NVIDIA’s Blackwell: The AI Thinking Machine Driving the Next Industrial Revolution

Blackwell: The AI Thinking Machine

NVIDIA’s CEO Jensen Huang unveils the Blackwell architecture, a significant leap in AI computing designed for reasoning models and agentic AI. This new generation of supercomputers, including the GB200 and GB300 systems, represents a fundamental shift in how computers are built and utilized, enabling unprecedented performance and scalability for AI factories and various industries like automotive and robotics.

Points clés

Moore’s Law provides about two times more performance every 3 to 5 years, while NVIDIA aims for 30-40 times more performance in one generation.
The new GB200 system weighs two to two and a half tons, has 1.2 to 2 million parts, costs about $3 million, consumes 120 kilowatts, is manufactured in 150 factories, and involves 200 technology partners.
NVIDIA invested approximately $40 billion in R&D to create the GB200 and GB300 systems.
The Blackwell architecture is designed as a “thinking machine” that reasons, plans, and engages in internal dialogue, functioning as one giant virtual GPU.
The MVLink system is a revolutionary memory semantics interconnect and compute fabric that directly connects CPUs and GPUs, offering a bandwidth of 130 terabytes per second.
The Blackwell architecture results in a significant performance leap over the Hopper generation, enabling higher throughput for AI factories and supporting more users simultaneously.
NVIDIA is mass-producing these supercomputers at a scale of a thousand systems per week.
The Grace Blackwell DGX Spark and DGX Station systems bring the Grace Blackwell architecture to desktop and desk-side form factors for developers.
NVIDIA’s DGX-1, the first AI supercomputer built in 2016, cost billions to develop and initially had no customers until OpenAI, a non-profit startup, expressed interest.
NVIDIA’s RTX Pro server is designed to run virtually all software, including video games like Crisis, and is powered by Blackwell RTX Pro 6000 GPUs and Super Nix switches.
NVIDIA’s NeMo platform enhances open-source AI models like Mistral, Llama, and DeepSeek through post-training, reinforcement learning, and extended context.
NVIDIA has partnered with Perplexity to integrate regional models into their reasoning search engine.
Agentic AI, a significant advancement from one-shot AI, integrates capabilities like access to real-time data, reasoning, and tool usage.
NVIDIA’s DGX Lepton system allows users to deploy models across various clouds (Lambda Cloud, AWS, GCP) using a single architecture and deployment method.
NVIDIA is collaborating with companies like BMW and Toyota to build digital twins of factories and warehouses in Omniverse for simulation and optimization.
Omniverse is built to create photorealistic and physically accurate digital twins where robots can learn and train.
NVIDIA provides the full stack for AI-driven robotics, including the computer (Thor), operating system, and AI models.
NVIDIA’s AV team has won the end-to-end self-driving car challenge at CVPR for two consecutive years.
NVIDIA’s Halo system for autonomous driving emphasizes safety from chip architecture to testing methodologies.
Humanoid robots are becoming viable due to advancements in AI learning from digital twins in Omniverse, with partnerships including Disney Research and DeepMind.
The number of people using inference has increased 100 times in a couple of years, from 8 million to 800 million.
The next wave of AI, characterized by agentic AI and robotics, will lead to an exponential increase in inference workloads and token generation.
Europe is significantly increasing its investment in AI infrastructure.

À retenir

So, apparently, the future of computing involves two-ton, multi-million dollar “thinking machines” that talk to themselves and generate “tokens” which, if I understand correctly, are the new food source for little AI robots. And if you’re a developer, and you need a GPU, the answer is always “yes.” Just don’t ask for a glass of whiskey from the robot without arms. It’s an industrial revolution, folks, powered by copper coax and a whole lot of R&D cash. Better start stockpiling those tokens!

Sources