Specifications
Model
SRS-GB200-NVL72-M1
Platform
NVIDIA GB200 NVL72 SuperCluster
GPUs
72x NVIDIA Blackwell B200
CPUs
36x NVIDIA 72-core Grace Arm Neoverse V2
GPU Memory
Up to 372 GB HBM3e per superchip, 744 GB per tray
CPU Memory
Up to 480 GB LPDDR5X per superchip, 960 GB per tray
Compute Trays
18x 1U liquid-cooled, 2 GB200 superchips and 4 B200 GPUs each
NVLink Switch Trays
9x, 4 ports per compute tray
GPU-to-GPU Bandwidth
1.8 TB/s per GPU
Power Shelves
8x 1U 33 kW (6x 5.5 kW PSUs each) with built-in capacitor, 132 kW total
Operating Power
125 kW to 135 kW
Rack Dimensions
2236 x 600 x 1068 mm
In-Rack Cooling
4U 250 kW capacity CDU with redundant PSU and dual hot-swap pumps
In-Row Cooling
Up to 1.3 MW capacity, 1.8 MW option supporting up to 8 racks
Liquid-to-Air Option
180 kW or 240 kW, no facility water required
Networking
4x NVLink Switch ports per compute tray, rear-accessible
Deployment Scope
Scalable to 30 racks per SuperCluster unit
Overview
The Supermicro SRS-GB200-NVL72-M1 is the 72-GPU scalable unit of the GB200 NVL72 SuperCluster. Eighteen 1U compute trays each hold two GB200 Grace Blackwell superchips - two 72-core Grace CPUs and four B200 GPUs - for 372 GB of HBM3e per superchip and 744 GB per tray, all tied together by nine NVLink switch trays at 1.8 TB/s per GPU.
Power comes from eight 1U 33 kW shelves, each built from six 5.5 kW PSUs with an integrated capacitor bank to ride through inference transients, bringing total rack input to 132 kW with an operating window of 125 to 135 kW. That density is what forces the liquid cooling design.
Supermicro offers three cooling topologies from the same rack: an in-rack 4U CDU rated to 250 kW with redundant PSUs and dual hot-swap pumps, an in-row CDU up to 1.3 MW serving up to eight racks, or a liquid-to-air variant at 180 or 240 kW for sites with no chilled water or cooling tower.
Key Benefits
Capacitor-buffered power shelves: 33 kW shelves with built-in capacitance absorb GPU transient spikes that would otherwise trip upstream protection. Three cooling paths, one rack: in-rack, in-row or liquid-to-air options let the same SKU land in a chilled-water hall or an air-cooled retrofit. Rack-level scalability: configurations extend to 30 NVL72 racks within a single SuperCluster deployment.
Applications
Large-model pre-training and fine-tuning, mixture-of-experts inference, scientific HPC at exascale, enterprise AI factory builds, and clusters that must scale from a single NVL72 rack into a multi-rack SuperCluster.
Request a Quote — Supermicro SRS-GB200-NVL72-M1 — 72-GPU GB200 NVL72 Scalable Unit
QS Compute — global B2B supply of AI computing hardware, edge AI systems and accelerators. Volume pricing, 15-day sample lead time.
Get Your Quote →