Nvidia Unveils First Discrete Grace CPU and Hopper H100 Data Center GPU
Nvidia introduces its first-ever discrete CPU Superchip alongside the powerful Hopper H100 GPU, both designed specifically for AI and data center workloads rather than consumer gaming.
Nvidia CEO Jensen Huang kicks off the Graphics Technology Conference (GTC) by unveiling the Grace CPU Superchip, marking the company's first-ever discrete CPU. This processor combines two Grace CPUs using NVLink technology, similar to Apple's M1 Ultra, but targets artificial intelligence, data science, and high-memory applications instead of general consumer computing. The 144-core CPU utilizes ARM architecture despite Nvidia's previously abandoned attempt to purchase the company.
In addition to the Grace Superchip, Nvidia reveals its next-generation Hopper architecture through the H100 GPU, a chip built specifically for data center accelerators rather than consumer gaming graphics cards. Manufactured on TSMC's N4 process, the H100 packs 80 billion transistors and stands as the first GPU to support PCIe 5.0 and HBM3 memory. Nvidia claims that just 20 of these GPUs handle the equivalent of the entire world's internet traffic.
Customers access the H100 through Nvidia's fourth-generation DGX servers, which combine eight GPUs to deliver 32 petaFLOPs of AI performance. For even greater demands, the DGX H100 SuperPod links 32 of these servers to provide 20TB of HBM3 memory and 1 exoFLOP of AI performance. Nvidia debuts this immense architecture in its own EOS supercomputer, which integrates 18 SuperPods containing a total of 4,608 H100 GPUs.