KB Article – Web Publication

Overview

NVIDIA HGX and DGX are both high-performance AI computing platforms built around multi-GPU NVIDIA accelerators, NVLink/NVSwitch technology, and high-speed networking. The key architectural difference is the level of integration: HGX is a modular platform/reference architecture that system manufacturers integrate into their own servers, whereas DGX is a complete NVIDIA-designed and validated AI system delivered as an integrated platform.

Architecture at a Glance

1. HGX Architecture

HGX is best understood as the GPU-centric building block used by NVIDIA ecosystem partners to create high-end AI and HPC servers. An HGX platform combines multiple SXM GPUs with high-bandwidth GPU-to-GPU interconnects and associated networking interfaces. The exact CPU, memory, storage, chassis, power, cooling, and network implementation can be integrated by the server OEM.

For example, current NVIDIA HGX reference configurations include 8-GPU systems. NVIDIA's certification documentation describes configurations such as 2-8-9-400 for HGX H100/H200, meaning 2 CPUs, 8 GPUs, 9 network adapters, and an average network bandwidth of 400 Gb/s.

2. DGX Architecture

DGX is a complete NVIDIA AI system. Instead of supplying primarily the accelerated compute building block, NVIDIA defines and validates the complete system architecture, including GPUs, NVSwitch, CPUs, system memory, storage, networking, power, thermal design, management, and the DGX software environment.

As an example, the DGX H100/H200 platform uses eight NVIDIA GPUs, NVSwitch-based GPU interconnects, high-speed ConnectX networking, NVMe storage, large system memory, and integrated out-of-band management.

3. Key Architectural Difference

The simplest way to remember the difference is:

HGX = NVIDIA GPU platform/building block → OEM integrates the complete server.

DGX = NVIDIA complete AI system → NVIDIA integrates and validates the complete platform.

4. GPU and NVLink Architecture

Both platforms can use the same class of NVIDIA SXM GPUs and NVSwitch technology. Therefore, HGX versus DGX is not primarily a difference in GPU architecture. The major distinction is the system-level integration around the GPU subsystem.

For example, NVIDIA documents HGX H100/H200/B200 as 8-GPU systems, while DGX H100/H200 systems also use eight GPUs. In an HGX deployment, the server partner builds the surrounding system to the NVIDIA reference and certification requirements; in DGX, NVIDIA supplies the complete validated system.

5. Networking Architecture

HGX provides flexibility for system partners to integrate high-speed networking according to the validated reference architecture. Depending on the generation and configuration, this can include ConnectX NICs or newer NVIDIA SuperNIC solutions and Ethernet or InfiniBand fabrics.

DGX provides a more prescriptive networking design. For example, NVIDIA's DGX H100/H200 documentation specifies high-speed ConnectX-7 networking for cluster, storage, and management functions.

6. Operational Perspective for Data Center Engineers

·         Hardware troubleshooting: HGX: OEM-specific components and firmware may require partner documentation. DGX: standardized NVIDIA system architecture makes troubleshooting more consistent.

·         BMC / management: HGX: implementation can vary by OEM. DGX: NVIDIA-defined management architecture.

·         Firmware / BIOS: HGX: coordinate NVIDIA GPU/platform requirements with OEM firmware packages. DGX: follow NVIDIA DGX-qualified firmware/software releases.

·         GPU/NVLink health: Both require checking GPU, NVLink and NVSwitch health; DGX provides a more standardized platform.

·         InfiniBand / Ethernet: Both can be deployed with high-speed NVIDIA networking. The exact topology depends on the system and cluster design.

·         Spare parts: HGX: often depends on the server OEM. DGX: system-level parts and support are tied to the NVIDIA DGX platform.

7. Summary

HGX and DGX can provide very similar GPU compute capabilities because both are built around NVIDIA's multi-GPU scale-up architecture. The architectural distinction is at the system-integration level. HGX is a flexible, partner-integrated platform designed to let OEMs build customized AI servers, while DGX is NVIDIA's complete, validated, ready-to-deploy AI system with a tightly integrated hardware and software stack.

Quick Reference

Sources

·         NVIDIA HGX Platforms Documentation – https://docs.nvidia.com/hgx-platforms/

·         NVIDIA Certified Systems – HGX Reference Configurations – https://docs.nvidia.com/certification-programs/latest/nvidia-certified-systems.html

·         NVIDIA HGX Platform – https://www.nvidia.com/en-in/data-center/hgx/

·         NVIDIA DGX H100/H200 User Guide – https://docs.nvidia.com/dgx/dgxh100-user-guide/dgxh100-user-guide.pdf

·         NVIDIA Enterprise Reference Architecture – HGX H100/H200/B200/B300