Nvidia launches DGX GH200, a supercomputer for artificial intelligence

At the preliminary event of Computex 2023, Jensen Huang, NVIDIA’s CEO, announced the launch of DGX GH200, a supercomputer for artificial intelligence, composed of 256 Grace Hopper superchips. He underscored that Grace Hopper superchips have entered full-scale production and, through a modular architecture, will cater to the diverse designs and accelerated computational needs of data centers worldwide.

Mr. Huang elucidated the superior computational capacity of GPU-accelerated computations, driving higher workloads than CPUs. He accentuated that generative artificial intelligence will revolutionize computational models and conceive novel interactive experiences such as translating text inputs into realistic video representations, creating music via textual descriptions, or even facilitating computer-generated singing of original compositions.

Nvidia DGX GH200

Alongside shifts in computational methods that are transforming traditional communication modalities, Mr. Huang anticipates further alterations in computation methods, propelled by the ubiquity of 5G and other network technologies.

The newly announced DGX GH200 AI supercomputer, featuring 256 Grace Hopper superchips, corresponds to a 144TB display memory capacity, a computational efficiency of 1 Exaflop, and caters to applications such as generative AI, recommendation systems, and data analysis.

NVIDIA has further declared the full-scale production of Grace Hopper superchips. NVIDIA’s MGX is providing system manufacturers with a modular architecture that meets the varied design and accelerated computational demands of data centers worldwide, including Cloudtech and Supermicro, which are leading the creation of more than 100 system configurations to expedite the processing of AI, HPC, and Omniverse workloads.

NVIDIA also announced the launch of NVIDIA Spectrum-X, an Ethernet platform designed for ultra-large-scale generative AI, which combines NVIDIA Spectrum-4, BlueField-3 DPU, and corresponding acceleration software to accelerate massive data transmission efficiency.

As for the development of the Omniverse platform, Jensen Huang elaborated that companies including Foxconn, Inventec, Pegatron, Quanta, and Wistron are using technologies such as Omniverse, Isaac Sim, and Metropolis to build virtual factories or expedite product manufacturing through robots and automated detection systems.

The collaboration between WPP and NVIDIA to develop a generative AI content engine for digital advertising is based on NVIDIA’s AI and Omniverse open platform, combined with creative 3D and AI tools, thus revolutionizing content and experiential perceptions.

In addition, Dah Ming Robotics has leveraged Isaac Sim on the Omniverse open platform to optimize the results of automated optical detection, increasing the speed of machine detection by 20%, and thereby enhancing the quality of electronic manufacturing products.

NVIDIA has also officially launched Isaac AMR, granting robots more advanced autonomous capabilities, and offering customers early access to experience this advancement. The robot reference design adopting the Nova Orin architecture will also be open for evaluation and testing.