Graphics processing unit (GPU) manufacturer Nvidia has announced a high-speed supercomputer called DGX GH200. The announcement was made at the Computex technology exhibition held in Taiwan this week.
Nvidia CEO Jensen Huang stated that DGX GH200 is designed to assist various technology companies in developing a range of artificial intelligence (AI) products or services, particularly for generative AI models.
Generative AI is an AI model that can create new art, works, or content (such as images, music, texts, videos, etc.) that have never existed or been created by humans.
Typically, generative AI relies on complex machine learning algorithms, which process stored data to combine and generate new content.
To support the performance of generative AI, DGX GH200 utilizes NVLink Switch System technology.
This technology allows 256 units of the high-speed GH200 Grace Hopper chips within DGX GH200 to be combined into a single GPU chip with exceptional performance.
With this technology, Nvidia claims that its latest supercomputer can achieve a performance of up to 1 exaflop and has a large shared memory capacity of up to 144 TB.
This performance achievement is comparable to the Frontier supercomputer located at the Oak Ridge National Laboratory in Tennessee, USA. According to the latest data, Frontier has a processing capability of 1.2 exaflops.
Compared to Nvidia’s previous generation of supercomputers, DGX GH200 is also claimed to provide 10 times more bandwidth and a performance that is 500 times faster.
The DGX GH200 supercomputer is expected to be available by the end of 2023.
Reportedly, several major companies involved in AI projects, such as Google, Meta, and Microsoft, have already shown interest in Nvidia’s new supercomputer.
They are expected to be among the first companies to utilize DGX GH200 in the future, as reported by KompasTekno from Engadget on Tuesday, May 30, 2023.
