NVIDIA Blackwell

The world’s most powerful chip architecture designed to run trillion-parameter generative AI models at 25x less cost and energy than its predecessor.

NVIDIA Blackwell marks a generational leap in computing, featuring 208 billion transistors and the second-generation Transformer Engine. The platform integrates two high-performance dies via a 10 TB/s chip-to-chip link, enabling massive throughput for LLM inference. By utilizing the GB200 Grace Blackwell Superchip, data centers can scale to 72-GPU liquid-cooled racks (NVL72) that act as a single massive GPU. This architecture delivers 20 petaflops of AI performance and reduces training times for frontier models like GPT-4 while maintaining energy efficiency through dedicated decompression engines and secure AI features.

https://www.nvidia.com/en-us/data-center/blackwell/