NVIDIA HGX AI Factory

NVIDIA Enterprise Reference Architecture

• Abstract
• Overview
• Components
  • NVIDIA HGX B300 Baseboard
  • NVIDIA HGX B300 Systems
  • Networking
    • Compute (Node East/West) Ethernet Networking
    • Converged (Node North/South) Ethernet Networking
• Networking Hardware
  • NVIDIA ConnectX-8 HCA (on baseboard)
  • NVIDIA BlueField-3 B3240 DPU
  • NVIDIA SN5610 Switch
  • NVIDIA SN2201 Switch
• Networking Physical Topologies
  • Compute (Node East/West) Network
    • Compute Fabric Excluded for Pure Inference Deployments
    • Multi Plane Topology Approach
      • Dual Plane Topology
      • Single Plane Topology
  • CPU Converged (Node North/South) Network
  • Out-of-Band Management (Node) Network
• Networking Logical Architecture
  • Enterprise RA Scalable Unit (SU)
  • Spine-Leaf Networking
  • Out-of-Band (OOB) Management Network
  • Optimized Use Case
    • Architecture Details
    • Compute (Node West/East) Fabric Table
    • Converged (Node North/South) Fabric Table
    • 32 Nodes with 256 NVIDIA HGX B300 SXM GPUs
    • 64 Nodes with 512 NVIDIA HGX B300 SXM GPUs
    • 128 Nodes with 1024 B300 HGX GPUs
• Software
• NVIDIA-Certified Storage
• Services
• Summary

Appendix

• Appendix: Node Configurations
  • Compute Nodes
  • Switches and Cables

Notices

• Notices
  • Notice