional programming methods.

• The concept of predictive layers in computing is introduced, where interactions are anticipated by AI based on database inputs.

Innovations in Graphics Rendering

• A demo showcased how most scenes are rendered using artificial intelligence, significantly changing video game graphics rendering techniques.

• It was revealed that 90% of frames in a demo were generated by neural networks, showcasing a major advancement over traditional ray tracing methods.

Announcements: RTX Blackwell and GPU Pricing

• The keynote introduced RTX Blackwell and the new series of GPUs, highlighting an impressive price point of $549 for performance equivalent to top-tier previous models.

Scaling Laws in Artificial Intelligence

• Jensen discusses scaling laws that indicate larger datasets lead to more effective models; this has been observed across generations in AI development.

• There’s an ongoing race among companies to acquire vast amounts of data from various sources including public internet data and paywalled content.

AI Data Production and Scaling Laws

The Nature of Data Creation

• Humanity has produced an impressive amount of data, with AI increasingly generating derivative data that saturates the internet.

• Models trained on human-generated data create additional data, which is then used to train subsequent models, raising questions about the quality of this new data.

Emergence of New Scaling Laws

• Two significant scaling laws have emerged: post-training scaling law and test time scaling.

Post-Training Scaling Law

• This law involves techniques like reinforcement learning from human feedback (RHF), where AI generates answers based on queries and receives feedback for improvement.

• Fine-tuning through RHF can significantly enhance model quality, comparable to pre-training methods.

Test Time Scaling Law

• Test time scaling focuses on resource allocation during AI usage rather than just improving parameters. It allows models to optimize computation for producing answers effectively.

Advancements in AI Models

• Advanced models like Gemini 2.0 demonstrate capabilities in complex tasks such as PhD-level mathematics and coding challenges, showcasing the effectiveness of these scaling laws.

• Reasoning processes within AI allow for multi-step problem-solving rather than simple direct inference, enhancing overall performance.

Technological Infrastructure and Future Prospects

GPU Technology Insights

• The Grace Blackwell GPU represents cutting-edge technology utilized by cloud providers, highlighting advancements in computational power essential for modern AI applications.

Historical Context of Technological Evolution

• A reference to "Chip Wars" illustrates the historical struggle in semiconductor development and how far technology has come since the invention of transistors.

The Rise of Agentic AI

Understanding AI Interactions and Future Developments

The Role of AI in Customer Interaction

• AI systems are evolving to interact with customers by retrieving information from various sources, including storage systems, the internet, and documents like PDFs.

• Future AI responses will involve multiple models working simultaneously rather than a simple question-and-answer format, enhancing the depth of interaction.

Computational Demands and Market Predictions

• The demand for computational resources for inference is expected to surge significantly, indicating a growing market for Nvidia chips as more agents enter the workforce.

• A model orchestration layer will be essential to manage interactions among different models effectively; investments in companies like Crew AI highlight this need.

Nvidia's New Model Series

• Nvidia announced the Neaton series of models based on Llama, specifically tuned for enterprise applications. These include Nano (local machine use), Super, and Ultra (high accuracy but higher latency).

Introduction of Nvidia Cosmos

• Nvidia is developing a world foundation model called "Nvidia Cosmos," designed to understand physical environments through simulations.

• This platform aims to advance physical AI by utilizing auto-regressive and diffusion-based models alongside an accelerated data pipeline.

Generating Synthetic Data for Training

• Developers can create virtual world states using text, image, or video prompts within Cosmos, generating photorealistic synthetic data crucial for training physical AI robots.

Nvidia's Innovations in AI and Robotics

The Power of Nvidia GPUs and Multiverse Generation

• Nvidia GPUs are central to advancements in robotics, enabling the generation of every possible outcome through a model system known as multiverse generation.

• The concept of digital twins for factories is introduced, where each factory will have a virtual counterpart that can simulate various future scenarios to optimize performance based on key performance indicators (KPIs).

Advancements in Autonomous Driving and Robotics

• Nvidia collaborates with numerous car manufacturers and autonomous vehicle companies, applying Omniverse technology not only to driving but also to robotics.

• A significant breakthrough in humanoid robots is anticipated, likened to the "ChatGPT moment," driven by new waves of artificial intelligence.

The Future Landscape of General Robotics

• Current developments showcase real robots from various companies, indicating rapid advancements in general robotics due to enabling technologies.

• Three types of adaptable robots are highlighted: agentic AI for information work, self-driving cars benefiting from existing infrastructure, and humanoid robots capable of operating within human environments.

The Emergence of a New Industry

• Jensen Huang predicts that the robotics industry will become the largest technology sector ever created if breakthroughs continue in these three areas.

Introduction of DJX1 Supercomputer

• Nvidia unveils the DJX1 supercomputer designed for consumer use, aimed at making AI accessible for researchers and startups without needing extensive infrastructure.

• The DJX1 was first delivered to OpenAI in 2016, marking a pivotal moment in revolutionizing artificial intelligence development.

AI Integration Across Industries

• Artificial intelligence has permeated various sectors beyond research labs; it is now essential for software engineers and creative professionals alike.

Project Digits: A New Era for Home Computing

• Nvidia announces Project Digits—an AI supercomputer that runs on their entire software stack while being compact enough for home use.

• This device emphasizes local computing power with privacy features since it operates entirely on open-source software locally.

Technical Specifications and Design Insights

NVIDIA's New CPU and AI Capabilities

Overview of the New CPU

• NVIDIA has collaborated with Mediatek to develop a new CPU, which is now in full production. This CPU is designed to connect seamlessly with the Blackwell GPU.

Specifications and Features

• The new system-on-chip (SoC) boasts impressive specifications: 1 petaFLOP AI compute, 20 ARM cores, and 128 GB of low-powered DDR5x memory. It also includes a 4 TB SSD for storage, along with Wi-Fi, Bluetooth, and USB connectivity.

Performance Comparison

• The performance of this new computer is compared favorably against Apple's Mac Mini, suggesting it may even outperform it in certain aspects.

Versatility and Use Cases

• This device functions as a cloud computing platform that can also serve as a Linux workstation. It offers flexibility for various user needs.

Advanced Computing Capabilities

• The system supports advanced features like GPU direct connections out-of-the-box. Notably, it can run models with up to 200 billion parameters effortlessly.

Pricing and Availability