Colocation sites for B300 server deployments.

What arrives: a 10U server, in two versions

NVIDIA’s user guide gives the DGX B300 as 10 rack units, 442 mm high, 482.6 mm wide and 904.2 mm deep, with inlet air of 10 to 35 °C. It ships in two versions: an AC version with 12 power supplies of 3.2 kW and 12 inlets at 200 to 240 V, up to 16 A each, and a DC version fed from a 54 V DC bus bar at up to 300 A. NVIDIA lists 168 kg for the AC version and 123 kg for the DC version; packaged, 207.6 kg and 162.6 kg.

NVIDIA’s safety guidance applies to the rack it goes in: anchor the rack to an unmovable support so it cannot tip when a server is extended, route cables so they do not block airflow, and use mechanical assistance when moving and lifting.

HGX B300 servers: same GPUs, the vendor’s spec sheet

NVIDIA sells these GPUs two ways. DGX B300 is NVIDIA’s own complete server, with one spec sheet, NVIDIA’s, and every figure in this guide comes from it. HGX B300 is NVIDIA’s GPU baseboard: eight Blackwell Ultra SXM GPUs on one board, joined by fifth-generation NVLink. Server makers build their own systems around that board, many of them on NVIDIA’s MGX modular reference design, and NVIDIA says the CPU and memory specifications are defined by the OEM offerings. HGX B300 is shipping now.

What NVIDIA fixes

Eight Blackwell Ultra SXM GPUs, 2.1 TB of total GPU memory, fifth-generation NVLink at 1.8 TB/s per GPU and 14.4 TB/s across the board, and 1.6 TB/s of networking bandwidth. This part is the same in every HGX B300 server and in DGX B300.

What the vendor sets

The CPUs and system memory, the chassis height, the power supplies and their count, airflow, inlet temperature, weight, the supported racks, and whether the server is air-cooled or liquid-cooled. These are what decide the site requirement, so take them from the vendor’s own spec sheet, not from this guide.

Using this guide for HGX

DGX B300 is the reference point: an eight-GPU air-cooled server of about 15 kW in 10U. Expect an HGX B300 server from any vendor to be in the same class, check each of its figures against the DGX B300 figures here, and then apply the same rack patterns, circuits and cooling plant.

• Ask the vendor for rated and peak system power, the number of power inlets and their voltage, and whether a DC bus bar version exists.
• Ask for airflow in CFM, the maximum inlet temperature, the altitude derating, and whether a liquid-cooled variant is offered.
• Ask for the system and packaged weight, the height in rack units, and which racks it is qualified for.

Cooling: air, assisted at four per rack

DGX B300 is air-cooled; there is no liquid-cooled version of the DGX system. NVIDIA’s user guide lists 1,500 CFM of airflow per system with power supplies, or 1,350 CFM with a bus bar, at 70% fan speed, and its best-practices guide says each system needs up to approximately 2,145 CFM of supply air at sea level. The specification allows inlet air of 10 to 35 °C; the best-practices guide recommends ASHRAE Class A2 supply air with a maximum of 30 °C, MERV 13 filtration, and an ISO 14644-1 Class 8 rack area. Heat output is about 49,500 BTU per hour per system at most.

NVIDIA’s two patterns differ in how that air is handled. The low-density pattern, two systems per rack, uses unassisted air cooling: about 4,200 CFM per rack with an 18 °C rise across the rack. The high-density pattern, four systems per rack, uses air cooling with an active rear-door heat exchanger: about 6,200 CFM per rack with a 20.6 °C rise at a 25 °C supply. NVIDIA recommends active rear-door heat exchangers for high-density deployments and says passive ones are not recommended for DGX B300.

Air

Hot-aisle containment, blanking panels, brush grommets and enough cold air at the rack face for every server. Ask for the supply temperature the operator guarantees at the rack, not the room, and whether it can hold 30 °C or below.

Rear-door heat exchangers

A rear-door unit is a liquid device: it needs a CDU, facility and technology water loops, leak detection, its own power, floor-load checks, a containment pitch that lets the door open fully, and coolant (PG25) handling and spill procedures. NVIDIA does not specify a model; the supplier has to confirm heat load and airflow.

Altitude

NVIDIA rates the system to 3,048 m, with the allowed inlet temperature falling as altitude rises: 30 °C up to about 300 m, 25 °C to about 1,500 m, 20 °C to 3,048 m.

Rack density: 16 or 32 GPUs per rack

Each server is 8 GPUs and 10U. NVIDIA’s patterns are two per rack, 16 GPUs, in a traditional 48U or 52U rack, or four per rack, 32 GPUs, in a 48U MGX-style rack with an integrated rear DC bus bar and power shelves. NVIDIA’s DGX SuperPOD reference architecture confirms up to four DGX B300 systems per rack where the data centre has the power and cooling for it.

NVIDIA gives the rack for both patterns as 48U, 2,236 mm high, 600 mm wide and 1,068 mm deep, before the rear-door heat exchanger. It estimates the four-server rack at 1,122 kg, again without the heat exchanger. A scalable unit is 72 systems on Quantum-X800 InfiniBand or 64 on Spectrum-X Ethernet, and a full DGX SuperPOD of eight units is 576 systems and 4,608 GPUs.

• Two per rack is the density an ordinary hall can offer; four per rack is a density the hall has to be built for.
• Four servers use 40U; the power shelves, PDUs and switches take the rest, so the rack is full.
• NVIDIA’s layout for a 64-system unit is two rows: the DGX racks, the network, management and storage racks, the CDUs and the rear-door units, all placed within the cable limits of the compute fabric.

Power: 15 kW a server, 30 or 58 kW a rack

NVIDIA estimates system power at 15 kW for the AC version and 14.5 kW for the DC version, with estimated peaks of 19.7 kW and 19 kW. Its rack figures follow: the two-server AC rack draws 30 kW on average and 39.4 kW at peak; the four-server DC rack draws 58 kW on average and 76 kW at peak. Commit the peak, not the average.

Two per rack (AC)

AC rack PDUs, 2N because each server has redundant power supplies. NVIDIA’s provisioning: one 60 A three-phase circuit per utility source at 400 V minimum, which gives a rack capacity of about 40 to 44 kVA depending on voltage.

Four per rack (DC)

Four 1U 33 kW DC power shelves on the rack’s bus bar, N+1, since four systems need at least three active shelves, and 2N circuit inputs to each shelf. NVIDIA’s provisioning: four 60 A three-phase circuits per utility source at 400 V minimum, which gives about 139 to 176 kVA per rack depending on voltage and breaker derating.

Breaker derating

In the United States and Canada the usable capacity of a circuit is 80% of the breaker rating, so NVIDIA’s tables show the same rack at 141 kVA on a derated 400 V, 63 A supply and 176 kVA without derating.

Whole unit

NVIDIA’s estimate for a typical 64-system high-density unit is 975 kVA average and 1,450 kVA peak, including the network, management and storage racks, two CDUs and 24 rear-door units.

Infrastructure: distribution, networking, plant and facility

Power distribution

Three-phase circuits at 400 V or more, one per rack for two servers and four per rack for four servers, from each of two utility sources. Each AC server needs 12 outlets, so count PDU outlets and breaker capacity per rack, not just kW. The DC pattern replaces PDUs with power shelves on a bus bar.

Networking

Eight OSFP ports per server serving eight ConnectX-8 adapters at up to 800 Gb/s, on InfiniBand or Ethernet, plus two dual-port BlueField-3 DPUs for storage and management. NVIDIA’s compute fabric runs on MPO-12 APC single-mode fibre into twin-port OSFP transceivers, with cable runs of 500 m at most and kept under 90% of the limit; management uses Cat6 copper up to 30 m. Cabinet spacing is governed by those cable limits. The fabric wants overhead cable trays sized for fill and weight.

Cooling plant

For two per rack, the hall’s own air: containment, filtration and audited airflow. For four per rack, add CDUs, technology and facility water loops, rear-door heat exchangers with their power and leak detection, and PG25 coolant storage, spill kits and disposal.

Facility

NVIDIA’s site checklist: a loading dock with hydraulic levellers, secure staging space to uncrate, a pathway with ramps, door sizes, lifts and turn radii that a 200 kg packaged server and a 1,100 kg rack can travel, aisle containment structures, cable management structures and seismic bracing where required. Confirm the floor rating for the rack you choose.

People

NVIDIA stresses that power, cooling and space are finite and interrelated, and that facility operations, IT operations, the systems operations centre, the application owner and network operations all have to plan together. Ask the operator who on its side owns each.

Deployment: existing hall or purpose-built?

NVIDIA’s two patterns map onto the two answers. The low-density pattern, two AC servers per rack with traditional PDUs and unassisted air cooling, is designed for existing halls: a 48U rack, one 60 A three-phase circuit per source, 30 to 40 kW and about 4,200 CFM per rack, inlet air at 30 °C or below. Many existing air-cooled colocation halls can offer that today, and NVIDIA states the system can be deployed in a colocation facility.

The high-density pattern, four DC servers in an MGX-style rack, is a different site: a DC bus bar and power shelves instead of PDUs, four 60 A three-phase circuits per source, 58 to 76 kW per rack, rear-door heat exchangers with CDUs and water loops, and coolant handling. That is purpose-built infrastructure, whether in a new hall or a retrofitted one, and the schedule for the loops and the power decides your delivery date.

• Growing a few servers at a time: start at two per rack in an existing hall, and ask for expansion rights to adjacent racks with lead times in writing.
• Building a cluster: a scalable unit of 72 systems is 36 high-density racks or 36 pairs, plus network, management and storage racks, within the fabric’s cable limits.
• Either way, ask for remote hands who have swapped a 168 kg GPU server before, and a dock and route that can take it.

B300 servers or GB300 NVL72?

The same Blackwell Ultra silicon also arrives as the GB300 NVL72 rack: 72 GPUs in one liquid-cooled unit. If your plan is a large, dense cluster in a liquid-cooled hall, read the GB300 NVL72 guide.