NVIDIA has announced the launch of the H200 and GH200 product lines. These are NVIDIA's most powerful chips ever, building on the existing Hopper architecture and adding more memory and computing power to power future generations of AI supercomputers.
The H200 comes with 141GB of HBM3e memory running at about 6.25 Gbps, with six HBM3e stacks bringing 4.8 TB/s of total bandwidth per GPU. The original H100 came with 80GB of HBM3, which corresponded to a total bandwidth of 3.35 TB/s, which is a huge improvement. Compared to the SXM version of the H100, the SXM version of the H200 increases memory capacity and total bandwidth by 76% and 43%, respectively. In terms of raw computing power, though, the H200 won't change much, except in individual application scenarios that will benefit from larger memory configurations.
This time NVIDIA also brought the GH200, which combines the H200 GPU and Grace CPU, combining the Hopper architecture GPU with the Arm architecture Grace CPU, using NVLink-C2C to connect the two. Each Grace Hopper Superchip contains 624GB of memory with 144GB of HBM3e and 480GB of LPDDR5x memory.
The Alps supercomputer at the Swiss National Supercomputing Center is likely to be one of the first Grace Hopper supercomputers to go into service next year, though it will still be a GH100.The first GH200 system to go into service in the U.S. will be the Venado supercomputer at Los Alamos National Laboratory, and the Texas Advanced Computing Center's (TACC) Vista The Texas Advanced Computing Center's (TACC) Vista system also uses the Grace Hopper Superchip, although it is not clear if it is a GH200.
The largest known supercomputer to use GH200 is the Jupiter at the Jϋlich Supercomputing Center, which will house nearly 24,000 GH200 chips and provide 93 ExaFLOPS of AI computational performance, in addition to 1 ExaFLOPS of traditional FP64 computational performance.