Nvidia's latest Grace Hopper superchip (GH200) processor has been unveiled in nine new systems around the world. The GH200 is a chip recently released by Nvidia that eliminates the PCI bus in the CPU/GPU communication path.
As announced by NVIDIA at ISC 2024, the new Grace Hopper-based supercomputers that will come online include CEA in France and EXA1-HE in Eviden; Helios, Cyfronet Academic Computer Center, Poland; Hewlett-Packard Enterprises' Alps at the Swiss National Supercomputing Center; JUPITER at the Jülich Supercomputing Center in Germany; DeltaAI at the National Center for Supercomputing Applications at the University of Illinois at Urbana-Champaign; Miyabi of the Japan Joint Center for Advanced High Performance Computing, which was jointly established by the Center for Computer Science at the University of Tsukuba and the Center for Information Technology at the University of Tokyo.
In April, the French Alternative Energies and Atomic Energy Commission (CEA) and Eviden, part of the Atos group, announced the delivery of the EXA1-HE supercomputer based on Eviden's BullSequana XH3000 technology. The BullSequana XH3000 architecture provides a new patented warm water cooling system, while the EXA1-HE is equipped with 477 compute nodes based on Grace Hopper.
"AI is accelerating research on climate change, accelerating drug discovery, and making breakthroughs in dozens of other areas," said Ian Buck, vice president of hyperscale and high-performance computing at NVIDIA. "Grace Hopper-powered systems are becoming an essential part of high-performance computing because of their ability to transform the industry while improving energy efficiency."
In addition, systems from Isambard-AI and Isambard 3 at the University of Bristol in the United Kingdom, as well as Los Alamos National Laboratory and Texas Advanced Computing Center in the United States, have joined the wave of NVIDIA's arm-based supercomputers using the Grace CPU and Grace Hopper platforms.
Eliminate the PCI intermediate path
The design of the Grace Hopper is a Hopper GPU combined with an arm-based Grace CPU. Prior to Grace Hopper, CPUs (typically x86) used one or more PCI bus-based GPUs. These additional GPUs must communicate via the PCI bus for data transfer, and as a result, two or more different memory domains are created: the CPU domain and the GPU domain. Data transfer between these domains must pass through the PCI bus, which often becomes a bottleneck.
The Grace Hopper uses the NVLink-C2C interconnect to connect the CPU and GPU, providing a single shared memory domain. It's a memory-consistent, high-bandwidth, and low-latency interconnect. It is the core of the Grace Hopper processor and provides up to 900Gb/s of total bandwidth.
Sovereign AI and high-performance computing
The drive to build new, more efficient AI-based supercomputers is accelerating as countries around the world recognize the strategic importance of sovereign AI – and countries are investing in domestically owned and hosted data, infrastructure, and tech talent to foster innovation.
Combining the arm-based Grace CPU and Nvidia's Hopper GPU architecture, the GH200 is a new optimized design for scientific supercomputing centers around the world. Many centers plan to go from system installation to real scientific research in a matter of months rather than years.
For example, the Isambard-AI Phase 1 consists of the HPE Cray supercomputer EX2500 and 168 NVIDIA GH200 superchips, making it one of the most efficient supercomputers ever built. When the remaining 5,280 Nvidia Grace Hopper superchips arrive at the University of Bristol's National Composites Centre this summer, the performance will be improved by a factor of 32.
Simon McIntosh-Smith, professor at the University of Bristol, said: "Isamard-AI positions the UK as a global leader in artificial intelligence and will help promote open science innovation at home and abroad." "Through our partnership with NVIDIA, we delivered the first phase of the project in record time, and when it is completed this summer, we will see a huge leap in performance to advance data analytics, drug discovery, climate research and more."