At the SC23 supercomputing conference, NVIDIA announced the launch of the HGX H200 computing platform, featuring the H200 Tensor Core GPU, to enhance various cloud execution entities and boost artificial intelligence performance.
Compared to the previously released H100, NVIDIA states that the H200 integrates HBM3e memory, enabling higher data processing throughput and synergizing with the GPU to accelerate AI computing and large-scale natural language models. With a total capacity of 141GB of HBM3e memory and a transfer rate of up to 4.8TB per second, this will double the memory capacity and increase the memory bandwidth by 100 times compared to the previously released A2.4.
By using H100-compatible software updates, NVIDIA emphasized that the H200 will be able to double the efficiency of running large natural language models such as NVIDIA TensorRT-LLM and Meta Llama 2.
As for the HGX H200 motherboard using H200, it can correspond to four-way or eight-way configurations and is compatible with the previous generation HGX H100, which means it can be upgraded directly by replacement. In addition, it can also be paired with the Grace Hopper Superchip GH200 to improve overall computing performance and can be deployed in cloud, ground, hybrid cloud, or edge environments. It is expected that Aeon, ASUS, Dell, Eviden, GIGABYTE, HPE, InnoBridge Technology, Lenovo, QCT, Supermicro, Wistron and Wiwynn Technology will launch application design server products, which are expected to be available in the global market in the second quarter of 2024.
