SuperServer

AI Factory Solutions with NVIDIA HGX™ B300

Key Applications

  • Artificial Intelligence (AI)
  • HPC
  • AI/Deep Learning Training
  • AI Inference
  • Massive-scale AI Training
  • LLMs & Multi-modal LLMs

Key Features

  1. Full-stack solutions based on reference architectures that include Supermicro systems, NVIDIA GPUs, NVIDIA software, and NVIDIA networking
  2. Up to 256x NVIDIA HGX™ B300 GPUs providing up to 73.7TB total HBM3e memory
  3. Compatibility with the NVIDIA software stack (NVIDIA AI Enterprise and NVIDIA Run:ai)
  4. Plug-and-play solution with systems fully integrated into racks and tested before shipment and on-site deployment
  5. NVIDIA Spectrum™-X Ethernet Compute Fabric, Converged Network, and Out-of-Band Management included
  6. Supermicro AI Factory solutions are endorsed by NVIDIA for Infrastructure Configuration, Spectrum-X networking, and Software Reference Stack and based on the NVIDIA Enterprise Reference Architecture for HGX B300.
...
...
Cluster Overview
Compute Nodes
  • 4, 8, or 32 nodes (32–256 GPUs) per scalable unit
Compute Racks
  • Up to 8x compute racks per scalable unit
Networking Racks
  • 1x networking rack per scalable unit
Rack Form Factor
  • 48U 19"-width rack enclosure
Rack Dimensions
  • 800 mm × 1200 mm
Total GPUs
  • Up to 256x NVIDIA B300 GPUs
Total GPU Memory
  • Up to 73.7TB HBM3e
Estimated Total Power
  • Up to 500 kW
Compute Node Configuration
System SYS-822GS-NB3RT
GPU
  • NVIDIA HGX B300 8-GPU
CPU
  • 2x Intel® Xeon® 6767P Processor 64-Core 2.40GHz 336MB Cache (350W)
Memory
  • 24x 64GB DDR5 6400MT/s ECC RDIMM
Local Node Storage
  • 8x Hot-swap E1.S NVMe, 2x NVMe M.2
Networking
  • 1x NVIDIA BlueField®-3 (B3240), 8x Integrated NVIDIA ConnectX®-8 SuperNICs, up to 800Gb/s
Power
  • 15kW (full load power draw)
Compute Rack Configuration
Compute Nodes
  • 4x Compute Nodes per rack
Total GPUs
  • 32x GPUs per rack
Total CPUs
  • 8x CPUs per rack
Estimated Rack Power
  • ~60 kW per rack
Networking Overview
Compute Fabric
  • Up to 12x NVIDIA Spectrum-4 800GbE Ethernet Switches (SN5610)
Converged Network
  • Up to 2x NVIDIA Spectrum-4 800GbE Ethernet Switches (SN5610)
Out-of-Band Management
  • Up to 4x NVIDIA Spectrum 1GbE/10GbE Ethernet Switches (SN2201)
Software Overview
Application Software
  • NVIDIA AI Enterprise
Infrastructure Software
  • NVIDIA Run:ai
Services
Onsite Deployment
  • Plug-and-play on-site deployment and validation
Rack Integration
  • Full rack-level design, component installation, and cabling
  • System-level testing of each node (L11)
  • Cluster-level multi-node testing (L12)

SuperCluster SRS-48UAC-B300SX New

Certains produits peuvent ne pas être disponibles dans votre région