Liquid Cooled Rack-Scale Solution with 72 NVIDIA B300 GPUs and 36 Grace CPUs
Key Applications
AI/Deep Learning Training and Inference
Key Features
Rack-scale solution with NVIDIA GB300 Grace™ Blackwell Superchip providing 72 NVIDIA B300 GPUs and 36 Grace CPUs per rack
NVIDIA Blackwell Ultra with 288GB HBM3e per GPU
Direct Liquid-Cooling for up to 40% reduction in electricity cost for the data center
Comprehensive Service from consultation to full-scale deployment, providing all necessary parts, networking solutions, and onsite installation services
Up to 800Gb/s NVIDIA Quantum-2 InfiniBand or Spectrum-X Ethernet with integrated NVIDIA ConnectX®-8 SuperNICs
Product SKUs
Part Number
SRS-GB300-NVL72
Rack-Level Specifications
GPUs
72 NVIDIA B300 GPUs via NVIDIA GB300 Grace Blackwell Superchips
CPUs
36 NVIDIA Grace CPUs via NVIDIA GB300 Grace Blackwell Superchips
GPU memory
Up to 21 TB HBM3e
System Memory
Up to 17 TB LPDDR5X
Storage
Up to 144 E1.S PCIe 5.0 drive bays
Rack Configuration
Nodes
18x 1U ARS-121GL-NB3
Networking
9x NVLink Switch, 4-ports per compute tray connecting 72 GPUs to
provide 1.8TB/s GPU-to-GPU interconnect
Power
8x 1U 33kW (6x 5.5kW PSUs), total power 132kW
Liquid Cooling
1x in-rack Supermicro 250kW capacity CDU with redundant PSU and dual hot-swap pumps
Enclosure
Rack Enclosure
48U, 19”-width
Dimensions
W 600 x D 1068 x H 2236
Networking
Compute Fabric
NVIDIA Quantum-X800 InfiniBand or NVIDIA Spectrum-X Ethernet with integrated NVIDIA ConnectX®-8 SuperNICs for up to 800Gb/s compute fabric
In-Band Network
Up to 200Gb/s In-Band Management speeds with NVIDIA BlueField-3 DPUs
Out-of-band Network
1G/10G Ethernet out-of-band management
Liquid Cooling
Liquid Cooling Options
In-Rack CDU: Up to 250 kW capacity CDU with N+1 redundant pumps
In-Row CDU: Up to 1.8 MW capacity CDU with N+1 redundant pumps (supports up to 8 racks)
L2A Sidecar CDU: Up to 200 kW capacity CDU with N+1 redundant pumps (no facility water required)
