GPU Compute Servers (GRICAD)
Description
NVIDIA H100 GPU-accelerated calculation servers allow massive parallelism of simulations carried out with the main codes used in materials science. Additionally, with the increasing use of machine learning models in materials science, they are becoming an indispensable tool for running AI models and massive data analysis. These servers will be set up during 2024 for the development and testing of codes and workflows for the DIADEM community, with a view to their deployment in national centers.
GPU servers also have a much more attractive cost (financial or energy)/performance ratio than purely CPU calculation servers.
Location
LPSC, 53 avenue des Martyrs, 38026 GRENOBLE Cedex.
Contacts
Platform manager: Pierre-Antoine BOUTTIER
Associated targeted project coordinator: François WILLAIME
Technical characteristics
- 2 Lenovo calculation servers, integrated into a supercomputer (>10,000 cores), hosted in a full Direct Liquid Cooling container
- For each server:
- 4xGPU HGX NVIDIA H100 SXM5 700W 94G HBM2e GPU Board94GB unified memory,
- 2xCPU AMD EPYC 9654 @ 2.4Ghz (96 cores),
- Connected in 100Gbps Infiniband.