NVIDIA unveils Blackwell and the GB200 NVL72 rack at GTC

NVIDIA/Blackwell

NVIDIA announced the Blackwell GPU platform, claiming up to 25 times lower cost and energy for running trillion-parameter language models than Hopper. Its flagship GB200 NVL72 is a liquid-cooled rack linking 72 Blackwell GPUs and 36 Grace CPUs acting as one large accelerator with 1.4 exaflops AI performance.

• Blackwell platform delivers up to 25x lower cost and energy versus Hopper for trillion-parameter LLM inference, with 208 billion transistors per GPU.
• GB200 NVL72 comprises 72 Blackwell GPUs and 36 Grace CPUs with 1.8 TB/s bidirectional NVLink bandwidth and 30 TB total system memory.
• GB200 achieves 30x performance increase and 25x lower cost/energy versus H100s for LLM inference; supports up to 576 GPUs for complex models.