HGX, DGX, MGX: NVIDIA's Server Platforms
NVIDIA offers three main server platforms.
- MGX is a general modular reference architecture for buyers that mix different CPUs and NVIDIA GPUs themselves.
- OEMs (server manufacturers like Dell, Supermicro, and HPE) build servers around HGX, which is NVIDIA's proprietary GPU baseboard bundling their SXM GPU form factor, NVLink high-bandwidth interconnect, and NVSwitch routing chips.
- DGX is NVIDIA's own complete server using the same HGX baseboard, plus a fixed CPU, memory, and support stack chosen by NVIDIA.
Most data center GPU buyers end up with HGX-based OEM servers.
NVIDIA server platform hierarchy (click a box for details)
What HGX actually is
The HGX board holds 8 SXM GPUs, NVIDIA's proprietary GPU form factor. [1]NVIDIA, "HGX Platform" (accessed March 2026)https://www.nvidia.com/en-us/data-center/hgx/ The GPUs are wired together through NVLink and connected by NVSwitch chips so every GPU can communicate with every other at full bandwidth. [2]NVIDIA, "Introducing NVIDIA HGX H100" Technical Blog (2022)https://developer.nvidia.com/blog/introducing-nvidia-hgx-h100-an-accelerated-server-platform-for-ai-and-high-performance-computing/ Without NVSwitch, NVLink connections would be point-to-point between neighbors only.
The baseboard supplies GPU compute and interconnect. Everything else, CPUs, system RAM, storage, NICs (network interface cards), DPUs (data processing units that offload networking and security tasks), chassis, cooling, and baseboard management controller (BMC), is chosen by the OEM that builds around it.
NVIDIA ships HGX boards to OEMs: Dell, HPE, Supermicro, Lenovo, and others. [1]NVIDIA, "HGX Platform" (accessed March 2026)https://www.nvidia.com/en-us/data-center/hgx/ OEMs cannot modify the baseboard itself. GPU count, memory configuration, NVLink wiring, and NVSwitch layout are all fixed by NVIDIA.
The GPUs on the HGX baseboard are the most expensive component in the server, accounting for roughly 60-70% of total hardware cost in a typical 8-GPU deployment. [18]PyTorch to Atoms, "Meta's 24K H100 Cluster: CapEx/TCO and BOM Analysis" (2024)https://pytorchtoatoms.substack.com/p/metas-24k-h100-cluster-capextco-and
What OEMs change
Servers that use PCIe (an older, lower-bandwidth expansion slot) for their GPUs cannot use HGX boards. NVIDIA has shipped both 4-GPU and 8-GPU HGX variants across multiple generations, including A100, H100, and H200. [3]NVIDIA, "HGX A100 Datasheet" (2020)https://www.nvidia.com/content/dam/en-zz/Solutions/Data-Center/HGX/a100-80gb-hgx-a100-datasheet-us-nvidia-1485640-r6-web.pdf [2]NVIDIA, "Introducing NVIDIA HGX H100" Technical Blog (2022)https://developer.nvidia.com/blog/introducing-nvidia-hgx-h100-an-accelerated-server-platform-for-ai-and-high-performance-computing/ [4]NVIDIA, "H200 GPU" (accessed March 2026)https://www.nvidia.com/en-us/data-center/h200/ As of early 2026, the HGX platform page focuses on 8-GPU Blackwell and Rubin systems. [1]NVIDIA, "HGX Platform" (accessed March 2026)https://www.nvidia.com/en-us/data-center/hgx/
| Fixed by NVIDIA | Chosen by the OEM |
|---|---|
| GPU count and model | Chassis height (how tall the server is) |
| NVSwitch chips and NVLink wiring | Cooling (air, liquid, or hybrid) |
| Power delivery to GPUs | Host CPUs (Intel Xeon or AMD EPYC) |
| System RAM (DDR5 capacity and speed) | |
| NICs, DPUs, NVMe storage, BMC firmware |
OEM pricing for a complete 8-GPU HGX server ranged from $200,000-$300,000 for the H100 generation and $250,000-$400,000 for B200/B300-class systems, depending on configuration and vendor. [19]Exxact, "TensorEX TS4-128275176-E 8U HGX B200 Server" (accessed March 2026)https://www.exxactcorp.com/Exxact-TS4-128275176-E128275176 The spread reflects OEM choices around cooling, storage, NICs, and support tiers rather than the GPU baseboard itself.
| OEM | Recent HGX system | CPUs | Cooling |
|---|---|---|---|
| Dell | PowerEdge XE9680 | 2x Intel Xeon (up to 64 cores each) | Air |
| Supermicro | SYS-821GE-TNHR | 2x Intel Xeon (up to 64 cores each) | Air |
| Lenovo | ThinkSystem SR675 V3 | 2x AMD EPYC (up to 128 cores each) | Neptune hybrid |
Sources: Dell [5]Dell, "PowerEdge XE9680 Spec Sheet" (accessed March 2026)https://www.delltechnologies.com/asset/en-us/products/servers/technical-support/poweredge-xe9680-spec-sheet.pdf, Supermicro [6]Supermicro, "SYS-821GE-TNHR Product Page" (accessed March 2026)https://www.supermicro.com/en/products/system/gpu/8u/sys-821ge-tnhr, Lenovo [7]Lenovo, "ThinkSystem SR675 V3 Technical Specifications" (accessed March 2026)https://pubs.lenovo.com/sr675-v3/server_specifications_technical
Dell and Supermicro are the most common HGX OEMs. The main differentiators between vendors are chassis height (measured in rack units, where a 3U server is three slots tall and an 8U is eight), cooling design, CPU vendor (Intel Xeon or AMD EPYC), NIC and DPU selection, firmware qualification, and support tiers.
Cooling matters most as power rises. H100 SXM GPUs draw up to 700W each, [2]NVIDIA, "Introducing NVIDIA HGX H100" Technical Blog (2022)https://developer.nvidia.com/blog/introducing-nvidia-hgx-h100-an-accelerated-server-platform-for-ai-and-high-performance-computing/ which most OEMs handle with air cooling. Blackwell raises that to up to 1,000W per GPU, [8]NVIDIA, "DGX B200 User Guide" (2025)https://docs.nvidia.com/dgx/dgxb200-user-guide/introduction-to-dgxb200.html pushing many new designs toward liquid or hybrid cooling with rack-level coolant distribution units (CDUs).
What DGX is
DGX is NVIDIA's own complete server. It uses the same HGX baseboard as OEM servers but adds a fixed set of CPUs, memory, networking, storage, software, and support. [9]NVIDIA, "DGX Platform" (accessed March 2026)https://www.nvidia.com/en-us/data-center/dgx-platform/
The DGX B200 ships with 8 Blackwell GPUs, 1.4TB total GPU memory, and 72 PFLOPS (petaflops, or quadrillions of floating-point operations per second) of FP8 (8-bit floating point, a reduced-precision format used to speed up AI training) performance with sparsity (a technique that skips zero-value computations for higher throughput). [8]NVIDIA, "DGX B200 User Guide" (2025)https://docs.nvidia.com/dgx/dgxb200-user-guide/introduction-to-dgxb200.html
| Component | DGX B200 |
|---|---|
| GPUs | 8x Blackwell SXM |
| GPU memory | 1,440 GB (1.4 TB) |
| Training (FP8) | 72 PFLOPS (with sparsity) |
| FP4 Tensor Core | 108 PFLOPS (with sparsity) |
| CPUs | 2x Intel Xeon 8570 (56 cores each) |
| System memory | 2 TB DDR5 (expandable to 4 TB) |
| Storage | 2x 1.92 TB NVMe M.2 (OS) + 8x 3.84 TB NVMe U.2 (data) |
| NICs | 8x ConnectX-7 (400 Gbps each) |
| DPUs | 2x BlueField-3 |
| Form factor | 10U |
| Power supplies | 6x 3.3 kW (5+1 redundancy) |
Sources: NVIDIA DGX B200 User Guide [8]NVIDIA, "DGX B200 User Guide" (2025)https://docs.nvidia.com/dgx/dgxb200-user-guide/introduction-to-dgxb200.html, HGX platform page [1]NVIDIA, "HGX Platform" (accessed March 2026)https://www.nvidia.com/en-us/data-center/hgx/
The DGX B300 uses the B300 NVL16 architecture with 2.3TB of HBM3e (High Bandwidth Memory, a type of memory stacked on the GPU for faster data access), 8 ConnectX-8 SuperNICs, and 2 BlueField-3 DPUs. [10]NVIDIA Newsroom, "Blackwell Ultra DGX SuperPOD" (March 2025)https://nvidianews.nvidia.com/news/blackwell-ultra-dgx-superpod-supercomputer-ai-factories The HGX B300 baseboard (2.1TB, 8 Blackwell Ultra SXM GPUs) is a different configuration from the DGX B300 system. [1]NVIDIA, "HGX Platform" (accessed March 2026)https://www.nvidia.com/en-us/data-center/hgx/
DGX is a fixed hardware system. Around that hardware, NVIDIA also offers a broader platform of software and support for DGX deployments. [9]NVIDIA, "DGX Platform" (accessed March 2026)https://www.nvidia.com/en-us/data-center/dgx-platform/
- A validated NVIDIA-defined system configuration, including the CPU, NIC, DPU, storage, and firmware stack
- NVIDIA Enterprise Support and direct access to NVIDIA engineers for the validated stack
- Platform software that can be used with DGX deployments, including NVIDIA AI Enterprise, Base Command Manager, and Mission Control
DGX systems always include 2 BlueField DPUs. OEM HGX builds vary in DPU count and model.
NVIDIA has not published list prices for the DGX B200 or B300.
DGX costs more than an equivalent OEM server because it includes NVIDIA's software stack and direct support. Teams that want lower cost per GPU, or already work with a specific OEM, buy OEM servers instead.
GB200 NVL72
The GB200 NVL72 is a full liquid-cooled rack with 72 Blackwell GPUs and 36 Grace CPUs. Unlike HGX, which is a baseboard that goes inside a server, the NVL72 ships as an entire rack. [11]NVIDIA, "GB200 NVL72" (accessed March 2026)https://www.nvidia.com/en-us/data-center/gb200-nvl72/
Inside the rack are 18 compute trays. Each tray holds 2 GB200 Superchips. Each Superchip pairs 2 Blackwell GPUs with 1 Grace CPU through NVLink-C2C, a chip-to-chip coherent interconnect that provides 900 GB/s of bandwidth. [12]NVIDIA, "DGX SuperPOD Reference Architecture: GB200" (June 2025)https://docs.nvidia.com/dgx-superpod/reference-architecture-scalable-infrastructure-gb200/latest/ Nine NVSwitches connect all 72 GPUs into a single NVLink domain, so the rack behaves like one tightly coupled scale-up system.
| Spec | GB200 NVL72 |
|---|---|
| GPUs | 72 Blackwell |
| CPUs | 36 Grace (Arm) |
| Compute trays | 18 |
| NVSwitches | 9 |
| Total NVLink bandwidth | 130 TB/s |
| GPU-to-GPU bandwidth | 1.8 TB/s |
| FP4 performance | 1,440 PFLOPS (with sparsity) |
| FP8 performance | 720 PFLOPS (with sparsity) |
| Networking per tray | 4x ConnectX-7 (400 Gbps) |
| Power | 120-132 kW per rack |
| Cooling | Liquid (mandatory) |
Sources: NVIDIA GB200 NVL72 [11]NVIDIA, "GB200 NVL72" (accessed March 2026)https://www.nvidia.com/en-us/data-center/gb200-nvl72/, DGX SuperPOD Reference Architecture [12]NVIDIA, "DGX SuperPOD Reference Architecture: GB200" (June 2025)https://docs.nvidia.com/dgx-superpod/reference-architecture-scalable-infrastructure-gb200/latest/
NVIDIA sells the NVL72 as the DGX GB200, and OEM partners like Supermicro offer equivalent configurations. The GB300 NVL72, announced in March 2025, is the Blackwell Ultra successor. [10]NVIDIA Newsroom, "Blackwell Ultra DGX SuperPOD" (March 2025)https://nvidianews.nvidia.com/news/blackwell-ultra-dgx-superpod-supercomputer-ai-factories
The NVL72 requires liquid cooling and over 120 kW per rack. Traditional air-cooled data centers cannot run it without retrofitting.
DGX SuperPOD and DGX Cloud
This is the deployment layer above the server itself. DGX SuperPOD and DGX Cloud are not alternatives to HGX, DGX, or MGX base systems. They are ways to deploy DGX-class infrastructure at larger scale or through a cloud provider.
SuperPOD
DGX SuperPOD is NVIDIA's multi-rack deployment architecture for DGX systems and DGX GB200 or GB300 rack systems. It combines those compute systems with validated networking and operations software. [12]NVIDIA, "DGX SuperPOD Reference Architecture: GB200" (June 2025)https://docs.nvidia.com/dgx-superpod/reference-architecture-scalable-infrastructure-gb200/latest/
- Compute nodes: DGX B200/B300 systems or DGX GB200/GB300 NVL72 racks, depending on deployment scale and cooling design [12]NVIDIA, "DGX SuperPOD Reference Architecture: GB200" (June 2025)https://docs.nvidia.com/dgx-superpod/reference-architecture-scalable-infrastructure-gb200/latest/ [10]NVIDIA Newsroom, "Blackwell Ultra DGX SuperPOD" (March 2025)https://nvidianews.nvidia.com/news/blackwell-ultra-dgx-superpod-supercomputer-ai-factories
- Networking: NVIDIA Quantum-X800 InfiniBand or Spectrum-X Ethernet, depending on the platform generation and design target [12]NVIDIA, "DGX SuperPOD Reference Architecture: GB200" (June 2025)https://docs.nvidia.com/dgx-superpod/reference-architecture-scalable-infrastructure-gb200/latest/ [10]NVIDIA Newsroom, "Blackwell Ultra DGX SuperPOD" (March 2025)https://nvidianews.nvidia.com/news/blackwell-ultra-dgx-superpod-supercomputer-ai-factories
- Operations layer: NVIDIA Mission Control for cluster operations, orchestration, and monitoring
- Scale: from enterprise deployments up to tens of thousands of GPUs [9]NVIDIA, "DGX Platform" (accessed March 2026)https://www.nvidia.com/en-us/data-center/dgx-platform/ [12]NVIDIA, "DGX SuperPOD Reference Architecture: GB200" (June 2025)https://docs.nvidia.com/dgx-superpod/reference-architecture-scalable-infrastructure-gb200/latest/
- Delivery: pre-built, cabled, and factory-tested (months compressed to weeks) [12]NVIDIA, "DGX SuperPOD Reference Architecture: GB200" (June 2025)https://docs.nvidia.com/dgx-superpod/reference-architecture-scalable-infrastructure-gb200/latest/
Customers include Shell, BMW, Sony, and Lockheed Martin. NVIDIA reports that 8 of the top 10 global telcos and 10 of the top 10 global car manufacturers run DGX-based infrastructure. [9]NVIDIA, "DGX Platform" (accessed March 2026)https://www.nvidia.com/en-us/data-center/dgx-platform/
DGX BasePOD is a simpler option: a reference architecture that combines DGX systems with third-party storage from validated partners, without the full NVIDIA networking stack. [9]NVIDIA, "DGX Platform" (accessed March 2026)https://www.nvidia.com/en-us/data-center/dgx-platform/
DGX Cloud
DGX Cloud is the cloud access path to the same ecosystem. It gives organizations access to DGX-class infrastructure through cloud providers instead of buying and racking hardware on-premises. [13]NVIDIA, "DGX Cloud" (accessed March 2026)https://www.nvidia.com/en-us/data-center/dgx-cloud/
As of early 2026, NVIDIA offers DGX Cloud through AWS, Google Cloud, Microsoft Azure, and Oracle Cloud. Each is a managed AI training platform with co-engineered NVIDIA compute clusters and access to NVIDIA engineers. [14]NVIDIA, "DGX Cloud on CSPs" (accessed March 2026)https://www.nvidia.com/en-us/data-center/dgx-cloud-on-csps/ Typical use cases:
- Burst capacity for large training runs without permanent hardware
- Avoiding upfront CapEx while accessing DGX-class performance
- Testing workloads before committing to on-premises deployment
What MGX is
Not every workload needs 8 SXM GPUs. Some need fewer GPUs, different CPUs, or PCIe cards instead of SXM. MGX is NVIDIA's reference architecture for those servers. [15]NVIDIA, "MGX" (accessed March 2026)https://www.nvidia.com/en-us/data-center/products/mgx/
MGX is a modular reference architecture. It is a set of standardized building blocks, including chassis, power, cooling, and connectors, that OEMs and ODMs (original design manufacturers, companies that design and build hardware on behalf of other brands) use to build accelerated servers faster and cheaper. [15]NVIDIA, "MGX" (accessed March 2026)https://www.nvidia.com/en-us/data-center/products/mgx/ Partners like Supermicro, QCT, ASUS, and GIGABYTE build and sell the actual servers. [16]NVIDIA Newsroom, "NVIDIA MGX Gives System Makers Modular Architecture to Meet Diverse Accelerated Computing Needs of World’s Data Centers" (May 2023)https://www.nvidia.com/en-sg/news/nvidia-mgx-server-specification/
The modular approach yields over 100 possible server configurations from a common set of components. [15]NVIDIA, "MGX" (accessed March 2026)https://www.nvidia.com/en-us/data-center/products/mgx/ It cuts OEM development costs by up to 75% and reduces time to market to about 6 months. [16]NVIDIA Newsroom, "NVIDIA MGX Gives System Makers Modular Architecture to Meet Diverse Accelerated Computing Needs of World’s Data Centers" (May 2023)https://www.nvidia.com/en-sg/news/nvidia-mgx-server-specification/
- CPUs: NVIDIA Grace (Arm), NVIDIA Vera (Arm), x86, and other Arm processors
- GPUs: Blackwell, Rubin, and other PCIe and rack-scale solutions
- Networking: NVIDIA Spectrum, Quantum InfiniBand, BlueField DPUs, ConnectX NICs
- Form factors: 1U, 2U, 4U with air or liquid cooling
- Standards: Open Compute Project (OCP) rack specifications
MGX is designed for multi-generational compatibility. A chassis designed for one generation of GPUs and CPUs can be reused across future generations with less redesign work than a one-off custom platform. [15]NVIDIA, "MGX" (accessed March 2026)https://www.nvidia.com/en-us/data-center/products/mgx/ This protects the OEM's engineering investment across hardware cycles.
NVIDIA announced MGX at Computex in May 2023, with the first systems from Supermicro and QCT shipping later that year. [16]NVIDIA Newsroom, "NVIDIA MGX Gives System Makers Modular Architecture to Meet Diverse Accelerated Computing Needs of World’s Data Centers" (May 2023)https://www.nvidia.com/en-sg/news/nvidia-mgx-server-specification/ The GB200 NVL4 (4 Blackwell GPUs + 2 Grace CPUs) is compatible with liquid-cooled MGX server designs. [11]NVIDIA, "GB200 NVL72" (accessed March 2026)https://www.nvidia.com/en-us/data-center/gb200-nvl72/
| HGX | MGX | |
|---|---|---|
| What it is | GPU baseboard (hardware) | Server reference architecture (design spec) |
| GPU type | SXM only | PCIe, SXM, and rack-scale (e.g., GB200 NVL72) |
| GPU count | 4 or 8 per board | Varies, from 1-GPU servers to 72-GPU racks |
| NVSwitch | Yes | Depends on config |
| CPU | Chosen by OEM | Specified in reference design |
| Target workload | Large-scale training | Training, inference, enterprise AI, HPC, edge |
| Form factors | Defined by baseboard size | 1U to full rack |
MGX is not a replacement for HGX. HGX targets 8-GPU training nodes with maximum interconnect bandwidth; MGX covers inference servers, mixed CPU-GPU nodes, and smaller deployments.
HGX platforms over time
NVIDIA has shipped HGX boards since 2017. Each generation brings more GPU memory and faster NVLink bandwidth. The current Blackwell boards (B200, B300) are 8-GPU only. Rubin, announced in 2026, is next. [1]NVIDIA, "HGX Platform" (accessed March 2026)https://www.nvidia.com/en-us/data-center/hgx/ [17]NVIDIA Newsroom, "NVIDIA Introduces HGX-2" (May 2018)https://nvidianews.nvidia.com/news/nvidia-introduces-hgx-2-fusing-hpc-and-ai-computing-into-unified-architecture-6696445
HGX platform timeline (click a row for details)
Which platform fits
The choice between HGX, DGX, MGX, and rack-scale systems depends on the buyer's scale, budget, and operational capacity. Hyperscalers at 10,000+ GPU scale bypass OEMs entirely, buying from NVIDIA or ODMs like Quanta and Foxconn. Neoclouds typically buy OEM servers built on HGX. DGX is a turnkey option with NVIDIA-validated software and direct support. [9]NVIDIA, "DGX Platform" (accessed March 2026)https://www.nvidia.com/en-us/data-center/dgx-platform/ MGX targets smaller or inference-focused deployments where fewer GPUs or different CPU architectures are needed. [15]NVIDIA, "MGX" (accessed March 2026)https://www.nvidia.com/en-us/data-center/products/mgx/
References
- NVIDIA, "HGX Platform" (accessed March 2026)
- NVIDIA, "Introducing NVIDIA HGX H100" Technical Blog (2022)
- NVIDIA, "HGX A100 Datasheet" (2020)
- NVIDIA, "H200 GPU" (accessed March 2026)
- Dell, "PowerEdge XE9680 Spec Sheet" (accessed March 2026)
- Supermicro, "SYS-821GE-TNHR Product Page" (accessed March 2026)
- Lenovo, "ThinkSystem SR675 V3 Technical Specifications" (accessed March 2026)
- NVIDIA, "DGX B200 User Guide" (2025)
- NVIDIA, "DGX Platform" (accessed March 2026)
- NVIDIA Newsroom, "Blackwell Ultra DGX SuperPOD" (March 2025)
- NVIDIA, "GB200 NVL72" (accessed March 2026)
- NVIDIA, "DGX SuperPOD Reference Architecture: GB200" (June 2025)
- NVIDIA, "DGX Cloud" (accessed March 2026)
- NVIDIA, "DGX Cloud on CSPs" (accessed March 2026)
- NVIDIA, "MGX" (accessed March 2026)
- NVIDIA Newsroom, "NVIDIA MGX Gives System Makers Modular Architecture to Meet Diverse Accelerated Computing Needs of World’s Data Centers" (May 2023)
- NVIDIA Newsroom, "NVIDIA Introduces HGX-2" (May 2018)
- PyTorch to Atoms, "Meta's 24K H100 Cluster: CapEx/TCO and BOM Analysis" (2024)
- Exxact, "TensorEX TS4-128275176-E 8U HGX B200 Server" (accessed March 2026)
Frequently Asked Questions
What is NVIDIA HGX?
HGX is NVIDIA's proprietary GPU baseboard bundling their SXM GPU form factor, NVLink high-bandwidth interconnect, and NVSwitch routing chips. The board holds 8 SXM GPUs. OEMs like Dell, Supermicro, and HPE build servers around HGX but cannot modify the baseboard itself; GPU count, memory configuration, NVLink wiring, and NVSwitch layout are all fixed by NVIDIA.
What is the difference between HGX and DGX?
DGX is NVIDIA's own complete server. It uses the same HGX baseboard as OEM servers but adds a fixed set of CPUs, memory, networking, storage, software, and support. DGX costs more than an equivalent OEM server because it includes NVIDIA's software stack and direct support. Teams that want lower cost per GPU, or already work with a specific OEM, buy OEM servers instead.
What is NVIDIA MGX used for?
MGX is a general modular reference architecture for buyers that mix different CPUs and NVIDIA GPUs. MGX is not a replacement for HGX. HGX targets 8-GPU training nodes with maximum interconnect bandwidth; MGX covers inference servers, mixed CPU-GPU nodes, and smaller deployments.
What is GB200 NVL72?
The GB200 NVL72 is a full liquid-cooled rack with 72 Blackwell GPUs and 36 Grace CPUs. Unlike HGX, which is a baseboard that goes inside a server, the NVL72 ships as an entire rack. The NVL72 requires liquid cooling and over 120 kW per rack; traditional air-cooled data centers cannot run it without retrofitting.
Coverage creates a minimum value for what your GPUs are worth at a future date. If they sell below the floor, the policy pays you the difference.
Learn how it works →