Nvidia ships first servers with Grace CPUs in the first half of 2023

Nvidia will release its Grace processor for data centers and supercomputing in the first half of 2023. Server OEMs such as ASUS, Foxconn, Gigabyte and Supermicro, among others, will release servers based on the Arm chip during that period.

Nvidia comes first with different Grace reference platforms, which can be used in the servers of OEMs and contain differences in their amount of CPUs, GPUs and DPUs. The company comes with a CGX platform for ‘cloud graphics and gaming’, two different HGX variant for AI and HPC, and an OVX platform for ‘digital twins and the omniverse’. The company currently supplies those server platforms, but based on x86 processors from AMD and Intel.

Nvidia’s Grace Superchips and its various reference platforms

Those platforms will be based on the previously announced Grace CPU and Grace-Hopper ‘Superchips’. The former features two Grace data center processors on a single module, for a total of 144 Arm Neoverse cores. In addition, the module is equipped with Lpddr5x memory with a bandwidth of up to 1TB/s. The second Superchip has a Grace processor with 72 cores and an H100 GPU. To connect those chips, Nvidia uses its own NVLink-C2C interconnect† In addition, the systems can be equipped with a BlueField-3 dpuwhich are intended to accelerate network, storage, and security services for data centers, among other things.

According to Nvidia, several Taiwanese OEMs have plans to launch Grace systems in the first half of 2023. These include ASUS, Gigabyte, Foxconn, Qiwynn, QCT, and Supermicro. These OEMs make servers based on the aforementioned reference platforms, which include ‘blueprints and server baseboards’ with which OEMs can design their own server. The chips will be available in 2U servers, Nvidia says.