NVIDIA launches GB200, its most powerful AI chip: 208 billion transistors

At NVIDIA's GTC 2024 conference, NVIDIA CEO Jen-Hsun Huang announced the launch of a new generation of GPUs, Blackwell.

The first Blackwell chip is called GB200 and will be available later this year.

The Blackwell GPUs are described as bulky, utilizing TSMC's 4-nanometer (4NP) process to integrate two chips with a total of 208 billion transistors using a 10 TB/sec NVLink 5.0 connection.

The previous generation GPU, the "Hopper" H100, used a 4nm process and integrated 80 billion transistors.

Jen-Hsun Huang said, "Hopper is great, but we need a bigger GPU. blackwell is not a chip, it's the name of a platform."

NVIDIA said Blackwell-based processors, such as the GB200, provide a huge performance upgrade for AI companies, with AI performance of 20 quadrillion floating-point operations per second, compared to the H100's 40 quadrillion floating-point operations per second.

The system can deploy a model with 27 trillion parameters, while the GPT-4 uses about 1.76 trillion parameters to train the system.

The GB200 contains a composition of two B200 Blackwell GPUs and an Arm-based Grace CPU, and inference of large language models improves performance by a factor of 30 over the H100, and reduces cost and energy consumption to a quarter.