NVIDIA is gearing up for the "AI Summit"! The focus is on the B100 chip and Blackwell architecture.

Wallstreetcn
2024.02.23 12:29
portai
I'm PortAI, I can summarize articles.

After a 5-year hiatus, this year's GTC conference will be held offline for the first time. NVIDIA may unveil "heavy artillery" such as the Blackwell architecture and B100 AI GPU, further enhancing the NVIDIA ecosystem.

NVIDIA's dazzling performance continues to ignite market enthusiasm, acting as a mirror reflecting the industry's confidence in AI development. Next month, NVIDIA's annual "AI Summit" is about to kick off. What surprises will it bring?

From March 18th to 21st, NVIDIA's annual AI conference GTC (GPU Technology Conference) 2024 will be held at the San Jose Convention Center in the United States. On March 18th (Monday) at 1:00 PM (Beijing time, early Tuesday, March 19th at 4:00 AM), NVIDIA CEO Jensen Huang will deliver a keynote speech titled "Number One AI Summit for Developers."

GTC, as one of NVIDIA's most important annual release platforms, has become a recognized "AI compass." This year's GTC is the first offline event in 5 years. According to NVIDIA, Jensen Huang will unveil the latest breakthroughs in accelerated computing, generative AI, and robotics.

Huang mentioned that governments, industries, and organizations are all committed to harnessing the transformative power of generative AI, which has become a global focus.

NVIDIA stated that this GTC conference is aimed at AI developers to share the latest achievements in the AI field. Huang will host the event, and special guests including the eight authors of Transformer (who proposed the Transformer language model in "Attention Is All You Need") will push the conference atmosphere to a climax. (Note: "Attention Is All You Need" was published in 2017 and is considered one of the most important papers in the AI field. It has been cited over 80,000 times so far, introducing the Transformer language model based on attention mechanisms to replace recurrent and convolutional neural networks.)

In the pre-event video, NVIDIA highlighted the applications of generative AI, including the WPP/NVIDIA engine for digital advertising, the newly released RTX Chat, the industrial metaverse supported by SyncTwin, and OpenAI's code for creating Blender animations.

The market generally believes that the focus of this conference revolves around three aspects: 1. the next-generation Blackwell GPU architecture, 2. the new B100 chip, and 3. humanoid robots.

All Eyes on the Blackwell Architecture

Media analysis believes that this year's NVIDIA GTC returning to offline, Huang Renxun should "bring out some real deal," such as the highly anticipated Blackwell architecture and the B100 GPU based on Blackwell.

As NVIDIA has not disclosed the specific parameters and listing schedule of B100, the outside world can only piece together a preliminary understanding of B100 from limited official information and comparisons with H200. According to speculation, the performance of B100 is at least twice that of H200, which is a full four times that of H100.

There are also reports citing informed sources that B100 will adopt TSMC's 3nm process, with Samsung as the main memory supplier. Among a series of ambiguous messages, the most attention-grabbing is undoubtedly the Blackwell architecture.

NVIDIA currently holds over 90% of the market share of AI GPU chips. The H200 and H100 that help it dominate the AI chip industry are both based on the Hopper architecture. According to official statements, the biggest advantage of this architecture is accelerated computing, capable of handling trillion-parameter AI large model training and expanding to various data centers.

The "evolved" Blackwell will not only further enhance AI acceleration capabilities but also feature high-speed memory interfaces, improved ray tracing technology, and parallel processing capabilities. Media analysis points out that the performance of the Blackwell architecture will largely determine the computing power limit of B100 and NVIDIA's subsequent products.

Some analysts believe that the Blackwell GPU will also be NVIDIA's first HPC/AI accelerator designed with chiplet technology, directly competing with AMD's Instinct MI300.

Morgan Stanley believes that NVIDIA wants to defend its computing power advantage, firmly bind those core customers, and B100 is the most useful "trump card." It is expected that B100 will become a game-changer in artificial intelligence, even more powerful than the previous flagship AI chip H100.

Bank of America analyst Vivek Arya believes that the pricing of the upcoming B100 from NVIDIA will be at least 10% to 30% higher than the H100 system, solidifying NVIDIA's "king" status.

Some analysts also believe that the issue with B100 may be the speed at which the new components of B100 SXM and B100 PCIe, along with the accompanying DGX servers, can increase production. If there is a huge market demand, it is very likely to repeat the widespread delays seen in the initial shipments of H100. The NVIDIA ecosystem is further enhanced.

According to NVIDIA, more than 1,000 companies will participate in this year's GTC, including but not limited to Adobe, Amazon, Aon, Anthropic, Blackrock, Cohere, Databricks, Dell Technologies, Genetec, Mercedes-Benz, Micron, Netflix, Oracle, Pixar, Runway, Saudi Aramco, Scale AI, Service Now, Siemens, Snowflake, Supermicro, and Disney Animation Studios.

The conference will host over 900 sessions, with more than 300 exhibitors showcasing how enterprises deploy the NVIDIA platform to achieve breakthroughs in various industries such as aerospace, agriculture, automotive and transportation, cloud services, financial services, healthcare and life sciences, manufacturing, retail, and telecommunications.

The application of generative AI in film and television will be a key highlight, with the exhibition featuring a giant multi-sensory interactive installation by world-renowned AI artist Refik Anadol, as well as a demonstration of the generative AI tool Cuebric designed by Seyhan Lee for filmmakers.

Furthermore, the development of robotics in 2024 continues to reach new heights. NVIDIA emphasizes in the preheating document that it will release achievements in the field of robotics. Many globally leading robotics companies will showcase next-generation robot products at the event. Companies such as Agility Robotics, Boston Dynamics, Disney, and Google DeepMind will exhibit 25 robots on-site, including humanoid robots and industrial robotic arms.

Guotai Junan Securities released a report stating that the development of humanoid robots will also be a focus of this conference. Huang Renxun stated at the 2023 ITF World Congress: the next wave of artificial intelligence will be a new type called Embodied AI.

Previously, NVIDIA has launched the NVIDIA Isaac platform, with hundreds of customers using it to develop high-performance robot solutions in agriculture, warehouse automation, last-mile delivery, service robots, and other fields. The advancement of AI technology will continue to be the core driving force behind the development of the humanoid robot industry.