NVIDIA DGX B300 specs: GPUs, memory, power consumption, weight, MIG profiles and rack requirements

NVIDIA DGX B300 specifications

The NVIDIA DGX B300 is a 10U system with eight Blackwell Ultra GPUs, 2.1 TB of HBM3e and two NVLink switches with 14.4 TB/s of aggregate NVLink bandwidth. Two Intel Xeon 6776P processors with 2 TB of system memory, configurable to 4 TB, host the GPUs, and eight ConnectX-8 ports at up to 800 Gb/s plus two BlueField-3 DPUs connect the system to the network. It comes in a version with twelve power supplies, at 15.1 kW and 168 kg, and a busbar version at 14.5 kW and 123 kg.

NVIDIA DGX B300 product page (updated 15 October 2026), DGX B300 datasheet 4868000 (February 2026), DGX B300 user guide, “Introduction to NVIDIA DGX B300 System” (updated 5 October 2026).

GPU memory, compute and FP64 on the B300

Each B300 carries 288 GB of HBM3e, and the user guide adds them up to 2.3 TB. The product page and the datasheet print 2.1 TB, and NVIDIA’s vGPU reference gives 269 GB to the largest instance on one GPU, so plan with about 270 GB per GPU and 2.1 TB per system. A 70B model in FP8, about 73 GB of weights, fits one GPU with close to 200 GB left for KV cache and runtime overhead.

The datasheet of February 2026 gives 62 TB/s of HBM3e bandwidth for the system, and NVIDIA’s Blackwell Ultra technical blog of 22 August 2025 gives 8 TB/s per GPU; for a DGX B300, plan with the datasheet’s system figure. The blog also lists 1.8 TB/s of bidirectional NVLink 5 bandwidth per GPU, and the two NVLink switches connect all eight GPUs, so a model split across them stays inside one NVLink domain; how that compares with eight PCIe cards is in DGX B300 vs a server with eight RTX PRO 6000.

NVIDIA prints tensor performance with sparsity, and the FP8 footnote says dense is half the figure shown, so compare dense with dense across systems. For FP4 the datasheet prints both values, 144 PFLOPS sparse and 108 PFLOPS dense, which is 13.5 PFLOPS dense per GPU by our division. The blog’s 15 PFLOPS of dense NVFP4 is its figure for the Blackwell Ultra GPU in general, not a DGX B300 rating.

The HGX B300 column on NVIDIA’s HGX page, updated 30 September 2026, lists FP64 at 10 TFLOPS for eight GPUs against 296 TFLOPS for the HGX B200, and INT8 at 3 POPS against 72 POPS (sparse figures). Plan codes that need double precision for other hardware. The other differences from the B200 are in DGX B300 vs DGX B200 vs HGX B300.

B300 MIG profiles and instance sizes

MIG divides each B300 into up to seven isolated GPU instances, which gives a DGX B300 up to 56. NVIDIA’s MIG user guide, updated 13 October 2026, has neither a B300 table nor the B300 in its list of supported GPUs. The B300 profiles are published in the vGPU reference of NVIDIA AI Enterprise 8.2, in its table of MIG-backed types for the “NVIDIA HGX B300 279 GB”.

NVIDIA AI Enterprise 8.2, vGPU reference “Blackwell”, Table 80 “MIG-Backed NVIDIA vGPU for Compute for NVIDIA HGX B300 279 GB”, updated 2 September 2026.

The number before the “g” is the count of compute slices out of seven. The 1g.67gb profile pairs one compute slice with as much memory as a 2g.67gb, which suits a model that needs memory more than compute. The 1g.34gb+me profile, available once per GPU, adds media extensions, which the MIG user guide defines as “at least one media engine”. All seven types are vGPU for Compute types, and the reference states “Required license edition: NVIDIA AI Enterprise” for them. The DGX B300 product page, updated 15 October 2026, lists DGX OS as “Included” and NVIDIA AI Enterprise as “Recommended”, so have the quote state whether the AI Enterprise entitlement is part of it.

B300 PCIe: what NVIDIA offers

As of October 2026 we found no NVIDIA product page for a B300 as a PCIe card, and neither the vGPU reference nor the MIG guide lists one. The B300 is an SXM module for the eight-GPU HGX B300 baseboard and the DGX B300, and the PCIe Gen6 interface at 256 GB/s bidirectional that the Blackwell Ultra blog lists is the GPU’s host interface, not a card form factor.

For a standard PCIe server, the data-centre cards we supply are the H200 NVL, with 141 GB of HBM3e at 4.8 TB/s, up to seven MIG instances and NVLink bridges across 2 or 4 cards, and the RTX PRO 6000 Blackwell Server Edition, with 96 GB of GDDR7 at 1,597 GB/s and up to four MIG instances. Both draw up to 600 W per card.

We build AI servers to order with 2 to 8 of these cards per node. Tell us the model, its precision and the peak number of users, and we reply with a configuration and quote within one business day.

DGX B300 power consumption: AC and busbar versions

The product page, updated 15 October 2026, says “~14 kW”. The user guide, updated 5 October 2026, gives a power consumption of 14.5 kW and, for the 54 V DC busbar input, 15 kW as the system maximum. Only the datasheet of February 2026 gives one figure per version: 15.1 kW with power supplies and 14.5 kW with the busbar. NVIDIA’s data-centre guide, version 1.0 of 27 February 2026, estimates 15 kW average and 19.7 kW peak for the AC version and 14.5 kW and 19 kW for the busbar version, at 25 °C supply air. Its rack peaks of 39.4 kW for two AC systems and 76 kW for four busbar systems are these system peaks multiplied, so size feeds and cooling to the peak.

The AC version has twelve 3.2 kW power supplies in N+N redundancy at 200 to 240 V. Each inlet draws up to 16 A through a C19/C20 cord of 1,200 mm rated at 250 VAC and 20 A, and the user guide allows only NVIDIA-provided cords. Six supplies are the minimum for operation, although the system powers on with as few as three working. For two AC systems per rack, the guide plans 2N rack PDUs and one 60 A three-phase circuit at 400 V or more per utility source. NVIDIA’s DGX SuperPOD reference architecture for the DGX B300 with AC power, updated 2 September 2026, places four systems in one rack at about 56 kW.

The busbar version takes 54 V DC at up to 300 A, and the user guide states that the power shelves are “not included with the DGX B300”. The guide places four such systems in an MGX-style rack with a rear DC busbar, fed by four 1U shelves of 33 kW in N+1 and four 60 A three-phase circuits at 400 V or more per utility source.

Weight, dimensions and floor loading

The DGX B300 weighs 168 kg with power supplies and 123 kg in the busbar version, which NVIDIA’s data-centre guide gives without the power shelves. Packed, it weighs 207.6 kg or 162.6 kg, in a package of 1,065 × 711 × 1,178 mm (height, width, length).

The data-centre guide’s rack for four busbar systems is 48U, 600 mm wide, 1,068 mm deep and 2,236 mm high, and weighs 1,122 kg without the rear-door heat exchanger. Spread over the rack’s own footprint of about 0.64 m², that is about 1,750 kg per square metre by our division, so compare it with the floor’s rating per square metre and with the load on each castor. The user guide states that the system “must be installed by NVIDIA partner network personnel or NVIDIA field service engineers”. Our GPU rack power and cooling checklist covers the same checks for smaller GPU racks.

Inlet temperature, airflow and rear-door cooling

The user guide, updated 5 October 2026, gives an inlet air temperature of 10 to 30 °C. The datasheet of February 2026 and the data-centre guide of 27 February 2026 print an operating temperature of 10 to 35 °C, but the guide says supply air “should generally comply with ASHRAE Class A2 specifications” and sets a maximum of 30 °C. The guide allows 30 °C only up to 1,000 ft (about 300 m), then 25 °C up to 5,000 ft (about 1,500 m) and 20 °C up to 10,000 ft (3,048 m).

At 70 per cent fan speed the system moves 1,500 CFM in the PSU version and 1,350 CFM in the busbar version. NVIDIA’s guide states that each system “requires up to approximately 2145 CFM of supply air at sea level”, and it gives a cooling demand of 4,200 CFM for a rack with two AC systems and 6,200 CFM for four busbar systems. It recommends active rear-door heat exchangers for high-density racks and says passive ones are not recommended for the DGX B300. Liquid cooling for GPU racks covers rear doors and water temperatures.

Network ports and cabling for the compute fabric

Each DGX B300 has eight OSFP ports, one per ConnectX-8 adapter at up to 800 Gb/s InfiniBand or Ethernet, for the cluster network. Two dual-port QSFP112 BlueField-3 DPUs at up to 400 Gb/s serve the storage and management networks. With a 1 GbE RJ-45 host port and a 1 GbE BMC port, a system can need up to 14 network cables, and the user guide states that it “is not shipped with network cables or adapters”.

For a SuperPOD compute fabric, NVIDIA’s guide names mainly MPO12 APC passive single-mode fibre, cables of 500 m or less and runs under 90 per cent of each media type’s maximum distance.

Site checklist for a DGX B300 rack

NVIDIA DGX B300 user guide (updated 5 October 2026) and “Data Center Best Practices with DGX B300”, version 1.0, 27 February 2026; the “What to check” entries are our checklist.

We supply the DGX B300 per project, and before we can quote, the distribution asks us for the company, the project and the intended use. Send us those three with the figures from this checklist, and we check price and availability for your project.

What we supply

We supply the NVIDIA DGX B300 and HGX B300 servers per project, and approvals take longer than an order for cards; buying a DGX B300 in the EU describes how such an order runs. For PCIe servers we build AI servers to order with H200 NVL or RTX PRO 6000 Server Edition cards, assembled and burn-in tested, with manufacturer warranty on every component. We check the rack, power and airflow before we quote, and NVIDIA AI Enterprise licences come on the same EU contract and invoice as the hardware.

FAQ