NVIDIA CEO Jensen Huang returned to SIGGRAPH to discuss the rise of generative AI and unveiled announcements including the next-generation GH200 Grace Hopper Superchip platform, NVIDIA AI Workbench and upgrades to NVIDIA Omniverse. These additions are aimed at combining AI, graphics and virtual worlds, according to an Aug. 8 NVIDIA blog post.
"The generative AI era is upon us, the iPhone moment if you will," Huang said in the post. "Graphics and artificial intelligence are inseparable. Graphics needs AI, and AI needs graphics."
These developments were presented at SIGGRAPH in Los Angeles, Calif., highlighting NVIDIA's continuous efforts in advancing AI and graphics integration, the blog reported. SIGGRAPH is the top computer graphics conference. Huang made his return to the conference as generative AI continues to spread throughout an increasingly digital and hyperconnected world.
The GH200 Grace Hopper Superchip platform, a new unified toolset that enables simpler model tuning and deployment on NVIDIA AI platforms, and a significant upgrade to NVIDIA Omniverse with generative AI and OpenUSD are among the news highlights, according to the blog. The announcements focus on uniting all of the technological advancements from the last 10 years, including AI, virtual worlds, acceleration, simulation, cooperation and more.
NVIDIA changed graphics five years ago at SIGGRAPH by introducing real-time ray tracing and AI to GPUs. Huang noted, "while we were reinventing computer graphics with artificial intelligence, we were also reinventing the GPU for AI," the blog reported.
As a result, systems with greater processing capacity, such as the NVIDIA HGX H100, which uses eight GPUs and a total of 1 trillion transistors to dramatically accelerate systems based on CPU, are now available, according to the blog. The Grace Hopper Superchip, or NVIDIA GH200, which combines a 72-core Grace CPU with a Hopper GPU and entered full production in May, was developed by NVIDIA to maintain the momentum of AI.
Huang said an extra version of the NVIDIA GH200, which is already in production, will be added with state-of-the-art HBM3e memory, the post reported.
He also revealed the GH200 Grace Hopper superchip platform, which has the capacity to connect several GPUs for outstanding performance and an easily expandable server design, the blog post noted. The new platform will be offered in a variety of configurations and is designed to handle the most complicated generative workloads in the world, including big language models, recommender systems and vector databases.
The dual configuration consists of a single server with 144 Arm Neoverse cores, eight petaflops of AI capabilities and 282GB of the most recent HBM3e memory technology, the blog said. It offers up to 3.5 times more memory capacity and three times more bandwidth than the current generation offering. In the second quarter of 2024, leading system manufacturers plan to deliver systems based on the platform.
