Grace Hopper for
OpenFOAM 

The NVIDIA Grace Hopper provides developers with a modular system that optimally supports even demanding CFD simulations with OpenFOAM. For the application of complex simulation models in product development, it is necessary to perform the calculations massively in parallel on state-of-the-art computing architectures.

OpenFOAM-based applications on Grace Hopper architectures make optimum use of the performance potential of the servers. Compared to other systems such as the x86 system without Hopper, the NVIDIA Grace Hopper architecture only requires 15 percent instead of 35 percent runtime. Overall, only 85 percent of the runtime is generally used on the CPU side, which provides a good basis for faster and shorter design cycles. No matter how many product variants exist: With the right IT infrastructure for OpenFOAM, engineers increase the quality of simulations and significantly accelerate virtual product development.

Grace Hopper for
Large Language Models

With Grace Hopper, NVIDIA delivers a solid server that scales optimally to the requirements of demanding AI workloads. Large language models are based on billions to quadrillions of pieces of data and therefore require large computing power to enable the understanding and creation of language.

NVIDIA Grace Hopper is specially developed for training large AI models and enables high data throughput due to its architecture. Because of the HBM3 memory used, Grace Hopper achieves a memory bandwidth of almost 100 percent with a batch size of 1. While other systems such as x86 systems experience a drop in performance from a batch size of 4, the NVLink in Grace Hopper supports workloads by a factor of 4.5. This enables trained large language models with Grace Hopper to capture complex queries in the shortest possible time and process the huge data sets quickly.

Grace Hopper for
Rekommender Systems

Recommender systems (also known as recommendation services or filters) use powerful AI infrastructures to help end users and customers find the content, products and services they are most interested in. The underlying combination of sophisticated AI models and large data sets often requires extensive computing resources.

NVIDIA Grace Hopper is the right infrastructure for managing large models and massive data in recommender systems and takes interaction with user-generated content to the next level. NVIDIA Grace Hopper drives high-throughput recommender system pipelines by delivering outstanding CPU memory performance. This is made possible by NVIDIA's NVLink: the direct, high-performance communication bridge between a pair of GPUs delivers high-bandwidth access to HBM3 local memory and LPDDR5X memory. This accelerates the speed of work and helps the AI models used to deliver more accurate, relevant and faster results to end users.

Grace Hopper for
Graph Neural Networks

Graph Neural Network (GNN) already offers impressive performance and high interpretability, especially in the area where relationship and interaction of data play an important role. However, high computing power is required to enable the models to accurately analyze and predict large amounts of data in a short period of time.

NVIDIA Grace Hopper offers the right infrastructure for use cases in the field of GNN. Grace Hopper provides users with the structural basis to quickly and efficiently construct the numerical representation of a graph, which can later be used to train machine learning algorithms. For the processing of structured data and the training of GNN, the user has high bandwidth access to LPDDR5X and an NVLink C2C connection. Grace Hopper's large memory capacities solve graph-based machine learning tasks in the shortest possible time.

Would you like to learn more?

You will find further information in the NVIDIA-Whitepaper
'Performance and Productivity for Strong-Scaling HPC and Giant AI Workloads'