NVIDIA RTX PRO Server explained: 2U to 6U designs, the 2-8-5-200 node, certification

What NVIDIA calls an RTX PRO Server

An NVIDIA RTX PRO Server is NVIDIA's name for a GPU server that its system partners build on NVIDIA MGX reference designs, with two to eight NVIDIA RTX PRO 6000 Blackwell Server Edition cards or eight RTX PRO 4500 Blackwell Server Edition cards on PCIe. NVIDIA does not sell the server itself: Cisco, Dell Technologies, Hewlett Packard Enterprise, Lenovo, Supermicro and other makers do, each under its own model name. NVIDIA's RTX PRO Server page, updated on 11 June 2026, states that "RTX PRO Servers are NVIDIA-Certified".

The name sorts servers by their GPU and their design, much as "HGX server" does for servers built around NVIDIA's SXM baseboard. NVIDIA introduced 2U RTX PRO Servers for mainstream data-centre racks on 11 August 2025 and named configurations with two, four or eight GPUs; Dell's PowerEdge R7725 with two RTX PRO 6000 Server Edition was one of the first. The page lists agentic and generative AI, industrial and physical AI, simulation and data analytics, and rendering as workloads.

The card decides most of what such a server can do. The figures below come from NVIDIA's datasheet and product pages, because the specification tables on the RTX PRO Server page itself are not reliable. Both language versions give 800 GB of GPU memory for eight cards; eight cards of 96 GB are 768 GB. The German version, updated on 15 June 2026, also shows a card table with 48 GB of GDDR6, PCIe Gen 4, 300 W and active cooling, which are the figures of an older card, not of the RTX PRO 6000 Blackwell Server Edition.

NVIDIA RTX PRO 6000 Blackwell Server Edition product page and product brief SP-12355-001_v02, read on 24 October 2026; eight-card totals are our arithmetic (8 × 1,597 GB/s = 12,776 GB/s). The liquid-cooled version for the 6U design is single-slot, per NVIDIA's card page.

The memory does not pool. NVIDIA's product brief lists NVLink as "Not supported", so eight cards are eight 96 GB pools that exchange data over PCIe, and a model larger than one card is split with tensor or pipeline parallelism. How much of each card serves model weights and conversations is set out in our guide to eight RTX PRO 6000 Server Edition cards and the LLMs that fit.

The four MGX designs: 2U, 4U and 6U

NVIDIA's RTX PRO Server page shows four configurations as examples of MGX reference designs. Our guide to NVIDIA MGX servers explains MGX itself and how it differs from HGX; here the point is what each design holds.

NVIDIA RTX PRO Server page, updated 11 June 2026, read on 24 October 2026. Wall power is our estimate: cards at full board power plus two processors (2 × AMD EPYC 9135 at 200 W for two cards, 2 × EPYC 9355 at 280 W for eight RTX PRO 4500, 2 × EPYC 9555 at 360 W for eight RTX PRO 6000) plus 15 per cent for the rest of the server.

The 2U design with two cards fits an ordinary rack position. By our estimate it stays below the 2.94 kW we plan per 230 V 16 A feed, as our page on a server with two RTX PRO 6000 Server Edition works out. The second 2U design uses the RTX PRO 4500 Blackwell Server Edition, a passive single-slot card of 165 W with 32 GB of GDDR7 at 800 GB/s, so eight cards fit where only two RTX PRO 6000 Server Edition do. Eight such cards give 256 GB in total and suit many small models or virtual desktops more than one large model; our article on a server with eight RTX PRO 4500 Server Edition covers it.

The 4U design is the eight-card server most buyers mean. Eight cards at 600 W draw 4.8 kW on their own, and our estimate for the server is about 6.3 kW, which needs a three-phase feed; a 230 V 32 A single-phase feed, planned at 5.89 kW, does not carry it. Dell's datasheet for its eight-card servers gives about 8 kW per node as an indicative figure, and the feed is planned on the maker's figure for the exact model. The 6U design carries the liquid-cooled card and needs a coolant loop in the rack, which our article on liquid cooling for GPU racks describes.

We build servers with two, four or eight RTX PRO 6000 Server Edition cards in these form factors; send us your models, user count and rack feed, and our team replies with a configuration and its power per feed.

The 2-8-5-200 reference node

NVIDIA's Enterprise Reference Architecture for the RTX PRO AI Factory, last updated on 18 May 2026, describes clusters of RTX PRO Servers. Its overview calls each node "NVIDIA-Certified RTX PRO Server equipped with eight RTX PRO 6000 Blackwell Server Edition GPUs" and says the design "is based on the 2-8-5-200 infrastructure configuration (2 CPUs, 8 GPUs, 5 NICs at 200 Gbps each)".

NVIDIA RTX PRO AI Factory Enterprise Reference Architecture, overview and components pages, last updated 18 May 2026, read on 24 October 2026.

The four SuperNICs carry traffic between servers and the fifth adapter, the DPU, carries storage, management and user traffic. Four servers form one scalable unit, and the document describes clusters of 16 servers with 128 GPUs and 32 servers with 256 GPUs on Spectrum-X Ethernet.

For one eight-card server the rules add up as follows, by our arithmetic: 56 cores at 7 per GPU, 72 with the 2 per GPU for the operating system or a hypervisor, and 136 if every card is split into four MIG instances. Memory comes to at least 1,024 GB. The scalable unit and the cluster networking matter for companies that plan several servers from the start. A single server for inference on a company's own models needs the per-node rules and one or two network ports to users and storage, not the full five-adapter layout. Our guide to the CPU for a GPU server applies these core counts to AMD EPYC and Intel Xeon processors.

NVIDIA-Certified: what is tested and who supports it

Certification is a separate NVIDIA programme for partner systems. NVIDIA's page on NVIDIA-Certified Systems, updated on 3 September 2026, states that "All systems undergo a rigorous suite of functional, security, and performance tests", from single GPU to multi-node, including network and storage offload, TPM security and remote management, and that "Hardware support for NVIDIA-Certified Systems is available from the system vendors". Certified models are listed in NVIDIA's Qualified System Catalog, by maker, model and GPU.

NVIDIA's configuration guide for NVIDIA-Certified Systems, last updated on 30 September 2026, sets the starting values for an inference server. They differ from the reference architecture in two places.

NVIDIA-Certified Systems Configuration Guide, Table 1, last updated 30 September 2026, and RTX PRO AI Factory RA, components, 18 May 2026, both read on 24 October 2026; eight-card rows are our arithmetic.

The guide also asks for GPUs spread evenly across sockets and PCIe root ports, with "NICs and NVMe drives" placed "within the same PCIe switch or root complex as the GPUs". It adds that "a PCIe switch may be optional for cost-effective inference servers". Our reference configurations on AI servers follow the guide: two AMD EPYC processors and at least twice the GPU memory in system memory, 384 GB for two cards, 768 GB for four and 1,536 GB for eight RTX PRO 6000 Server Edition.

A certified model is a model that passed with a given GPU and count. A server bought as "RTX PRO Server" should appear in the catalog with exactly that card and number of cards, and the warranty for the cards should come with the server.

Switch boards or ConnectX-8 SuperNICs

Eight cards at Gen5 x16 need 128 lanes. Two AMD EPYC 9005 processors provide up to 160 lanes in a two-socket server, by AMD's datasheet, which leaves little for network adapters and drives. Most eight-card servers therefore place PCIe switches between the processors and the cards.

The older layout puts the cards behind discrete PCIe switch boards. NVIDIA's technical blog of 18 May 2025 describes it as two CPUs, eight GPUs and five network adapters, a layout that "also requires two to four discrete PCIe switches". HPE's ProLiant Compute DL380a Gen12 uses a GPU kit with four PCIe switchboards, and Dell's PowerEdge XE7745 four Broadcom switches, by their technical documents.

In the 4U and 6U RTX PRO Server designs, each ConnectX-8 SuperNIC integrates "48 lanes of PCIe Gen6 connectivity with an integrated PCIe Gen6 switch". The network adapter is then also the switch for two GPUs. NVIDIA gives up to 400 Gb/s of network bandwidth per RTX PRO 6000 Server Edition in this layout and up to 50 GB/s per GPU for traffic between GPUs on different CPU sockets, against 25 GB/s or less in the older one, and claims "up to 2x higher NCCL all-to-all performance".

Both layouts are sold under the same 4U description. Supermicro, for example, shows a SYS-422GL-NR with dual-root PCIe 5.0 switches and a SYS-422GL-FNR2 with a ConnectX-8 switch board and eight 400GbE ports, both for eight double-width cards, both with four 3,200 W power supplies in 3+1, as our MGX guide lists.

NVIDIA technical blog of 18 May 2025 and Supermicro system pages as cited in our MGX guide (read 16 October 2026); NVIDIA-Certified Systems Configuration Guide, 30 September 2026.

An inference server that runs one model copy per card sends almost nothing between its cards, and the layout matters little. A model split over four or eight cards with tensor parallelism exchanges data at every layer, and so does training across servers. For those, ask which layout the quote names and check the path between each pair of cards with nvidia-smi topo -m; our guide to P2P, IOMMU and NCCL on RTX PRO 6000 servers explains the settings, and RTX PRO 6000 without NVLink the layouts of a split model.

Which makers' servers qualify

NVIDIA's page names Cisco, Dell Technologies, Hewlett Packard Enterprise, Lenovo and Supermicro among the system partners, with further makers such as Advantech, ASRock Rack, Compal, Eviden, Foxconn, MiTAC Computing, Quanta Cloud Technology, Wistron and Wiwynn. Makers' documents show how far their models go.

NVIDIA press release of 11 August 2025; Lenovo Press LP2263 (28 July 2026); HPE QuickSpecs a00047453enw; Cisco UCS C845A M8 data sheet (24 August 2026); Dell PowerEdge XE7745 Technical Guide Rev. A09 (September 2026); all read on 24 October 2026 or as cited in our compatible-servers list.

The same card comes with different limits in each model. Dell's XE7745 guide gives a maximum ambient temperature of 30 °C for configurations with the RTX PRO 6000 Server Edition and states that the server "does not support user power cap". Lenovo uses the card's 450 W mode to fit four cards into 2U. Our list of RTX PRO 6000 Server Edition compatible servers covers cables, slots and power modes per model.

What the label changes for a buyer

The label tells you that the server follows an NVIDIA layout and, where the model is in the catalog, that NVIDIA tested it with that card. It does not fix power, cooling, management, support or the topology between the cards. Those come from the maker's documents for the exact model, and a quote should name them.

NVIDIA-Certified Systems page (3 September 2026) as quoted in our MGX guide; makers' documents cited above.

Our guide to 2, 4 or 8 RTX PRO 6000 Server Edition per server helps choose the size, and MLPerf® Inference v6.1 results for PCIe GPU servers shows how close eight-card servers from different makers came on the same benchmarks. For a cluster of several servers, the reference architecture's scalable unit of four nodes is a sensible planning step. We check rack, power and airflow before we quote; describe your room and the feeds per rack, and we say which design and card count fit.

What we supply

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. We build AI servers to order with two, four or eight RTX PRO 6000 Server Edition cards, eight RTX PRO 4500 Server Edition, or H200 NVL with NVLink bridges, on platforms whose makers list the card for that slot and power mode. Each server is assembled and tested before delivery, and every component carries the manufacturer warranty on one EU contract and invoice. Quantities and lead time are confirmed in writing with the quote. Operating system, drivers, CUDA and a container runtime are installed on request.

FAQ