NVIDIA has released its new H200Hopper GPU, which now features the world’s fastest HBM3e memory from Micron. In addition to the new artificial intelligence platform, NVIDIA also announced a major supercomputing win with its GraceHopper superchip, which now powers the ExaflopJupiter supercomputer.
NVIDIA's H100 GPU is by far the most in-demand artificial intelligence chip in the industry, but NVIDIA wants to provide customers with even higher performance. HGXH200 is the latest artificial intelligence high-performance computing platform, driven by H200TensorCoreGPU. These GPUs feature the latest Hopper optimization technology in both hardware and software, while delivering the world's fastest memory solution to date.
NVIDIAH200GPU is equipped with Micron's HBM3e video memory solution, with a capacity of up to 141GB and a bandwidth of 4.8TB/s. Compared with NVIDIAA100, the bandwidth has increased by 2.4 times and the capacity has doubled. In applications such as Llama2 (70 billion parameter LLM), this new memory solution nearly doubles NVIDIA’s AI inference performance compared to the H100 GPU.
In terms of solutions, NVIDIAH200GPU will be widely used in HGXH200 servers with 4-way and 8-way GPU configurations. The 8-way H200 GPU configuration in the HGX system provides 32PetaFLOPs of FP8 computing performance and 1.1TB of memory capacity.
The GPUs will also be compatible with existing HGXH100 systems, making it easier for customers to upgrade their platforms. NVIDIA partners, such as ASUS, ASRockRack, Dell, Eviden, GIGABYTE, Hewlett Packard Enterprise, Inglas (Ingrasys), Lenovo (Lenovo), QCT, Wiwynn (Wiwynn), Supermicro (Supermicro) and Wistron (Wistron), these manufacturers will provide updated solutions when the H200GPU is launched in the second quarter of 2024.
In addition to announcing the H200 GPU, NVIDIA also announced a large-scale supercomputer project powered by its GraceHopper superchip (GH200). This supercomputer is called "Jupiter" and is located at the Forschungszentrum Jülich in Germany. It is part of the European High-Performance Computing Joint Undertaking (EuroHPC Joint Undertaking) and is contracted by Eviden and ParTec. This supercomputer will be used in materials science, climate research, drug discovery and other fields. This is also the second supercomputer released by Nvidia in November. The previous one was Isambard-AI, which can provide up to 21 Exaflops of artificial intelligence performance.
In terms of configuration, the Jupiter supercomputer is based on Eviden's BullSequanaXH3000 and adopts a fully liquid-cooled architecture. It has a total of 24,000 Nvidia GH200 GraceHopper super chips, which are interconnected through the company's Quantum-2 Infiniband. Considering that each Grace CPU contains 288 Neoverse cores, we're looking at Jupiter having nearly 7 million ARM cores in the CPU alone (6.912 million to be exact).
Performance metrics include 90Exaflops for AI training and 1Exaflop for high-performance computing. The supercomputer is expected to be installed in 2024. Overall, these are major updates for the future as Nvidia continues to lead the world of artificial intelligence with its powerful hardware and software technologies.