The latest generation of Nvidia flagship GPUs is ready. The H200 Hopper and GH200 are intended to accelerate the development of AI.

Nvidia announces new GPUs today. These are not the first, as the H200 and GH200 immediately become the new spearhead of Nvidia’s portfolio. The H200 GPU is based on the Hopper architecture, making it a souped-up version of the current flagship (and cash cow) H100. With new GPUs, Nvidia wants to drive a “performance boost” through the development of artificial intelligence next year.

The most important upgrade can be found in the H200 Hopper’s memory. This is the manufacturer’s first GPU to integrate faster HBM3e memory. The H200 Hopper can provide 141GB of storage at 4.8 terabytes per second; almost twice the capacity and 2.4 times the bandwidth compared to the previous model H100. The GPU will be available in two configurations: An eight-tier configuration gives you 32 petaflops of FP8 and 1.1TB of aggregated memory.

With HBM3e memory, the H200 Hopper is also better able to support increasingly complex AI models. As an example, Nvidia Metas LLaMA 2, a model with 70 billion parameters. With this model, the H200 achieves twice as high an inference speed as its predecessor, the manufacturer proudly writes in the press release.

Great chip

In addition, Nvidia is also launching a successor to its “super chip” Grace Hopper today, the GH200. It combines a CPU and a GPU in one socket and was designed for supercomputers. The GH200 delivers two hundred exaflops of computing power to solve complex scientific problems.

HPE, among others, has already announced a new Cray supercomputer with four GH200 chips on board. The most complex machine that will feature in Nvidia’s new offering is Jupiter, Europe’s first exascale computer, which will be operational starting next year. The first appearances with the Nvidia H200 may have to wait until the end of 2024.

The H200 and GH200 may also be the last chips to carry the Hopper brand name. Starting next year, Nvidia will name its flagship CPUs after the American statistician and mathematician David Harold Blackwell. The B100 will be the first chip of this new generation and is expected next year.