DGX B300 vs DGX B200 vs HGX B300 servers: what differs and which to choose

DGX B300 vs DGX B200: what differs

The DGX B300 and the DGX B200 are both NVIDIA systems with eight SXM GPUs, two NVLink switches and 14.4 TB/s of aggregate NVLink bandwidth in 10U. In a B300 vs B200 comparison, each B300 (Blackwell Ultra) carries 288 GB of HBM3e against 180 GB per GPU in the DGX B200, delivers 50 per cent more dense FP4 throughput and, on NVIDIA’s rating, twice the attention throughput. The DGX B300 moves the cluster network from ConnectX-7 at 400 Gb/s to ConnectX-8 at 800 Gb/s per GPU, and its GPUs give up most of the FP64 and INT8 throughput. HGX B300 vs DGX B300 is a separate question: NVIDIA prints the same GPU figures for both, but an HGX B300 server is built by a server manufacturer, with its own processors, memory, cooling and support.

Per eight-GPU system unless marked per GPU. NVIDIA HGX platform page (updated 30 September 2026), DGX B200 and DGX B300 product pages (updated 16 and 15 October 2026), DGX B300 user guide (5 October 2026); attention and bandwidth per GPU from NVIDIA’s Blackwell Ultra technical blog (22 August 2025) and the DGX B300 datasheet (February 2026); compute capability from NVIDIA’s CUDA GPU list, read 16 October 2026.

B300 vs B200 performance: dense FP4, FP8 and attention

NVIDIA prints most Tensor Core figures with sparsity, and its footnote says “Dense is ½ sparse spec shown.” For FP4 the HGX table gives both figures, and the sparse one is the same for both systems, 144 PFLOPS. The dense figure differs: 108 PFLOPS for eight B300 against 72 for eight B200, or 13.5 and 9 PFLOPS per GPU by our division. NVIDIA’s Blackwell Ultra blog gives the chips 15 and 10 PFLOPS of dense NVFP4, the same ratio of 1.5. NVFP4 checkpoints do not use the 2:4 structured sparsity behind the sparse figure, so compare the dense column, as our TFLOPS comparison of dense and sparse figures explains.

FP8 and BF16 are unchanged, at 72 and 36 PFLOPS of sparse throughput per system. A model served in FP8, or a training run in BF16 or FP8, gains nothing from the B300’s tensor rates on paper. NVIDIA lists up to 8 TB/s of memory bandwidth per GPU on both, so the ceiling for one user’s token rate, bandwidth divided by the bytes read per token, is about the same for a model that fits either GPU. The B300 gains where FP4 compute, attention or memory capacity limit the work: prompt processing and large batches in NVFP4, long contexts, and more concurrent sessions per GPU.

For attention, NVIDIA’s blog lists the special function units at 10.7 TeraExponentials per second on Blackwell Ultra against 5 on Blackwell, “delivering up to 2x faster attention-layer” compute. Attention grows with the context, so the gain is largest for long prompts, RAG requests with many retrieved passages and reasoning models that write long answers.

The B300 gives up double precision and INT8. NVIDIA’s HGX table lists 10 TFLOPS of FP64 per eight-GPU system against 296 for the B200, and 3 POPS of INT8 against 72. A simulation, CFD or quantum chemistry code that needs FP64 runs far slower on the newer system. Models quantised to INT8 weights and activations should be moved to FP8 on the B300, which runs at the same rate as on the B200.

DGX B200 vs DGX B300 specs: power, cooling, processors and network

On NVIDIA’s figures the two DGX models draw similar power and differ in network, weight, power distribution and inlet limit.

NVIDIA DGX B200 user guide (updated 8 September 2026) and product page (16 October 2026); DGX B300 user guide (updated 5 October 2026), product page (15 October 2026), datasheet (February 2026) and “Data Center Best Practices with DGX B300” (version 1.0, 27 February 2026).

NVIDIA offers the DGX B300 in AC/PDU and DC/busbar versions, and its product page says it is “deployable in NVIDIA MGX racks for the first time.” The lower inlet limit matters when both generations share a room, since supply air between 30 and 35 °C suits the DGX B200 but not the DGX B300. NVIDIA’s data-centre guide sizes a rack with two AC-powered DGX B300 at 39.4 kW peak and 30 kW average, so plan the power feed to the peak rather than to the nameplate figures in the table. The cluster network doubles per port, so switches, optics and cables have to be planned for 800 Gb/s. NVIDIA’s DGX SuperPOD reference architecture for the DGX B300 with AC power, updated on 2 September 2026, uses Quantum-X800 InfiniBand switching and scalable units of 72 systems. The room, floor load and power distribution for one or two systems are covered in our DGX B300 specifications and power requirements.

HGX B300 vs DGX B300: NVIDIA’s system or a manufacturer’s server

NVIDIA describes HGX as “a single baseboard with eight” Rubin, Blackwell or Blackwell Ultra SXM GPUs, and server manufacturers build their HGX B300 servers around it. The GPU figures in NVIDIA’s HGX table are the same as the DGX B300’s, and the table adds that “CPU and memory specs are defined by OEM offerings.” When NVIDIA announced Blackwell Ultra on 18 March 2025, it named Cisco, Dell Technologies, Hewlett Packard Enterprise, Lenovo and Supermicro among the companies expected to deliver servers based on it. Their HGX B300 servers differ from the DGX B300 and from each other in processors, memory, chassis height and cooling.

Supermicro’s 8U air-cooled HGX B300 server, for example, takes two AMD EPYC 9005 processors and up to 6 TB of DDR5. Its 4U version is cooled by direct liquid cooling, which Supermicro says captures up to 98 per cent of the system’s heat, and fits up to 64 GPUs in a standard 19-inch rack (press release of 9 December 2025).

NVIDIA DGX B300 product page (15 October 2026) and HGX platform page (30 September 2026); Supermicro AS-8126GS-NB3RT product page, read 16 October 2026, and press release of 9 December 2025.

Choose the DGX B300 when one company should answer for the whole system, when you plan to grow along NVIDIA’s DGX BasePOD or SuperPOD reference architectures, and when you want Mission Control for cluster operations. Choose an HGX B300 server when you need AMD processors, more than 4 TB of system memory, direct liquid cooling or a lower chassis, or when your other servers already run under one manufacturer’s management tools and support contract. NVIDIA AI Enterprise is licensed per GPU on either, eight licences per system, and the DGX B300 page lists it as recommended rather than included. Our guide to buying a DGX B300 in the EU covers the order itself.

We supply the DGX B300 and HGX B300 servers per project. Before we can quote, the distribution asks us for the company, the project and the intended use, so send us those three and we check price and availability for your project.

Which to choose for inference, training, HPC and mixed fleets

For large-model inference with long contexts, the B300’s memory decides. Llama 3.1 405B in FP8 takes about 405 GB of weights at one byte per parameter. On a DGX B200 that leaves about 1 TB for KV cache before any reserve, on a DGX B300 about 1.7 TB against the 2.1 TB NVIDIA lists, so the B300 holds about 60 per cent more concurrent long sessions. A 70B model in BF16, about 141 GB, fits one B200 GPU with under 40 GB to spare and one B300 GPU with more than 100 GB, which changes how many conversations each copy serves.

For training and fine-tuning in BF16 or FP8 the tensor rates are equal, so the B300’s advantage is memory per GPU, for larger batches or fewer GPUs per model shard, and 800 Gb/s per GPU between nodes. For HPC codes that need FP64, the B200 has about 30 times the B300’s rate on NVIDIA’s figures, 296 against 10 TFLOPS per system, and the H200 NVL lists 30 TFLOPS per card.

Mixed fleets work when jobs are placed by what each generation does well: FP64 and INT8 work on the B200 nodes, long-context and NVFP4 inference on the B300. NVIDIA’s CUDA list gives the B200 compute capability 10.0 and the B300 10.3, so check that the framework and container versions you run support the B300 before both share one scheduler queue.

When a PCIe server with H200 NVL or RTX PRO 6000 fits better

An eight-GPU SXM system is sized for models that must be split across GPUs at low latency and for training that exchanges gradients every step. When each model copy fits one 96 GB RTX PRO 6000 Server Edition or one 141 GB H200 NVL, a PCIe server runs one copy per card, draws 4.8 kW of GPU power with eight 600 W cards and stands in an ordinary server room. Our DGX B300 vs eight RTX PRO 6000 comparison sets the two side by side, and the RTX PRO 6000 vs H200 NVL comparison covers the choice between the two cards. The H200 NVL also suits FP64 work and bridges 2 or 4 cards over NVLink, and the RTX PRO 6000 adds graphics and virtual workstations, which the B300 lacks.

We build AI servers to order with four or eight of these cards and check the rack, power and airflow before we quote. Describe your models, their precision and your user count for a configuration and quote within one business day.

What we supply

We supply the DGX B300 and HGX B300 servers per project, on one EU contract and invoice. Before we can quote, the distribution asks us for the company, the project and the intended use, and approvals take longer than an order for cards. For workloads that fit PCIe cards, we build AI servers to order with H200 NVL and RTX PRO 6000 Server Edition cards, assembled and burn-in tested, with manufacturer warranty on every component. The cards themselves are on our professional GPU line-up, and NVIDIA AI Enterprise licences come on the same invoice.

FAQ