AI Compute Pod, Scale-Up (NVLink) vs Scale-Out Fabric, and Structural Floor Loading
Plain-English definition
A tightly coupled group of high-density GPU racks interconnected by ultra-low-latency copper/optical scale-up fabrics within a rack or row (such as NVLink) and lossless RDMA/Ethernet/InfiniBand scale-out fabrics across pods, concentrating extreme kW, coolant flow, and slab weight.
Why it matters on shift
Because scale-up copper backplanes have strict reach limits of just a few meters, operators cannot casually spread 120 kW GPU racks across empty rooms to ease cooling; power, CDU manifolds, and structural slab support must meet the pod where it sits.
Illustrative operational example
In an illustrative eight-rack AI pod totaling 960 kW and over 14 metric tons of populated rack mass, engineers verify structural slab capacity, overhead busways, and an in-row N+1 CDU skid before uncrating hardware.
Scope, code & safety boundary
Raised-floor tiles rated for legacy 6 kW air-cooled cabinets can buckle under concentrated 1,800 kg liquid-cooled GPU racks unless heavy-duty pedestals, load-spreading ramps, or slab-on-grade designs are engineered.
Related courses & interactive lab practice
High-density AI cooling & commissioning
Practice live readings and switching in "NOC, Spine-Fiber & White-Space Walkdown" with no login required.
Primary standards & technical references