Step 5 · Facility
Power, cooling, network, floor
Check the room before ordering: from the team tier up, power and cooling – not the GPU – decide where you can install.
Power by tier
| System | Power | Electrical | Source |
|---|---|---|---|
| Mac Studio | ≤ 480 W | Ordinary socket, near silent | apple.com |
| DGX Spark | 240 W PSU | Ordinary socket | nvidia.com |
| 1 × RTX 5090 | ≥ 1,000 W system PSU | One dedicated 10–16 A circuit is comfortable | nvidia.com |
| 4 × RTX PRO 6000 | 2.4 kW GPUs alone | 220–240 V, dual PSUs; office-room heat (rule of thumb) | nvidia.com |
| 8 × H100 / H200 | 8 × 700 W GPUs + host | Server room / colocation; node total to be verified per OEM | nvidia.comnvidia.com |
| HGX / DGX B300 | ≈ 14.5 kW, 15 kW max per node | 12 × C19/C20 at 200–240 V 16 A (N+N) or 54 VDC busbar | docs.nvidia.com |
| SuperPOD rack (4 × DGX B300) | ≈ 56 kW / rack | Three 2U rack PDUs per rack in the reference design | docs.nvidia.com |
| GB200 / GB300 NVL72 | To be verified | Fully liquid-cooled rack; facility water and CDU | nvidia.comnvidia.com |
Cooling
Air-cooled DGX B300: inlet 10–30 °C, ≈ 1,500 CFM, 49,476 BTU/h of heat per node – practically all input power becomes heat.docs.nvidia.com
Liquid (direct-to-chip) B300 servers need facility water, CDUs and manifolds; Supermicro states complete rack and on-site service are required for its 4U liquid model.supermicro.comdell.combuy.hpe.com
NVL72 racks are fully liquid-cooled.nvidia.com
Network
Desk tier: 10GbE (Mac Studio, DGX Spark); DGX Spark adds ConnectX-7 200 Gb/s for linking units.apple.comnvidia.com
HGX B300: 8 × 800 Gb/s east-west (one ConnectX-8 per GPU) plus BlueField-3 north-south. SuperPOD uses four fabrics – compute, storage, in-band, out-of-band – and needs > 80 GBps storage I/O per node and ≥ 2 × 100GbE uplinks.docs.nvidia.comdocs.nvidia.com
Switch families: Quantum-X800 InfiniBand or Spectrum-X Ethernet.nvidia.comnvidia.comnvidia.com
Space & floor
DGX B300 is 10U, 904 mm deep and 168 kg (PSU version); Supermicro 8U air version 119 kg. Plan deep racks, lifts for installation and floor loading for 4 nodes per rack.docs.nvidia.comsupermicro.comdocs.nvidia.com
Storage
Reference minimums for HGX B300: inference ≥ 1TB NVMe per CPU socket, training ≥ 2TB per socket, plus a 1TB boot drive; clusters add a shared high-speed storage tier and NFS home storage.docs.nvidia.comdocs.nvidia.com