NVIDIA has introduced the GH200 Grace Hopper platform, a next-generation solution designed to cater to the demands of accelerated computing and generative AI tasks.
This innovative platform is anchored by the Grace Hopper Superchip, featuring the world's first HBM3e processor. Engineered to tackle intricate generative AI workloads, such as large-scale language models, recommender systems, and vector databases, the GH200 Grace Hopper™ platform offers an array of configurations to suit diverse needs.
At the heart of this platform is a dual configuration that significantly outperforms its predecessor. With a single server housing 144 Arm Neoverse cores, delivering eight petaflops of AI performance and backed by 282GB of cutting-edge HBM3e memory technology, it boasts up to 3.5 times greater memory capacity and 3 times more bandwidth than its predecessor.
Jensen Huang, NVIDIA's founder and CEO, emphasizes the need for specialized accelerated computing platforms to meet the escalating demand for generative AI. The GH200 Grace Hopper Superchip platform is tailored to this demand, offering exceptional memory technology and bandwidth that enhances throughput. Its unique server design allows for seamless deployment across entire data centers.
The Grace Hopper Superchip at the core of this platform can be interconnected with additional Superchips via NVIDIA NVLink™ technology. This setup allows for collaborative deployment of large-scale models crucial to generative AI tasks. This high-speed, coherent technology grants GPUs full access to CPU memory, delivering a combined memory capacity of 1.2TB in the dual configuration setup.
Moreover, the HBM3e memory, boasting 50% faster performance than conventional HBM3 memory, contributes to a remarkable total combined bandwidth of 10TB/sec. This improved memory performance empowers the new platform to handle models 3.5 times larger than its predecessor, enhancing overall performance with 3 times faster memory bandwidth.
In light of the growing demand for the Grace Hopper platform, leading manufacturers have already begun offering systems built around the Grace Hopper Superchip. The forthcoming next-generation Grace Hopper Superchip platform, featuring the HBM3e processor, is fully compatible with the NVIDIA MGX™ server specification revealed at COMPUTEX. This compatibility ensures that any system manufacturer can swiftly and cost-effectively incorporate Grace Hopper into a wide range of over 100 server variations.
The platform is projected to be adopted by leading system manufacturers, with systems expected to launch in the second quarter of the calendar year 2024.
