Overview

Instructor(s): James Chegwidden

Time: Tuesday, March 4, 2025 — 1:30PM-2:30PM CT

Location: online using Zoom

Prerequisite(s): Active ACCESS ID, general understanding of computer architecture

This short-course will discuss and examine the NVIDIA GH200 Grace Hopper Superchip. The Grace CPU architecture is designed to provide exceptional performance for AI, machine learning (ML), and scientific computing applications. Its key innovation lies in the integration of Grace processors with the Hopper GPU, forming a hybrid system optimized for massive parallelism and computational workloads.

In this short-course, we’ll explore the technical aspects of the Grace Chip, its advantages, challenges, and architectural design. The benchmarks conducted by HPRC researchers will be presented, highlighting the Grace Hopper system's performance across a variety of HPC workloads.

Course Materials

Presentation slides

• ACES: Intro to the Grace Hopper Superchip (Spring 2025) PDF

Learning Objectives and Agenda

In this class, participants will learn:

• The fundamental architecture and design principles of the NVIDIA GH200 Grace Hopper Superchip.
• How the integration of Grace CPUs with Hopper GPUs creates a powerful hybrid system optimized for parallel processing.
• Key performance metrics and real-world benchmarks of the Grace Hopper system.
• The applications and advantages of using the Grace Hopper Superchip in high-performance computing (HPC) environments.
• Potential challenges and considerations in deploying and utilizing the Grace Hopper system in production environments.

This course focuses, among others, on the following topics:

• The NVIDIA GH200 Architecture
• HPC Benchmarks and Performance
• Energy Efficiency and Scalability
• Challenges and Opportunities