NVIDIA MGX servers explained: what the modular design means for a buyer

What NVIDIA MGX is

NVIDIA MGX is a modular reference architecture that server manufacturers use to build GPU servers from building blocks NVIDIA has defined and validated, from 2U and 4U nodes up to rack-scale systems. NVIDIA calls it “an open modular reference architecture” for OEMs, ODMs and ecosystem partners, and its MGX page, updated on 31 May 2026, counts “over 100 combinations of modular designs”. A manufacturer picks the CPU, x86 or NVIDIA Grace, the PCIe GPUs, for example the RTX PRO 6000 Blackwell Server Edition or the H200 NVL, and the networking, such as ConnectX SuperNICs and BlueField DPUs. HGX is NVIDIA’s other server design, a baseboard with four or eight SXM GPUs joined by NVLink; the HGX B300 has eight.

For a buyer, the MGX label means that the manufacturer started from an NVIDIA layout and component set. The server model, its power supplies, fans, management controller, card list and warranty still come from the manufacturer. When NVIDIA announced MGX at Computex 2023, it listed chassis of “1U, 2U, 4U (air or liquid cooled)” and wrote that “MGX differs from NVIDIA HGX™ in that it offers flexible, multi-generational compatibility with NVIDIA products”.

On the current MGX page, the CPU options are x86 and Arm, NVIDIA Grace and Vera among them, the GPUs are PCIe and rack-scale parts, and the networking includes BlueField DPUs, Spectrum-X Ethernet and Quantum InfiniBand. The page also shows NVIDIA’s NVL72 racks as rack-scale MGX designs.

MGX vs HGX vs a generic PCIe GPU server

The three designs differ in how the GPUs are mounted and connected.

NVIDIA MGX page (31 May 2026), RTX PRO Server page (11 June 2026), RTX PRO 4500 Blackwell Server Edition page (9 July 2026), H200 product page (27 August 2026) and HGX page (30 September 2026), all read on 16 October 2026.

For sizing, the main difference is the GPU-to-GPU link. The RTX PRO 6000 has no NVLink, so eight cards in an MGX server are eight separate 96 GB pools that exchange data over PCIe 5.0 x16, and the server is planned card by card. Our example allocation for an eight-card RTX PRO 6000 server shows how a company of 1,000 staff spreads its models across the cards. An HGX B300 server joins its eight GPUs in one NVLink domain, which suits training and models too large for one card. Our comparison of DGX B300, DGX B200 and HGX B300 covers the HGX side in detail.

RTX PRO Server: the MGX designs in 2U, 4U and 6U

NVIDIA uses the name RTX PRO Server for servers with RTX PRO Blackwell Server Edition cards that its system partners build. Its RTX PRO Server page, updated on 11 June 2026, shows four configurations as “Examples based on NVIDIA MGX™ modular reference designs”.

NVIDIA RTX PRO Server page, updated 11 June 2026, read on 16 October 2026; only the 6U entry is labelled liquid-cooled.

The RTX PRO 4500 Blackwell Server Edition in the second 2U design is a passive, single-slot card of 165 W with 32 GB of GDDR7, per NVIDIA’s product page of 9 July 2026.

The page states that “RTX PRO Servers are NVIDIA-Certified™”. Among the system partners it names Cisco, Dell Technologies, Hewlett Packard Enterprise, Lenovo and Supermicro. Which models of these manufacturers take the card, and how many, is listed in our overview of RTX PRO 6000 Server Edition compatible servers.

NVIDIA’s RTX PRO AI Factory reference architecture, last updated on 18 May 2026, describes an eight-card RTX PRO 6000 server as a “2-8-5-200 infrastructure configuration”. It asks for at least seven physical CPU cores, 128 GB of system memory and one Gen5 x16 link per GPU. Its network is four BlueField-3 SuperNICs at up to 400 Gbps for traffic between servers and one BlueField-3 DPU for north-south traffic. It asks for Gen5 PCIe switches only where more links are needed and prefers direct attachment otherwise.

ConnectX-8 SuperNICs in place of PCIe switches

Eight cards at Gen5 x16 need 128 lanes, which two processors also share with the network adapters and NVMe drives, so most eight-card servers use discrete PCIe switches and a few dual-socket designs attach all eight directly, as our article on how many GPUs fit in one server explains. NVIDIA’s technical blog of 18 May 2025 describes the older layout as two CPUs, eight GPUs and five network adapters, which “also requires two to four discrete PCIe switches”. In the newer MGX layout, each ConnectX-8 SuperNIC integrates “48 lanes of PCIe Gen6 connectivity with an integrated PCIe Gen6 switch”, so the network adapter also serves as the PCIe switch for the GPUs.

With two GPUs per SuperNIC, NVIDIA gives up to 400 Gb/s of network bandwidth per RTX PRO 6000. For traffic between GPUs on different CPU sockets, the blog names “up to 50 GB/s per GPU of IO bandwidth”, against 25 GB/s or less in the older layout, and it claims “up to 2x higher NCCL all-to-all performance”. An inference server that runs one model copy per card sends almost nothing between its cards, while a model split across cards and multi-node work benefit from the newer layout.

Both layouts are sold, so the quote has to name the one you get. Supermicro, which NVIDIA lists as an MGX partner, shows two 4U systems for up to eight double-width GPUs. The SYS-422GL-NR is described as a “4U MGX Dual-Root PCIe GPU System” with PCIe 5.0 x16 switches. The SYS-422GL-FNR2 carries an NVIDIA ConnectX-8 PCIe switch board and eight 400GbE ports. Both pages list the RTX PRO 6000 and RTX PRO 4500 Server Edition and the H200 NVL. Both give four 3,200 W power supplies in a 3+1 configuration and an operating temperature of 10 to 35 °C.

MGX servers with the H200 NVL

The H200 NVL is the PCIe version of the H200, and NVIDIA’s product page, updated on 27 August 2026, gives its server options as “NVIDIA MGX™ H200 NVL partner and NVIDIA-Certified Systems with up to 8 GPUs”. The H200 SXM, by contrast, goes into “NVIDIA HGX™ H200 partner and NVIDIA-Certified Systems™ with 4 or 8 GPUs”. The NVL card is a dual-slot, air-cooled card of up to 600 W, configurable, with 141 GB of memory at 4.8 TB/s. Unlike the RTX PRO 6000, it can join 2 or 4 cards over an NVLink bridge at 900 GB/s per GPU.

NVIDIA’s enterprise reference architecture for the H200 NVL, as NVIDIA’s technical blog of 11 December 2024 describes it, uses “a PCIe Optimized 2-8-5 reference configuration” of two CPUs, eight cards and five network adapters. It places one BlueField-3 SuperNIC for every two H200 NVL cards and a BlueField-3 DPU per server. In an MGX server with H200 NVL cards, ask which cards the bridges join, 2-way or 4-way, and whether the manufacturer lists the bridge for that model; Supermicro lists the NVLink bridge as an option on both 4U systems above.

For an MGX server with H200 NVL cards, the configuration we propose names the bridge layout, and we check the rack, power and airflow before we quote. Send us the cards, how many per server and the power feed, and we propose a configuration.

NVIDIA-Certified Systems: what the label covers

NVIDIA’s page on NVIDIA-Certified Systems, updated on 3 September 2026, says “All systems undergo a rigorous suite of functional, security, and performance tests”. The tests run from single-GPU to multi-node and also cover network and storage offload, TPM security and remote management. NVIDIA adds that “Hardware support for NVIDIA-Certified Systems is available from the system vendors”. The Qualified System Catalog lists the certified models, filtered by category, so check the exact model with your cards there.

An MGX design is where the manufacturer starts, and certification is a separate NVIDIA programme. A server first has to be qualified, which NVIDIA describes as “thermal, mechanical, power, and signal integrity qualification”, before it can be submitted for certification. NVIDIA’s configuration guide for NVIDIA-Certified Systems, updated on 30 September 2026, gives the inference server it describes at least two CPU sockets, six physical cores per GPU and twice the total GPU memory in system RAM. It calls 2, 4 or 8 GPUs per server balanced and places NICs and NVMe drives on the same PCIe switch or root complex as the GPUs. NVIDIA presents these figures as “a starting point for addressing workload-specific needs”. Our guide to what to specify when buying an AI server turns them into a request.

Where MGX servers from different manufacturers differ

The MGX modules fix the layout. Power, cooling, management and support are each manufacturer’s own, and the same card can come with different limits in two servers. Some models cap a slot at 450 W, and some require a particular power supply size whenever the RTX PRO 6000 is configured. Eight cards at 600 W draw 4.8 kW before processors and fans.

Building blocks from NVIDIA’s MGX, RTX PRO Server and configuration guide pages; power supply and temperature examples from Supermicro’s SYS-422GL-NR and SYS-422GL-FNR2 pages; read on 16 October 2026.

Ask for the inlet temperature limit for your exact GPU configuration, and compare it with what the rack position holds in summer. For the warranty, ask whether a card bought outside the manufacturer’s option part numbers is covered.

MGX PCIe server or HGX B300 server

For a company of 1,000 to 2,000 staff serving models that fit 96 or 141 GB, an MGX PCIe server with eight cards runs one model copy per card. Our allocation example serves 1,000 staff from one such server and 2,000 from two. One server has no failover for a host fault, so a service that must run through maintenance windows needs two. The H200 NVL adds NVLink bridges for 2 or 4 cards where a model or its cache needs more than one card.

An HGX B300 server fits training, fine-tuning of large models and inference of models that need eight GPUs in one NVLink domain, with 2.1 TB of GPU memory in total. We supply 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, and approvals take longer than an order for cards.

If your workload needs an HGX B300 server, describe the company, the project and the intended use in the form below, and we check price and availability for your project.

What we supply

We build AI servers to order with 2 to 8 RTX PRO 6000 Server Edition, RTX PRO 4500 Server Edition or H200 NVL cards, assembled and burn-in tested, with manufacturer warranty on every component and delivery anywhere in the EU, on one EU contract and invoice. We check the rack, power and airflow before we quote, and the configuration and quote follow within one business day. We also supply the professional NVIDIA GPUs on their own for servers you already run, after a compatibility check of the platform, power and cooling. NVIDIA AI Enterprise and vGPU licences come on the same invoice.

FAQ