All articles tagged with #grace hopper
NVIDIA's Grace Hopper (GH200) and Grace Superchip, both based on Arm Neoverse V2, are not compatible and cannot be used interchangeably in servers. The two parts have different connectors, signaling differences, and power requirements. NVIDIA does not sell a standardized half-width carrier board for these modules, which has led to confusion in the industry.
NVIDIA has announced that their Grace Hopper "superchip" has entered full production, combining a Grace CPU and Hopper H100 GPU to deliver a semi-integrated CPU/GPU product for AI models. The Grace CPU packs 72 CPU cores and comes with up to 480GB of LPDDR5X memory, while the Hopper GPU brings just shy of 1 EFLOPS of FP16 matrix math throughput for AI workloads, as well as 80GB of HBM3 memory. NVIDIA is also building its first DGX system around the chip, a full-on multi-rack computational cluster called the DGX GH200 AI Supercomputer, which is a complete, turn-key, 256 node GH200 cluster designed for training large AI models.