Nvidia has made a major leap in the field of supercomputing with the unveiling of the Grace Hopper GH200 superchip. This innovative chip represents a significant advancement in artificial intelligence (AI) performance, reaching an impressive 200 exaflops.
The GH200 is based on the integration of the HGX H200 GPU with Hopper architecture and a Nvidia Grace CPU. This superchip is the result of combining the advanced capabilities of these two technologies, creating an integrated module designed for high-performance computing and large-scale AI applications.
The heart of the GH200, the HGX H200 GPU, is characterized by its Hopper architecture, which significantly improves performance compared to previous models. It incorporates NVIDIA H200 Tensor cores and HBM3e memory technology, optimized to accelerate generative AI tasks and large language models. This GPU offers 141GB of memory with a speed of 4.8 TB per second, doubling the bandwidth of the previous model, Nvidia A100.
The GH200 stands out for its ability to double the inference speed in large language models, such as Meta’s Llama 2. It also improves high-performance computing workloads thanks to high-speed interconnect technologies NVIDIA NVLink and NVSwitch. The GH200’s eight-way configuration provides over 32 petaflops of FP8 deep learning compute and 1.1TB of aggregate high-bandwidth memory.
This superchip is designed for a variety of data centers, including those operating on-demand, in hybrid cloud, and at the edge. Its compatibility with the infrastructures of leading companies in the sector such as Amazon Web Services, Google Cloud, Microsoft Azure, and Oracle Cloud expands its application field.
The GH200 is already in use in prestigious supercomputing centers worldwide, driving scientific innovation with its enormous processing capability. The early adoption of Grace Hopper by more than a hundred companies, organizations, and government agencies around the world highlights the importance and impact of this development in current supercomputing.
In summary, Nvidia’s Grace Hopper GH200 superchip marks a milestone in the history of supercomputing, opening new possibilities for research and development in the field of AI and high-performance computing.