Highlights:

  • The GB200 Blackwell Superchip integrates two Blackwells and a Grace CPU on a single board, forming a powerful chip suitable for individual systems, clusters, or racks.
  • Project Ceiba, initially unveiled by Nvidia and AWS in 2023 with the goal of constructing one of the world’s fastest supercomputers, will leverage Blackwell technology hosted on AWS infrastructure.

At the outset of this week in San Jose, Nvidia Corp. inaugurated its annual developer conference, GTC, and disclosed a series of product announcements. Among these, Nvidia debuted the Blackwell platform, the highly anticipated product. While Blackwell might initially seem like just another graphics processing unit, it transcends that definition, earning it the designation of a “platform.”

Blackwell is intriguing due to its ability to blur the distinction between chips and systems. Essentially, a single chip comprises two interconnected chips. The GB200 Blackwell Superchip consists of two Blackwell units and a Grace central processing unit integrated onto a board, effectively forming a colossal chip. This can then be utilized to construct a singular system, a cluster, or even a rack. The remarkable aspect lies in the seamless transfer of information between the chips, irrespective of the configuration, without any loss in performance.

The company emphasizes its capacity to assist businesses in “developing and operating real-time generative AI on trillion-parameter large language models.” Blackwell is expected to achieve this at a significantly reduced cost and energy consumption compared to the previous platform. In certain respects, Blackwell represents a nostalgic return to Nvidia’s roots. During a Q and A session with analysts, Chief Executive Jensen Huang elaborated on this aspect.

“We are going back to where we started. Blackwell generates content, just like our GPUs. They were used to create graphics. This is different in that it creates content of all types for everyone. Blackwell was built for the generative AI era,” he said.

Furthermore, it was revealed that Amazon Web Services Inc. will provide Nvidia Grace Blackwell GPU-powered Amazon EC2 instances and Nvidia DGX Cloud.

Details of each announcement are as follows:

The Blackwell Platform

The Blackwell platform, as described by Nvidia, is built upon six groundbreaking technologies capable of scaling up to 10 trillion parameters, which include:

  • This chip boasts 208 billion transistors, establishing it as the most potent in the world. It utilizes a custom-built 4NP TSMC process with GPU dies constrained to a two-reticle limit. These chips are interconnected through a 10-terabits-per-second chip-to-chip link, culminating in a unified GPU.
  • Introduction of support for double the compute and model sizes through new four-bit floating-point AI inference capabilities.
  • Enhanced performance for multitrillion-parameter and mixture-of-experts AI models.
  • Implementation of a dedicated engine to ensure reliability, availability, and serviceability.
  • Integration of protection mechanisms for AI models and customer data without compromising performance.
  • Inclusion of a dedicated decompression engine supporting the latest formats for high-performing data analytics and data science.

Broadening Partnership with AWS

In a bid to democratize AI, Nvidia will collaborate closely with AWS to provide access to the Blackwell platform. By integrating Blackwell with Amazon’s EFA networking, Nitro virtualization, and EC2 UltraClusters, customers will gain the capability to scale up to thousands of GB200 Superchips, thus accelerating inference workloads for resource-intensive, multitrillion-parameter language models. Since it’s offered as an AWS service, customers have the flexibility to begin with a relatively modest environment instead of investing in a DGX server upfront.

Additionally, Project Ceiba, a collaboration unveiled by Nvidia and AWS in 2023 with the aim of constructing one of the world’s fastest supercomputers, will leverage Blackwell, hosted on AWS infrastructure. Ceiba is set to feature 20,736 B200 GPUs equipped with the new Nvidia GB200 NVL72 and fifth-generation NVLink, interconnected with 10,368 Nvidia Grace CPUs.

The company states that its R and D teams will utilize Ceiba to push the boundaries of AI in various domains such as large language models, graphics encompassing image, video, and 3D generation, simulation, digital biology, robotics, self-driving cars, Nvidia Earth-2 climate prediction, and beyond. This endeavor is aimed at driving Nvidia’s future generative AI innovation forward.

AWS and Nvidia are collaborating on the development of novel applications, which include the creation of Nvidia BioNeMo foundation models. These models are designed for generative chemistry, protein structure prediction, and gaining insights into the interactions between drug molecules and targets.Nvidia plans to offer these models through AWS HealthOmics, a platform designed to assist healthcare and life sciences organizations in storing, querying, and analyzing various types of omics data, including genomic and transcriptomic data.

The teams are working together to introduce generative AI microservices tailored for applications in drug discovery, medical technology, and digital health.