Nvidia reveals a whole new kind of Ethernet for generative AI
Nvidia founder and CEO Jensen Huang delivered an electrifying keynote that brought many shocking new developments. At COMPUTEX Taipei conference, Huang presented products and systems using accelerated computing for everything from manufacturing to 5G connectivity.
Nvidia’s DGX GH200 AI supercomputer stands out among these new offerings as one of the most remarkable. Combining energy-efficient Nvidia Grace CPU cores with Nvidia Quadro P5000 GPUs, this supercomputer is known for being highly effective yet cost effective at running generative AI and can train massive deep learning models much more rapidly than traditional computers.
Nvidia has unveiled another of its offerings – Spectrum-X networking platform – designed specifically to support multi-tenant, hyperscale generative AI data centers. Using Ethernet networks for artificial intelligence workload transfer and performance enhancement of up to four times over existing solutions. Spectrum-X is expected to become a key technology driver behind an imminent wave of multi-tenant AI clouds that promise more affordable cloud services with higher-quality and more precise results for customers.
Nvidia made another exciting announcement with cuLitho, an artificial intelligence-powered design tool that will enable chip fabs to produce three to five times more photomasks used for lithographic processes that print circuit widths onto silicon chips. Nvidia estimates this will also result in significantly increased production yields and reduced energy consumption by up to nine times, all part of their plan for moving away from computing being focused purely on performing tasks to one that uses computers as information generators.
Generative AI is revolutionizing all industries from advertising to genomics and beyond. By processing inputs like images, audio files and text into outputs such as photos, videos or new information that can be shaped the user wants it, Generative AI enables businesses to improve processes that touch on everything from advertising and genomics. Generative AI even transforms that information into physical products like proteins or DNA molecules for transformation!
Nvidia’s new MGX servers provide an agile platform for creating AI, HPC and omniverse apps. Nvidia has already partnered with system vendors such as Dell, Lenovo and Supermicro to provide customized options tailored specifically to individual requirements.
Nvidia’s MGX servers support a comprehensive selection of GPUs, CPUs and deep learning accelerators from Nvidia as well as its new accelerated Ethernet platform Spectrum-X. In addition, SDKs such as Cumulus Linux pure SONiC NetQ allow developers to build AI applications such as Nvidia Isaac for robot simulation testing as well as Metropolis for factory quality control, vision AI solutions or factory quality control applications on these servers. Furthermore, SoftBank and Nvidia have collaborated on building new generative AI and 5G data centers using multi-tenant servers powered by Nvidia Grace Hopper superchips from Nvidia.