News｜The most powerful AI chip GB200 from Nvidia

At 4 am Beijing time today, Renxun Huang the founder of Nvidia, boarded the podium at the GTC2024 conference, wearing his signature leather jacket, and delivered a keynote speech entitled "Witnessing the transformative Moment of AI." He introduced the audience to an eye-catching new product - the most powerful AI chip in history, known as the GB200. Huang has experienced 30 years of ups and downs with Nvidia, and since Nvidia was officially founded in 1993, he has always focused on the GPU field, and now has become the dominant player in the global GPU market, and the market value has soared to $2.2 trillion, and the amazing rise speed is shocking.

This new generation GPU chip "Blackwell" uses TSMC's 4nm process, integrates two independently manufactured bare crystals (dies) with a total of 208 billion transistors, and connects the chips together through NVLink 5.0 technology. According to the introduction, each Blackwell Die has 25% more floating-point computing power than the previous generation Hopper Die, and there are two Blackwell chips in each package, which improves the total performance by 2.5 times.

The GB200 chip consists of two B200 Blackwell Gpus and a Grace CPU based on the Arm architecture. It is said to be capable of 20 petaflops of floating-point operations per second, while the H100 chip has only 4 petaflops. At the same time, the GB200 costs and energy consumption are reduced by 25 times.

For comparison, Nvidia said that to train a model with 1.8 trillion parameters would have required 8,000 Hopper Gpus running for 90 days, consuming 15 megawatts of power. Now, only 2,000 Blackwell Gpus are needed to do the job, consuming just 4 megawatts of power.

Nvidia is also working with cloud service giants such as Amazon, Google, Microsoft and Oracle to sell access to GB200 chips through cloud services. According to reports, Amazon's cloud service AWS will build a server cluster with 20,000 GB200 chips.

During his presentation, Huang stressed that "while Hopper is already great, we need a more powerful GPU." Blackwell is not just a chip, it's a platform." He believes that artificial intelligence is the driving force for fundamental changes in the economy, and the Blackwell chip will be the engine driving this new industrial revolution.

Huang also revealed that Nvidia is working with the world's most dynamic companies to bring the promise of artificial intelligence to various industries.

Not only on the hardware side, Nvidia is also making big moves in the software field. Huang said they will add a product called NIM to Nvidia's enterprise software subscriptions. NIM makes it easier to reason with older Nvidia Gpus and allows companies to continue using the 100 million Nvidia Gpus they already have. Nvidia's strategy is to require customers who buy Nvidia servers to sign up for Nvidia Enterprise and charge $4,500 per GPU per year.

It's also worth mentioning that NVIDIA Omniverse Cloud will be able to connect with Apple's mixed reality headset, the Vision Pro.

At the end of his talk, Huang left a cliffhanger: "We now have the technology we need to create humanoid robots." Currently working for 1X Technologies, Agility Robotics, Apptronik, Boston Dynamics, Figure AI, Fourier Intelligence, Sanctuary AI, Unitree Humanoid Robotics companies such as Robotics and XPENG Robotics build AI platforms. This could mean that Nvidia humanoid robots are coming soon.