NVIDIA MGX: A Modular Blueprint for AI Data Centers

The rapid advancement of generative AI, large language models (LLMs), and high-performance computing has placed unprecedented demands on data center infrastructures. In response, NVIDIA has introduced MGX, a modular reference architecture aimed at transforming how enterprises and cloud providers build scalable AI factories, according to NVIDIA.

Modular Architecture: A Necessity

NVIDIA MGX leverages a building-block approach, enabling partners to design multiple systems efficiently, reducing both development costs and time-to-market. The architecture supports various product generations and offers hundreds of combinations of GPUs, DPUs, CPUs, storage, and networking, catering to AI, high-performance computing (HPC), and digital twin applications.

Three primary trends are propelling the adoption of NVIDIA MGX:

• Power Density and Cooling: Modern AI computations necessitate increased power density and liquid-cooled infrastructure. For instance, NVIDIA Blackwell GPUs demand up to 120 kW per rack. MGX addresses these needs with liquid-cooled busbars and manifolds, facilitating efficient high-density deployments.
• Heterogeneous Workload Support: Enterprises are managing diverse workloads within single data centers. MGX’s modular compatibility allows organizations to tailor infrastructure for specific workloads without redesigning entire systems.
• Supply Chain Agility: Pre-integration of around 80% of components streamlines the build process, reducing deployment timelines from 12 months to under 90 days.

Standardized architectures like MGX ensure stable, reliable server deployments that support evolving performance needs while maintaining interoperability. The ecosystem allows flexible component selection, reducing investment risks and lead times.

Inside the MGX Rack System

The NVIDIA MGX rack system comprises compute trays and NVLink switch trays. Compute trays integrate powerful combinations of CPUs and GPUs, delivering core performance for AI training and simulation workloads. NVLink switch trays provide the high-speed interconnect fabric necessary for efficient GPU-to-GPU communication.

Beyond compute and switch trays, the MGX rack includes a robust foundation of mechanical, electrical, and cooling infrastructure, ensuring operational efficiency and scalability.

Transforming AI Factory Design

NVIDIA MGX offers significant advantages across the data center ecosystem. For system builders, it reduces R&D costs by leveraging shared reference designs and allows comprehensive certification for the NVIDIA software stack. Data center operators benefit from seamless scalability and reduced total cost of ownership, while AI workloads achieve unprecedented performance levels.

With over 200 ecosystem partners adopting MGX components, enterprises now have a future-proof path to exascale AI, ensuring that AI factories can evolve alongside silicon innovations.

Image source: Shutterstock