EnglishCurrent language: English
SearchBriefing

Resources: IT equipment

IT equipment that changes the facility

What a configuration requires from your electrical, thermal and space envelope, from the vendor's own documentation. Each value keeps the configuration and conditions stated by the source. We do not compare performance claims.

Board power for one accelerator is not server or rack power. Rack figures depend on the exact model and revision. Use the engineering pack to turn a rack figure into load and heat for your own assumptions.

Records

  • AI rack system, NVIDIA

    GB300 NVL72 rack

    Rack-scale system with 72 Blackwell Ultra GPUs and 36 Grace CPUs (18 compute trays, 9 NVLink switch trays).

    Cooling
    Liquid cooled rack (MGX architecture) with tray-level and rack-level leak detection
    Full rack power requirement
    up to 142 kW (As stated in NVIDIA's Enterprise Reference Architecture)
    Power shelves
    8 × 33 kW (Each shelf has six 5.5 kW PSUs)
    Compute network per tray
    4 × ConnectX-8, up to 800 Gb/s per SuperNIC (Plus one BlueField-3 B3240 DPU per tray for north-south traffic)
    Dimensions
    not disclosed
    Weight
    not disclosed

    For the facility: Needs a liquid cooling loop to the rack and rack power distribution sized for the stated requirement. In-rack management switches run from the DC busbar.

    Source: NVL72 AI Factory Enterprise Reference Architecture: System Hardware and Components (opens in a new tab) (accessed 18 Oct 2026)

  • AI rack system (reference design value), NVIDIA

    DSX reference design compute cabinet range (MGX Gen 1.1 to Vera Rubin NVL72)

    Cabinet thermal design power range used in NVIDIA's facilities reference design.

    Cooling
    Liquid to liquid CDUs with a technology cooling loop of at least 1.5 L/min per kW in the reference design. CRAHs serve the remaining air-cooled equipment
    Cabinet TDP range
    198 to 330 kW (MGX Gen 1.1 at 198 kW to Vera Rubin NVL72 at 330 kW, reference design values)
    Technology cooling loop design flow
    at least 1.5 L/min per kW (Reference design value)
    Dimensions
    not disclosed
    Weight
    not disclosed

    For the facility: Reference design values for planning conversations, not a product rating. Confirm with the exact system specification.

    Source: DSX Facilities Infrastructure Reference Design Overview (design guide v2.0, 19 August 2026) (opens in a new tab) (accessed 18 Oct 2026)

  • GPU server, NVIDIA

    DGX B200

    Eight Blackwell GPU system with two Intel Xeon Platinum 8570 CPUs.

    Cooling
    Air cooled (NVIDIA's performance notes describe an air-cooled system)
    System power usage
    ~14.3 kW (Maximum, as stated)
    Rack units
    10 RU
    Operating temperature
    10 to 35 °C
    Dimensions (H × W × L)
    444 × 482.2 × 897.1 mm
    Dimensions
    444 × 482.2 × 897.1 mm
    Weight
    not disclosed

    For the facility: Rack power and air cooling must cover the stated maximum per system multiplied by systems per rack.

    Source: NVIDIA DGX B200 product page, specifications (opens in a new tab) (accessed 18 Oct 2026)

  • Accelerator (OAM module), AMD

    Instinct MI355X

    GPU accelerator module for servers, 288 GB HBM3E.

    Cooling
    Listed as passive and active
    Typical board power
    1,400 W (Per GPU module, as stated)
    Form factor
    OAM module
    Dimensions
    not disclosed
    Weight
    not disclosed

    For the facility: Board power is per module only. Server and rack power depend on the OEM system configuration.

    Source: AMD Instinct MI355X GPUs product page (opens in a new tab) (accessed 18 Oct 2026)

  • Accelerator (OAM module), AMD

    Instinct MI350X

    GPU accelerator module for servers, 288 GB HBM3E.

    Cooling
    Passive OAM, as listed
    Typical board power
    1,000 W (Per GPU module, as stated)
    External power
    54V UBB
    Dimensions
    not disclosed
    Weight
    not disclosed

    For the facility: Board power is per module only. Server and rack power depend on the OEM system configuration.

    Source: AMD Instinct MI350X GPUs product page (opens in a new tab) (accessed 18 Oct 2026)

In research, not yet published

These are held back until each value can be tied to a current vendor source with its exact configuration. Listing here is not an endorsement.

  • NVIDIA GB200 NVL72 rack (vendor page did not load during research)
  • OEM rack-scale systems from Dell, HPE, Lenovo and Supermicro
  • Data centre switches and optics (for example Arista, Cisco and NVIDIA Spectrum)
  • AI storage systems
  • Rack power shelves, busbars and manifolds from rack integrators

Updated 18 Oct 2026.