Nvidia achieved immense sales growth last year, which was largely made possible by the AI accelerators it offers. One of these models, the Grace Hopper superchip, has now appeared outside of supercomputers for the first time. The massive accelerator card, which when fully expanded combines 72 ARM cores with a GPU with 16,896 shaders, is being offered by a German dealer on the website gptshop.ai.
Workstation with GH200
The model is not sold individually, but as part of finished workstations. These can be configured in various ways. For example, there are two variants of the GH200 to choose from: the older model with 96 GiByte HBM3 memory and the newer variant with 144 GiByte HBM3 memory. In both cases, 480 GiBytes of LPDDR5X memory are also available for the built-in ARM CPU.
Source: gptshop.ai A look into the system: The edge of the case is equipped with Noctua fans throughout. The cheapest basic model of the workstation is offered for 45,900 euros, and the most expensive basic variant costs 62,500 euros. In both cases, the majority of the price is probably for the GH200. There are also additional accessories, such as two 2,000 W power supplies and numerous Noctua fans for cooling. A 1 TB SSD is also installed ex works, and optionally there are additional SSDs, hard drives, an Nvidia network card or a Geforce RTX 4060 to enable the system to output graphics.
Equally interesting: Custom chips: Nvidia will also serve game consoles with “custom silicon” in the future
However, given the high price and lack of intended use, the workstation is unlikely to be an option for end customers, and supercomputers will probably continue to be the more economical choice for scientific use. The workstation use of the GH200 is definitely exciting, even if the high purchase costs can hardly be justified purely by technical interest.
Those: via Tom’s Hardware