NVIDIA MGX is a standardized modular server platform and reference design that NVIDIA will introduce at Computex 2023. Its goal is to simplify the server gas pedal development process for NVIDIA and its OEM and ODM partners.
Key features of NVIDIA MGX:
- Accelerated time to market: MGX is designed to dramatically reduce the time required to develop, test, and qualify server systems from the typical 18-24 months to a few months. This is achieved by standardizing internal components and allowing partners to choose from a menu of certified components that are guaranteed to work together.
- Modularity and flexibility: MGX allows server designers to choose different combinations of processors (CPUs), graphics processing units (GPUs) and specialized network adapters (DPUs) in a single chassis. This allows for more than 100 application-specific configuration options.
- Standardization and Consistency: The MGX platform standardizes the chassis in which HGX/DGX components are enclosed, while providing the ability to upgrade systems in the field. It also ensures a degree of hardware consistency to accelerate the deployment of AI and HPC systems.
Supported components and configurations:
- The MGX platform supports NVIDIA's full portfolio of GPU gas pedals, including the Hopper H200 and Lovelace L4/L40S.
- On the CPU side, it supports both x86 processors (presumably Intel and AMD) and Grace-Grace Superchip (GG200) and Grace-Hopper Superchip (GH200) processors.
- On the networking front, MGX includes ConnectX-7 (hybrid InfiniBand/Ethernet) network interface cards and BlueField-3 DPUs.
- Example configurations include a 4U server with two x86 CPUs, four H200 NVL GPUs and, a pair of BlueField-3 DPUs.
