Nvidia Blackwell GPU Powers New AI Supercomputers

Nvidia announced its next-generation Blackwell GPU architecture on March 18, 2024, engineered to significantly boost performance for artificial intelligence and data analytics tasks. The new architecture is built on a custom 4-NP process and features a 192 billion transistor design, promising up to 5x faster training and 30x faster inference speeds compared to its predecessor, Ampere. Blackwell integrates 192GB of high-bandwidth memory (HBM3e) per GPU, offering 1.7x more memory capacity and 1.4x greater bandwidth.

The Blackwell platform introduces several key innovations, including a second-generation Transformer Engine that dynamically optimizes floating-point precision to accelerate deep learning training and inference. It also features a new NVLink switch capable of delivering 1.8 terabytes per second of bidirectional bandwidth per GPU, enabling the connection of up to 576 Blackwell GPUs for massive AI models. The architecture supports FP4 precision, a new format that further enhances performance and reduces memory footprint for large language models.

Major cloud service providers, including Amazon Web Services (AWS), Google Cloud, and Microsoft Azure, have committed to deploying systems powered by the Blackwell GPU. AWS announced plans to offer Amazon EC2 instances featuring the GB200 Superchip, which combines two B200 GPUs with a Grace CPU, starting in late 2024. Microsoft Azure will also offer Blackwell-powered virtual machines, and Google Cloud will integrate the technology into its AI infrastructure. These collaborations aim to provide customers with enhanced capabilities for developing and deploying advanced AI applications.

Nvidia also introduced the DGX SuperPOD, a new AI supercomputer built with Blackwell GPUs, designed to deliver up to 115 petaflops of AI performance. The company highlighted the potential of Blackwell to accelerate scientific discovery, drug development, and autonomous systems. The Blackwell GPU is expected to be available in systems from Nvidia's partners starting in the second half of 2024, with the GB200 NVL72 server configuration shipping in the same timeframe.