“When a company’s stock rises 260 percent in a year, its CEO can behave like a rock star. That’s how Jensen Huang appeared yesterday, in a black leather jacket, in front of thousands of attendees in a venue where major concerts are regularly held.
NVIDIA’s GPU Technology Conference (GTC) kicked off yesterday in San Jose, California, and lasts for two more days. The goal is to announce a range of new products and platforms aimed at solidifying NVIDIA’s dominance in the artificial intelligence market. The conference is attended by developers, scientists, and technologists from leading companies, universities, research institutions, and government agencies around the world. The who’s who of the artificial intelligence world are present, including OpenAI, xAI, Meta, Google, and Microsoft, as well as L’Oréal, Lowe’s, Jaguar Land Rover, Shell, Verizon, as they also use NVIDIA’s products. In his opening speech, Huang announced the next generation of NVIDIA chip architecture codenamed Blackwell, and related products, including the company’s latest AI chip called the B200. The GPU platform is named in honor of mathematician David Harold Blackwell and succeeds the Hopper architecture, which was launched two years ago. Huang said the new chips are much faster and larger than their predecessors. They will be available later this year. UBS analysts estimate that NVIDIA’s new chips could cost up to $50,000, roughly double the price of the previous generation.
During a detailed technical presentation, Huang illustrated the various applications of the new chips in creating and deploying artificial intelligence systems. He also introduced new AI-specific network switches and supercomputing systems that integrate these new chips.
He emphasized that the new chips could be used in training the latest and largest artificial intelligence models. If two thousand Blackwell GPUs are used over approximately 90 days, the power consumption will be only 4 megawatts. For comparison, the same task with the previous generation of chips would require the use of 8,000 GPUs and 15 megawatts of energy in the same time period. Blackwell GPUs contain 208 billion transistors and can enable AI models that scale to 10 trillion parameters.
For comparison, OpenAI’s GPT-3 consists of 175 billion parameters. Blackwell will be embedded in Nvidia’s GB200 Grace Blackwell Superchip, which connects two B200 Blackwell GPUs with a Grace CPU. Nvidia has announced plans to use Blackwell GPUs in Amazon Web Services, Dell Technologies, Google, Meta, Microsoft, OpenAI, and Tesla. Nvidia also announced the GB200 NVL72 liquid-cooled rack system, which contains 36 GB200 Grace Blackwell Superchips. The system is 30 times more powerful compared to its current H100 GPU.
The company also announced that it will have a server board, named HGX B200, which connects eight B200 GPUs. The Blackwell GPU architecture includes six transformative technologies for accelerated computing. Among the significant achievements are new Tensor Cores and TensorRT-LLM Compilers, which together reduce LLM operating costs and energy consumption by up to 25 times.
Analysts estimate that NVIDIA’s new chips could cost up to $50,000, roughly double the price of the previous generation.
In addition to introducing new products, Huang outlined his vision for future technologies that are still in development, including drug discovery, genomics, and humanoid robots. On stage next to him, at the end of his speech, stood nine humanoid robots of different appearances.
Nvidia is at the top, the company is worth more than $2 trillion, just behind Apple and Microsoft, Huang has become rich, with around $70 billion, but the competition is not sleeping. Among the newcomers challenging Nvidia are chip startups with names like Cerebras, Groq, and SambaNova Systems, as well as better-known competitors like Intel and Advanced Micro Devices. We’ve written about how current customers are craving their own chips now. Still, when you dominate 80 percent of the market, you deserve to behave like a rock star.
