NVIDIA HGX

Type: Platform Tags: NVIDIA, GPU, Hardware, HGX, Blackwell, Rubin, Data Center, SXM, Multi-GPU Related: NVIDIA-Blackwell-Architecture, NVIDIA-Vera-Rubin, NVIDIA-Vera-Rubin-POD, NVIDIA-Vera-CPU, NVLink, NVIDIA-GB200-NVL72, NVIDIA-GB300-NVL72, NVIDIA-HGX-AI-Factory, NVIDIA-DGX, NVIDIA-Spectrum-X, NVIDIA-Spectrum-X-Validated-Solution-Stack, NVIDIA-Quantum-X800-InfiniBand, NCCL, NVIDIA-ConnectX-InfiniBand, NVIDIA-ConnectX-9, NVIDIA-BlueField-DPU, NVIDIA-BlueField-4 Sources: NVIDIA official product page (live fetch 2026-04-10; updated from https://www.nvidia.com/en-us/data-center/hgx/, https://www.nvidia.com/en-us/data-center/technologies/rubin/, https://docs.nvidia.com/enterprise-reference-architectures/hgx-ai-factory/latest/index.html, https://docs.nvidia.com/networking/software/spectrumx-solution-stack/index.html, https://www.nvidia.com/en-us/networking/products/infiniband/quantum-x800/) Last Updated: 2026-05-09

Summary

NVIDIA HGX is a high-performance multi-GPU baseboard platform designed for AI training, inference, and HPC in data center servers. It connects 8 GPUs via NVLink in an SXM form factor, enabling OEM and ODM server builders to create NVIDIA-validated AI compute nodes. The platform spans multiple GPU generations including Blackwell (B200, B300) and Rubin.

Detail

Purpose

Provides a standardized, NVIDIA-validated multi-GPU baseboard that OEMs and ODMs use to build GPU-accelerated servers. Unlike DGX (NVIDIA’s complete turnkey system), HGX is the GPU board that goes into third-party server designs.

Current Configurations

SystemGPU GenNVLink GenGPU-to-GPU BWTotal NVLink BWTotal Memory
HGX Rubin NVL8Rubin6th Gen3.6 TB/s28.8 TB/s2.3 TB
HGX B300Blackwell Ultra5th Gen1.8 TB/s14.4 TB/s2.1 TB
HGX B200Blackwell5th Gen1.8 TB/s14.4 TB/s1.4 TB

Key Features

  • 8x SXM GPU form factor per baseboard
  • Fifth/sixth-generation NVLink for all-to-all GPU communication
  • Optional NVIDIA Vera CPU or x86-based CPU baseboard pairing
  • Advanced networking: up to 800 Gb/s (Quantum-X800 InfiniBand or Spectrum-X Ethernet)
  • NVIDIA BlueField-3 DPU integration for networking and security services
  • Validated reference designs for major server OEMs (Dell, HPE, Lenovo, Supermicro, etc.)

Performance (HGX Rubin NVL8 vs HGX B200)

  • 5.5x more NVFP4 inference performance
  • 3.5x higher inference performance vs. prior generation
  • 2.6x higher LLM training performance
  • 2x attention (transformer) performance improvement
  • Current public materials connect HGX B300 and HGX Rubin NVL8 to Blackwell Ultra and Vera Rubin platform roadmaps rather than requiring separate wiki pages for every HGX baseboard variant.
  • The current NVIDIA-HGX-AI-Factory Enterprise RA turns HGX B300 into a 2-8-9-800 AI factory pattern with Spectrum-X, ConnectX-8, BlueField-3, AI Enterprise, Run:ai, and NetQ.
  • NVIDIA-Spectrum-X-Validated-Solution-Stack includes B300/HGX-adjacent Spectrum-X stack validation signals, while NVIDIA-Quantum-X800-InfiniBand is the InfiniBand counterpart for 800 Gb/s AI networking.

Use Cases

  • Large language model training (up to multi-node scale)
  • Generative AI inference
  • High-performance computing and simulation
  • Enterprise AI at scale

Target Customers

OEM server builders (Dell, HPE, Lenovo, Supermicro, etc.) and their data center customers who want NVIDIA-validated GPU compute in their own server chassis and management ecosystem.

Relationship to DGX

HGX is the GPU baseboard; DGX is NVIDIA’s complete turnkey system built around the same GPU technology. HGX gives OEMs flexibility on chassis, cooling, and management software.

Connections

Resources