The Supermicro NVIDIA GB300 NVL72 Rack
The centerpiece of any discussion about large-scale AI infrastructure today is the NVIDIA GB300 NVL72. This is not a server in the conventional sense.
Supermicro Liquid Cooled Supermicro NVIDIA GB300 NVL72 Rack 16It is an interconnected rack system containing 72 NVIDIA Blackwell Ultra GPUs paired with their Grace CPUs across 18 compute nodes, along with all of the NVLink switching fabric, ConnectX-8 networking, power infrastructure, and liquid cooling required to operate the entire rack as a single unified compute entity.
Power Infrastructure: ORV3 Power Shelves
At the top and bottom of the NVL72 rack are Supermicro power shelves, part of the ORV3 open rack specification. These shelves eliminate the need for individual power supplies inside each compute node or switch tray.
Power is instead distributed via a bus bar at the rear of the rack, and compute nodes slide in to make simultaneous contact with both the power and liquid-cooling connections.
Supermicro designs and manufactures these power shelves. Each power supply unit in this generation is a 5.5-kilowatt module, with higher-capacity versions in development for future platforms.
The NVLink Spine and Compute Nodes
The middle section of the rack is occupied by the NVIDIA NVLink switches and the 18 GB300 compute nodes.
The NVLink spine at the rear of the rack contains the cabling infrastructure that connects the entire GPU array into a single shared memory and compute fabric. This is what allows the 72 GPUs in the rack to operate as one large GPU rather than as 18 independent nodes.
Each GB300 compute node features NVIDIA ConnectX-8 networking, which, in this rack context, is used for scale-out.
When a single NVL72 rack is insufficient for a given workload, the ConnectX-8 interfaces connect those racks via high-speed switches, extending the logical GPU boundary across multiple racks.
NVIDIA Spectrum-4 Switching
Switching to the Ethernet networking handled by the NVIDIA Spectrum-X platform and Spectrum-4 switches.
Each Spectrum-4 switch provides 64 ports of 800Gbps Ethernet, with the ability to split those ports into 2x400Gbps configurations. This aggregates all ConnectX-8 uplinks from the GB300 compute nodes and connects the cluster to external infrastructure or to additional NVL72 racks.
NVIDIA Spectrum-X Ethernet has become the dominant fabric choice for large-scale AI deployments. It implements adaptive routing and congestion control at the network level, along with deep integration into observability tooling and AI-specific software libraries, including NCCL. Spectrum-X is tested, qualified, and certified in conjunction with Supermicro systems.
NVIDIA BlueField-3 DPUs for North-South Traffic
Within the NVL72 rack, the NVIDIA BlueField-3 DPU handles north-south traffic flows, meaning access to storage systems and other infrastructure outside the GPU cluster.
The NVLink fabric and ConnectX-8 handle east-west GPU-to-GPU traffic, while the BlueField-3 DPU manages connections to data pipelines, object storage, and other services that feed the training or inference workloads running on the cluster.
Next, let us get to putting PCIe GPUs in standard servers and the management software before getting more in-depth on the liquid-cooling.
I’m in awe of how much you cover. Small networking to 1.6T optical DSP’s. Small GB10 box to giant Supermicro AI Factory.