NVIDIA Blackwell: B200 and GB200 NVL72, the rack-scale AI era
In one sentence At GTC 2024 NVIDIA announces Blackwell (208 billion transistors, two dies acting as one GPU) and the GB200 NVL72 rack with 72 GPUs and 36 Grace CPUs. Claimed up to 30x LLM inference versus the same number of H100s.
Every year NVIDIA unveils its news at the GTC conference. On March 18, 2024 Jensen Huang announces Blackwell, the new family of AI chips.
The headline number: the Blackwell GPU has 208 billion transistors and is made of two chips joined by a link so fast that the system sees them as a single GPU. But the real product is the rack: the GB200 NVL72 puts 72 GPUs and 36 Grace CPUs in one liquid-cooled cabinet, linked via NVLink as if they were one big computer.
NVIDIA claims up to 30 times the LLM inference performance of the same number of H100s, and up to 25 times lower cost and energy. Among expected customers it names AWS, Google, Meta, Microsoft, OpenAI, Oracle, Tesla and xAI; products arrive from partners starting late 2024.
Companies
NVIDIA
Tools
B200, GB200 NVL72, NVLink, NIM
Tags
Sources