DGX B300 or a server with eight RTX PRO 6000: when you need NVLink and HBM, and when you do not

Two machines that look alike on a slide

Both are eight-GPU Blackwell boxes from NVIDIA’s catalogue. One is a system, the other a reference design that Dell, HPE, Lenovo, Cisco and others build. The similarities end at the count.

NVIDIA DGX B300 user guide and product page, HGX B300 page, RTX PRO AI Factory reference architecture, September 2026. The B300 carries 288 GB of HBM3e physically; NVIDIA’s MIG table and its own MLPerf entry describe the exposed capacity as 270 GB, so size models against 270.

Two footnotes an engineer will want. NVIDIA quotes DGX and HGX tensor figures in sparse form and prints the dense value beside it; the RTX PRO 6000 page prints 4 PFLOPS FP4 with no basis stated, so treat the per-card figure as sparse until NVIDIA says otherwise. And FP64 is not the B300’s job: the HGX B300 baseboard is rated at 10 TFLOPS against 296 for the B200, so a CAE code that needs double precision should not assume the newer box is faster.

The memory arithmetic that decides most cases

Weights take two bytes per parameter in FP16, one in FP8 and roughly half in FP4; the KV cache for open conversations comes on top, and the VRAM guide has the formula. Run three models through both machines.

Weights only, computed; parameter counts per the model cards. Nobody submitted an eight-card RTX PRO 6000 result for the 405B or DeepSeek-R1 benchmarks in MLPerf Inference v6.0.

The first two rows are the PCIe server’s case: one model copy per card, data-parallel, the interconnect irrelevant. The last two are the DGX’s case: the model has to be split across GPUs, and every token then crosses the link between them. Tensor parallelism inserts an all-reduce per layer; across NVLink at 1.8 TB/s that is a rounding error, across PCIe Gen5 x16 at 128 GB/s it is the bottleneck. Pipeline parallelism, which only sends activations at stage boundaries, and data parallelism tolerate PCIe; tensor parallelism does not.

What MLPerf Inference v6.0 measured

MLCommons’ closed division is the only place both machines appear under the same rules, and the results are public in the MLCommons repository. NVIDIA submitted the DGX B300; Dell, HPE and Lenovo submitted eight-card RTX PRO 6000 systems.

MLPerf Inference v6.0, closed division, available systems; per-node figures. The Whisper row compares two Lenovo systems. The RTX PRO 6000 entries run FP4, the B300 entries FP4 as well; no power measurements were submitted, so there is no measured tokens-per-watt comparison.

Two readings. Per node, the DGX is three to four times faster on 70B-class inference and about seven times on the 120B mixture-of-experts model, where 8 TB/s of HBM and a 15 PFLOPS FP4 GPU dominate even though the model fits a single RTX PRO card. Per GPU, that is 14,119 against 3,738 tokens per second on Llama 2 70B. The PCIe server is not a slower DGX; it is a different class of machine that happens to hold the same number of GPUs, and the H200 NVL comparison shows the same bandwidth logic one generation down.

Training makes the gap wider. In MLPerf Training v6.0 an eight-card RTX PRO 6000 system pre-trained the Llama 3.1 8B benchmark in roughly 375 minutes; eight B300 in an OEM HGX node took about 72 to 75, five times faster, and no PCIe RTX PRO system submitted the 70B LoRA benchmark at all. Training of 405B-class models appears in the results only on hundreds to thousands of NVL72 GPUs, which is outside both boxes.

The facility side

A DGX B300 is a 10U, 168 kg system in its AC version that draws up to 14.5 kW from twelve 3.2 kW supplies in N+N, wants inlet air no warmer than 30 °C, and moves 1,500 CFM at 70 per cent fan speed. NVIDIA’s best-practice document puts two of them in a rack on AC power distribution at 39.4 kW peak, or four on a 54 V busbar at 76 kW peak, and recommends active rear-door heat exchangers while advising against passive ones. That is rear-door territory from the first rack.

An eight-card PCIe server is a 4U Dell XE7745 at 68.5 kg or a 3U Lenovo SR675 V3 at under 40 kg, with 4.8 kW of GPU power plus processors, fans and memory, on ordinary 200 to 240 V feeds and ordinary rack cooling. Four of them approach 50 kW in one rack, which is where air stops, but one or two of them go into a standard row without a facilities project. The ratio to remember is roughly one DGX to three PCIe servers in power and weight, for three to seven times the inference throughput.

Software, support and the licence line

The DGX ships with DGX OS, Mission Control (Base Command Manager, Run:ai, UFM and NetQ, with an autonomous recovery engine) and three years of Business Standard support from NVIDIA, with advanced replacement of parts on the next business day included and an on-site engineer as an option. The reference architecture for RTX PRO servers is managed with Base Command Manager and Run:ai; Mission Control’s supported-systems list does not include them, and support comes from the OEM.

NVIDIA AI Enterprise is per GPU on both machines: eight licences for eight GPUs, whatever the cards cost. NVIDIA’s licensing guide states that for DGX systems with Blackwell GPUs the licences are purchased separately, while the DGX B300 product page lists the software as included and the SuperPOD architecture says it is included there. The honest instruction for a buyer is to have the quote state the entitlement and its term; our licensing guide explains what it covers.

When the DGX is the right answer

When the model at the precision you need, plus its cache, does not fit one 96 GB card and has to be split across GPUs at interactive latency: 405B-class dense models in FP8, 70B in FP16, the large mixture-of-experts reasoning models. When you fine-tune or pre-train and gradients are exchanged every step. When throughput per rack unit matters more than the price of the facility work, because one DGX replaces three to seven PCIe nodes on the benchmarks above. When you are building towards a BasePOD or SuperPOD and want NVIDIA-direct support and Mission Control operations. And, by NVIDIA’s own positioning, when the workload is beyond the “small and medium model sizes” the RTX PRO reference architecture is written for.

When eight PCIe cards are the better buy

When every model copy fits one card: 70B in FP8, 120B mixture-of-experts in FP4, and every 32B and smaller model that most companies actually run. When the estate is mixed: rendering, video, digital twins, virtual workstations and vGPU alongside inference, which the B300 cannot do at all and the RTX PRO 6000 does natively. When the work needs FP32: 120 TFLOPS per card against about 75 on the B300. When the room is an ordinary server room. When four MIG slices per card and a vGPU plan matter. And when a vendor choice and an OEM support contract are worth more than a single-source system. Our chassis guide covers the lanes, switches and power supplies of these servers.

DGX B200 or B300

The B300 keeps the B200’s FP8 peak, adds half again in dense FP4, doubles attention throughput, carries half again the memory and moves from ConnectX-7 at 400 Gb/s to ConnectX-8 at 800 Gb/s. It gives back FP64 and INT8, tightens the inlet limit from 35 to 30 °C and weighs more in the PSU version. MLPerf v6.0 puts the measured gain at 45 per cent on the 405B benchmark, 23 per cent on DeepSeek-R1 and 8 per cent on Llama 2 70B. For an inference estate the B200 is a sensible buy while it is offered; for the largest models and for training the B300’s memory decides.

What we supply

Eurokommerz supplies NVIDIA DGX systems and eight-card RTX PRO 6000 servers from Dell, HPE, Lenovo and Supermicro across the EU, with the facility checklist attached: feed, weight, inlet temperature, rear-door cooling where the DGX needs it. Tell us the models and the users, and we will say which class of machine the arithmetic points to before anyone opens a catalogue.

FAQ