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

DGX B300 and an eight-card RTX PRO 6000 server: specifications

Both are eight-GPU Blackwell configurations from NVIDIA’s catalogue. The DGX B300 is a complete NVIDIA system; the RTX PRO 6000 server is a reference design that Dell, HPE, Lenovo, Cisco and others build. Apart from the GPU count, almost every specification differs.

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; MLPerf® entries name the accelerator B300-SXM-270GB and NVIDIA’s vGPU tables give 269 GB as the largest profile, so size models against about 270.

Two further points matter when reading these figures. 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. FP64 is also much lower on the B300: 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 system is faster.

MIG splits one B300 into up to seven isolated GPU instances. NVIDIA’s MIG user guide, updated 13 October 2026, has no B300 table yet; the profiles are listed in the vGPU reference of NVIDIA AI Enterprise 8.2 for the HGX B300. They are 1g.34gb (up to seven per GPU), 1g.67gb (four), 2g.67gb (three), 3g.135gb (two), 4g.135gb and 7g.269gb (one each), so the eight GPUs of a DGX B300 give up to 56 instances. The RTX PRO 6000 Server Edition splits into up to four instances of 24 GB, 32 in an eight-card server. The full B300 MIG table, with the vGPU profile names, is in our article on DGX B300 specs and power requirements.

Memory: which models fit on each machine

In FP16, weights take two bytes per parameter, in FP8 one byte, and in FP4 half a byte plus block scales, with some layers kept in BF16; NVIDIA’s NVFP4 checkpoint of DeepSeek-R1, for example, is 424 GB. The KV cache for open conversations comes on top, and the VRAM guide has the formula. The table applies this to four models on both machines.

Weights only, computed; parameter counts per the model cards. DGX B300 about 2,160 GB exposed; eight RTX PRO 6000 768 GB, of which the driver shows 764.8 GiB.

In the first two rows the PCIe server is the better fit, because each card holds a full copy of the model, the copies run data-parallel and the interconnect does not matter. In the last two rows the DGX is the better fit, because the model has to be split across GPUs and every token then crosses the link between them. Tensor parallelism inserts two all-reduces per layer in the forward pass; across NVLink at 1.8 TB/s that is a small cost, across PCIe Gen5 x16 at 128 GB/s it is the main one. Pipeline parallelism, which only sends activations at stage boundaries, and data parallelism suit PCIe better; tensor parallelism runs over it at a cost. We supply the DGX B300 per project and build AI servers to order with up to eight RTX PRO 6000 Server Edition cards. Describe your task and our team will help.

What MLPerf® Inference v6.1 measured

MLCommons’ closed division is where both machines appear under the same rules, and the results are public in the MLCommons repository. In MLPerf® Inference v6.1 of 16 September 2026, NVIDIA submitted the DGX B300 (entry 6.1-0071), and other submitters entered systems with eight RTX PRO 6000 Server Edition cards. The table shows Llama 2 70B and gpt-oss-120b, two benchmarks that both machines ran.

MLPerf® Inference v6.1, Datacenter, closed division, available systems, offline scenario; per-node figures. Entries: NVIDIA DGX B300 6.1-0071; eight RTX PRO 6000 Server Edition 6.1-0070 (Llama 2 70B, benchmark llama2-70b-99.9) and 6.1-0082 (gpt-oss-120b). Retrieved from mlcommons.org on 16 October 2026; results verified by MLCommons Association. The ratios are our division of the per-node results, not an MLPerf® metric. The MLPerf name and logo are registered and unregistered trademarks of MLCommons Association in the United States and other countries. All rights reserved. Unauthorized use strictly prohibited. See www.mlcommons.org for more information.

Per node, the DGX is three to four times faster on 70B-class inference and about seven times faster on the 120B mixture-of-experts model, even though that model fits a single RTX PRO card. The difference comes from 8 TB/s of HBM per GPU and 108 PFLOPS of dense FP4 per node in the datasheet, 13.5 PFLOPS per GPU by our division. Per GPU, the Llama 2 70B figures are 13,660 against 3,754 tokens per second, our division of each system result by its eight GPUs, not an MLPerf® metric. The two are different classes of machine that happen to hold the same number of GPUs, and the H200 NVL comparison shows the same bandwidth logic one generation down.

Power, cooling and weight

A DGX B300 is a 10U system weighing 168 kg in its AC version with power supplies and 123 kg in the busbar version. NVIDIA’s user guide gives its power consumption as 14.5 kW, and 15 kW as the system maximum for the 54 V DC busbar input. The datasheet of February 2026 lists 15.1 kW for the AC version and 14.5 kW for the busbar version. The AC version has twelve 3.2 kW power supplies in N+N, of which at least six are required. It needs inlet air no warmer than 30 °C and moves 1,500 CFM at 70 per cent fan speed; NVIDIA’s data-centre guide for the DGX B300 asks for up to about 2,145 CFM of supply air per system. The same guide 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 for high-density racks while advising against passive ones. Even the two-system AC rack averages 30 kW.

An eight-card PCIe server is, for example, a 4U Dell XE7745 at 68.5 kg or a 3U Lenovo SR675 V3 at under 40 kg. It needs 4.8 kW of GPU power plus processors, fans and memory, on ordinary 200 to 240 V feeds and ordinary rack cooling, and one or two of them go into a standard row without a facilities project. Four of them put about 19 kW of GPU power alone into one rack (4 × 4.8 kW), before processors, fans and memory, and at that point containment and a sized power feed have to be planned. In power and weight, one DGX corresponds roughly to two or three PCIe servers, and on the language-model benchmarks above it delivers three to seven times the inference throughput of one.

Software, support and NVIDIA AI Enterprise licences

The DGX includes DGX OS and three years of NVIDIA Business Standard support, with next-business-day advance replacement and an on-site engineer as an option. NVIDIA lists Mission Control (Base Command Manager, Run:ai, UFM and NetQ, with an autonomous recovery engine) as recommended, not included, so it is a separate line in the quote. 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 licensed per GPU on both machines: eight licences for eight GPUs, regardless of the card price. NVIDIA’s licensing guide states that for DGX systems with Blackwell GPUs the licences are purchased separately, while the DGX B300 product page, updated 15 October 2026, lists it as “NVIDIA AI Enterprise (Recommended)” and only DGX OS as “Included”. Have the quote state the entitlement and its term; our licensing guide explains what it covers.

When a DGX B300 is the better choice

A DGX B300 is the better choice 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 and the large mixture-of-experts reasoning models. It is also the better choice when you fine-tune or pre-train and gradients are exchanged every step, and 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 language-model benchmarks above. The same applies if you are building towards a BasePOD or SuperPOD and want NVIDIA-direct support and Mission Control operations, or if, by NVIDIA’s own positioning, the workload is beyond the “small and medium model sizes” the RTX PRO reference architecture is written for.

When an eight-card PCIe server is the better choice

An eight-card RTX PRO 6000 server is the better choice when every model copy fits one card: 70B in FP8, 120B mixture-of-experts in FP4, and every model of 32B or smaller, the size most companies run. It also suits a mixed environment with rendering, video encoding, digital twins and virtual workstations alongside inference: the RTX PRO 6000 handles these natively and the B300 does not; it supports compute vGPU under NVIDIA AI Enterprise, not graphics. The same applies when the work needs FP32 (120 TFLOPS per card against about 75 on the B300), when the servers will stand in an ordinary server room, when four MIG slices per card and a graphics vGPU plan matter, and when a choice of vendor 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.

We build these servers to order with RTX PRO 6000 Server Edition cards. Tell us the models and whether the server also runs rendering or virtual workstations, and we reply within one business day with a configuration and a quote.

DGX B200 or B300

Compared with the B200, the B300 keeps the same FP8 peak, offers 50 per cent more dense FP4 and 288 GB of HBM3e per GPU against 180 GB in the DGX B200, doubles attention throughput and moves from ConnectX-7 at 400 Gb/s to ConnectX-8 at 800 Gb/s. It has lower FP64 and INT8 performance, a lower inlet limit (30 °C instead of 35 °C) and a higher weight in the PSU version. For FP8 inference and for FP64 work the B200 is equal or stronger on NVIDIA’s figures; for the largest models and for training, the B300’s memory decides. We do not supply the DGX B200; our DGX B300 vs DGX B200 comparison has the full table.

What we supply

We supply the RTX PRO 6000 Server Edition as cards and in AI servers built to order with up to eight of them, across the EU, on one EU contract and invoice. Before we quote a server, we check rack, power and airflow. From NVIDIA’s DGX line we supply the DGX Spark, and the DGX B300 per project; before we can quote a DGX B300, the distribution asks us for the company, the project and the intended use, as our guide to buying a DGX B300 in the EU describes. Send us the models, their precision and the number of concurrent users, and we will work out whether each model copy fits one 96 GB card or has to be split across GPUs.

FAQ