NVIDIA, the leader in accelerated computing and generative AI, has announced its next-generation GH200 Grace Hopper Superchip platform, which features the world’s first HBM3e processor. The platform is designed to handle the most complex and large-scale generative AI workloads, such as natural language processing, recommender systems and vector databases.

The GH200 Grace Hopper Superchip platform consists of a 72-core Grace CPU and a GH100 compute GPU, both based on Arm architecture. The CPU has 480 GB of ECC LPDDR5X memory, while the GPU has 141 GB of HBM3e memory, which is 50% faster and offers 25% more bandwidth than the previous HBM3 memory. The GPU can also be connected with other GPUs using NVIDIA NVLink, a high-speed, coherent interconnect that gives the GPU full access to the CPU memory. This enables the platform to run models up to 3.5 times larger than the previous version, with a combined memory capacity of 1.2 TB and a bandwidth of 10 TB/s.

The GH200 Grace Hopper Superchip platform is compatible with the NVIDIA MGX server specification, which allows any system manufacturer to easily integrate it into their servers. The platform is expected to be available in Q2 of 2024, while the current version with HBM3 memory will be available starting next month. NVIDIA claims that the new platform will deliver exceptional performance and scalability for generative AI applications in data centers.

Source: NVIDIA