Nvidia 2024 AI Event: Everything Revealed in 16 Minutes

Nvidia's 2024 AI Event unveiled the new Blackwell platform, a revolutionary GPU with 28 billion transistors and a unique design that eliminates memory locality issues. The platform is compatible with Hopper systems, allowing for a seamless transition to Blackwell. It also introduced the MV link switch chip with 50 billion transistors, enabling full-speed communication between GPUs. Nvidia highlighted partnerships with major companies like AWS, Google, and Microsoft, focusing on AI acceleration and secure AI development. The event also discussed the AI Foundry, a service that provides pre-trained models and tools for AI development, and the Omniverse, a digital twin platform for AI agent training and evaluation. Finally, the new Jetson Thor robotics chips were introduced, designed to power the next generation of AI-powered robotics.

Takeaways

• 🚀 **Blackwell Platform Introduction**: Nvidia introduces Blackwell, a new computing platform that significantly changes the architecture of GPUs, with 28 billion transistors and a design that allows two dies to function as a single chip with no memory locality issues.
• 🔗 **Memory Coherence**: Blackwell features 10 terabytes per second of data transfer between its two sides, creating a unified experience for the dies, which is a first for its kind in computing.
• 💻 **Compatibility with Hopper**: Blackwell is designed to be form-fit and function-compatible with Hopper, allowing for a seamless upgrade path for existing systems.
• 🔍 **Content Token Generation**: A key component of the new processor is the content token generation in the fp4 format, highlighting the importance of generative AI in the era of advanced computing.
• 🏗️ **System Integration**: The Blackwell chip can be integrated into two types of systems, one being a Hopper version for current HGX configurations, and the other a more advanced prototype.
• 🔗 **MV Link Switch**: Nvidia has developed an impressive chip called the MV Link Switch with 50 billion transistors, capable of connecting every GPU to each other at full speed simultaneously.
• 🌐 **Partnerships for AI Acceleration**: Nvidia is collaborating with major companies like AWS, Google, and Microsoft to accelerate AI services, databases, and other critical enterprise systems.
• 📦 **Nvidia AI Foundry**: The AI Foundry initiative aims to provide an end-to-end AI solution, including the NIMS (Nvidia Inference Microservice), Nemo, and DGX Cloud, to help companies build and scale their AI capabilities.
• 🤖 **Omniverse for Robotics**: Nvidia's Omniverse platform is central to creating digital twins for training AI agents and robots, streamlining workflows across different departments and tools.
• 🧠 **General Robotics Learning**: Project Groot is a foundation model for humanoid robot learning, capable of taking multimodal instructions and past interactions to produce actions for robots to execute.
• 📈 **Jetson Thor Chips**: The new Jetson Thor robotics chips are designed to power the next generation of AI-powered robotics, as demonstrated by the Disney Research robots that learned to walk using Isaac Sim.

Q & A

What is the name of the new platform announced by Nvidia at the 2024 AI Event?

The new platform announced by Nvidia is called Blackwell.

How many transistors does the Hopper chip have?

The Hopper chip has 28 billion transistors.

What is unique about the Blackwell chip's architecture?

The Blackwell chip has a unique architecture where two dies are abutted together in such a way that they function as one chip with no memory locality issues and no cache issues.

What is the data transfer rate between the two sides of the Blackwell Chip?

The data transfer rate between the two sides of the Blackwell Chip is 10 terabytes per second.

What is the significance of the MV link switch chip?

The MV link switch chip is significant because it allows every single GPU to communicate with every other GPU at full speed simultaneously, facilitating high-speed, efficient data processing.

How many billion transistors does the MV link switch chip have?

The MV link switch chip has 50 billion transistors.

What is the name of the format for content token generation in the generative AI era as mentioned in the event?

The format for content token generation in the generative AI era is called FP4.

Which companies are mentioned as partners gearing up for Blackwell?

Partners gearing up for Blackwell include AWS, Google, Oracle, Microsoft, and Dell.

What is the name of the pre-trained model that is optimized to run across Nvidia's install base?

The pre-trained model is called NM, also known as Nvidia Inference Microservice (NIMS) within the company.

What is the purpose of Nvidia AI Foundry?

Nvidia AI Foundry is designed to work with companies to build, optimize, and package AI solutions, similar to how TSMC manufactures chips for Nvidia's ideas.

What is the name of the simulation engine that represents the world digitally for robots?

The simulation engine that represents the world digitally for robots is called Omniverse.

What is the name of the general-purpose foundation model for humanoid robot learning developed by Nvidia?

The general-purpose foundation model for humanoid robot learning is called Project Groot.

Outlines

00:00

🚀 Introducing Blackwell: The Next-Gen GPU Platform

The first paragraph introduces the Blackwell platform, emphasizing its revolutionary design and capabilities. Blackwell is a significant departure from traditional GPUs, with 28 billion transistors and a unique architecture that allows two dies to function as a single chip with no memory locality or cache issues. It supports 10 terabytes per second of data transfer. The platform is designed to be form-fit and function-compatible with existing Hopper systems, facilitating an efficient transition. The paragraph also highlights the creation of a processor tailored for the generative AI era, focusing on content token generation in a new format called FP4. Additionally, the development of another chip, MVY Link Switch, with 50 billion transistors and four MV links capable of 1.8 terabytes per second data transfer is mentioned. The potential of connecting every GPU to every other GPU at full speed is also discussed, along with partnerships with companies like AWS, Google, and Microsoft to integrate and accelerate AI services.

05:00

🤖 Nvidia's AI and Robotics Initiatives

The second paragraph delves into Nvidia's AI and robotics initiatives. It discusses the collaboration with various companies to build AI systems, such as the partnership with AWS to integrate Nvidia Health, and the use of Nvidia Omniverse and Isaac Sim by Amazon Robotics. Google's preparation for Blackwell and its existing fleet of Nvidia GPUs are highlighted, along with the announcement of Google's Gemma model. Oracle's and Microsoft's readiness for Blackwell and their collaborations with Nvidia are also mentioned. The paragraph further covers the Nvidia inference microservice (NIMS) and the AI Foundry concept, which includes NIMS, Nemo microservice, and DGX Cloud. It outlines the AI Foundry's work with companies like SAP, Cohesity, Snowflake, and NetApp to build AI-driven solutions. The importance of Dell in building AI factories for enterprises is acknowledged, and the need for an end-to-end system for AI at scale is emphasized.

10:00

🌐 The Power of Omniverse and AI-Driven Robotics

The third paragraph focuses on the digital representation of the world through Nvidia's Omniverse platform and the OVX computer hosted in the Azure Cloud. It discusses the use of digital twins in industrial spaces to train AI agents for navigating complex environments. The announcement of Omniverse Cloud's integration with the Vision Pro is highlighted, enabling seamless connection to Omniverse portals and streamlined workflows across various design tools. The development of Nvidia Project Groot, a general-purpose foundation model for humanoid robot learning, is introduced. Isaac Lab, a robot learning application, and the new compute orchestration service, OSMO, for training and simulation are also mentioned. The paragraph concludes with the introduction of the Jetson Thor robotics chip, designed to power AI-driven robotics, and the showcase of Disney's BDX robots powered by Jetson, demonstrating the practical application of these technologies.

15:02

🔋 Blackwell: The Future of GPU Technology

The final paragraph summarizes the key points about the Blackwell platform. It reiterates the innovative aspects of Blackwell, including its high-performance processors, MV link switches, and networking systems. The paragraph emphasizes the system design as a marvel and reflects on the presenter's vision of what a GPU should represent in the modern era, encapsulating the essence of the Blackwell platform.

Keywords

Blackwell

Blackwell is referred to as a platform in the script, not a chip. It signifies a new era of GPU architecture that is form-fit and function-compatible with Hopper, another Nvidia technology. The term 'Blackwell' is central to the video's narrative as it represents a significant advancement in GPU technology, with the ability to seamlessly integrate two dies (components of a chip) to function as a single unit, showcasing a leap in data transfer and processing capabilities.

Hopper

Hopper is a term used to describe a previous generation of Nvidia's GPU architecture that is succeeded by Blackwell. It is mentioned as having '28 billion transistors,' indicating its high level of complexity and capability. Hopper is significant as it laid the groundwork for the innovations found in Blackwell, with its design and infrastructure being compatible and upgradable to the newer platform.

Transistors

Transistors are the fundamental building blocks of electronic devices, including GPUs. The script mentions '28 billion transistors' in Hopper and '50 billion transistors' in the MV link switch, emphasizing the vast scale and complexity of these computing components. The number of transistors often correlates with the processing power and efficiency of a chip, making it a key concept in understanding the advancements discussed in the video.

MV link switch

The MV link switch is a new chip introduced in the script, described as having '50 billion transistors' and the capability of handling '1.8 terabytes per second' of data transfer. It represents a significant development in interconnectivity for GPUs, allowing for every GPU to communicate with every other GPU at full speed simultaneously, which is crucial for the generative AI era discussed in the video.

Generative AI

Generative AI refers to the technology that can create new content based on existing data. In the context of the video, it is an era that Nvidia is preparing for with the creation of processors like Blackwell. Generative AI is highlighted as an important application for the new hardware, as it requires significant computational power and efficient data handling, which the new Nvidia technologies are designed to provide.

FP4

FP4 is mentioned as a format or technology related to content token generation, which is a part of the advancements in computing discussed in the video. While not elaborated upon in detail, it implies a specific approach or standard for handling data that is optimized for the new generation of Nvidia hardware, contributing to the speed and efficiency of generative AI processes.

DGX

DGX refers to Nvidia's Data Center systems designed for AI workloads. The script mentions 'one DGX' looking like a certain configuration and talks about an 'exaFLOPs AI system in one single rack,' indicating a level of performance that is orders of magnitude beyond current capabilities. DGX systems are a key part of the infrastructure that can harness the power of Nvidia's advanced GPUs for AI applications.

Nvidia AI Foundry

Nvidia AI Foundry is a service announced in the video that is likened to a foundry for manufacturing chips, but for AI. It includes three pillars: NIMS (Nvidia inference microservice), Nemo (microservice for data preparation), and DGX Cloud. The AI Foundry is designed to help companies build and optimize AI applications by providing pre-trained models, software optimization, and cloud services, which are all integral to the development of AI in various industries.

Omniverse

Omniverse is Nvidia's virtual world simulation platform, which is crucial for the development and training of AI agents, particularly in robotics. The script discusses the integration of Omniverse with Azure Cloud and its use in creating digital twins for complex industrial environments. It represents a significant step towards simulating real-world scenarios for AI training and evaluation.

Jetson Thor

Jetson Thor is a new robotics chip announced in the video, designed to power the next generation of AI-powered robots. It is mentioned in the context of Project Groot, a general-purpose foundation model for humanoid robot learning. The chip is a building block for enabling robots to learn from human demonstrations and to execute tasks, signifying a leap in the capabilities of AI in robotics.

Digital Twin

A digital twin is a virtual representation of a real-world entity or system. In the context of the video, digital twins are used to simulate complex industrial environments for AI agents and robots. This allows for the training and evaluation of AI systems in a controlled and scalable manner before deployment in the physical world, which is a key aspect of the future of heavy industries as discussed in the video.

Highlights

Nvidia introduces Blackwell, a new platform with a focus on generative AI era.

Blackwell features 28 billion transistors and a unique design that connects two dies as one chip.

10 terabytes per second of data transfer between the two sides of the Blackwell Chip.

Compatibility with current Hopper systems allows for a seamless upgrade path.

Blackwell's architecture eliminates memory locality and cache issues.

Introduction of the MV link switch with 50 billion transistors and 1.8 terabytes per second data transfer.

MV link switch enables full-speed communication between every GPU simultaneously.

Nvidia's partnership with major companies like AWS, Google, and Microsoft to integrate and accelerate AI services.

Nvidia AI Foundry aims to be an AI manufacturing platform, similar to TSMC for chips.

Nvidia inference microservice (NIMS) and Nemo microservice for data preparation and AI fine-tuning.

Collaboration with SAP, Cohesity, Snowflake, and NetApp to build AI-driven solutions.

Omniverse Cloud and its integration with design and simulation tools for a seamless workflow.

Project Groot, a general-purpose foundation model for humanoid robot learning.

Isaac lab and Osmos for training and scaling AI models for robotics.

Jetson Thor, a new robotics chip designed for AI-powered robotics.

Disney's BDX robots showcased, powered by Jetson and trained in Isaac Sim.

Nvidia's commitment to advancing computing at an incredible rate to meet AI demands.