AI & Tech – NVIDIA’s Blackwell Platform Launch
NVIDIA announced its newest platform, Blackwell, for the era of generative AI, at the GTC conference on March 18, 2024. Featuring groundbreaking technologies for the development and operation of massive AI models, Blackwell boasts as much as a 30x performance increase with up to a 25x reduction in energy consumption compared to its predecessor, the Hopper architecture.
Key Features and Innovations The Blackwell platform, named after mathematician David Blackwell, introduces the following new technologies:
- Superchip Design: Blackwell GPUs contain 208 billion transistors and are fabricated using a custom TSMC 4NP process. The GB200 Grace Blackwell Superchip connects two B200 GPUs to an NVIDIA Grace CPU using a high-speed NVLink interconnect.
- Second-Generation Transformer Engine: This engine adds support for new FP4 and MXFP6 data formats, improving the performance of large language models and Mixture-of-Experts models.
- Fifth-Generation NVLink: This high-speed interconnect provides 1.8 TB/s of bidirectional bandwidth per GPU and supports up to 576 GPUs communicating with each other.
- Dedicated Engines: The platform consists of a Decompression Engine that speeds up data analytics and a RAS Engine for reliability and diagnostics.
- Secure AI: State-of-the-art confidential computing capabilities protect AI models and data.