DGX vs HGX comes down to who builds the system around the GPUs. HGX is NVIDIA's GPU baseboard platform, which NVIDIA describes as "a single baseboard with eight NVIDIA Rubin, NVIDIA Blackwell, or NVIDIA Blackwell Ultra SXMs" [2]. Server makers build their own servers around it. DGX is NVIDIA's own complete system. NVIDIA's Fabric Manager guide gives DGX B200/B300 and HGX B200/B300 8-GPU systems one GPU baseboard topology: two NVSwitch chips and eight GPUs [3].
The GPU side therefore matches. NVIDIA's DGX B200 and HGX B200 pages both list FP4 performance of 144 PFLOPS sparse and 72 PFLOPS dense, and 14.4 TB/s of total NVLink bandwidth [2, 7]. The rest of the machine differs. NVIDIA sets a DGX system's CPUs, networking, storage, software and support term, and offers one memory upgrade [7, 12]. For an HGX server, the server maker chooses each of those, and its price covers what it chose [28, 29, 30].
MGX is a modular reference architecture that manufacturers use to build servers and racks [18]. The names overlap: the DGX B300 sits in an "NVIDIA MGX-compliant, air-cooled chassis" [11].
What is NVIDIA HGX?
HGX began as a reference design. In May 2017 NVIDIA gave four original design manufacturers early access to "the NVIDIA HGX reference architecture", which it called "the same data center design used in" its DGX-1 [4].
Today the name covers the GPU board and its NVLink switches. NVIDIA's HGX table lists 2.1 TB of GPU memory and 1.6 TB/s of networking bandwidth for HGX B300, and 1.4 TB and 0.8 TB/s for HGX B200 [2]. For the Rubin generation, a footnote states that "CPU and memory specs are defined by OEM offerings" [2]. The host side belongs to the server maker.
Eight GPUs is the current form. Earlier boards also came with four. NVIDIA's H200 page offers HGX H200 systems "with 4 or 8 GPUs" [5], and its certified list includes four-GPU HGX H100 and H200 servers from Dell, Lenovo and Supermicro [22].
DGX vs HGX: what the full system adds
A DGX is a fixed configuration. NVIDIA's DGX B300 brief says the system ships with two terabytes of memory, which the customer can raise to four, and that "the rest of the components in the system are in a fixed configuration" [12]. An HGX server is a range of configurations that its maker offers.
DGX B200 vs HGX B200
The table sets NVIDIA's DGX B200 beside two HGX B200 servers from NVIDIA's certified list [22].
| Part | NVIDIA DGX B200 | Supermicro SYS-A21GE-NBRT | Dell PowerEdge XE9780 (B200) |
|---|---|---|---|
| Host CPUs | Two Xeon Platinum 8570, 112 cores in total | Two 4th or 5th Gen Xeon, up to 64 cores each | Two 6th Gen Xeon, up to 86 cores each |
| Host memory | 2 TB, configurable to 4 TB | 32 DIMM slots, up to 4 TB | 32 DIMM slots, up to 4 TB |
| Cluster networking | Eight single-port ConnectX-7 at up to 400 Gb/s, plus two BlueField-3 DPUs | Two 10GbE ports onboard, plus 8 low-profile and 2 FHHL PCIe slots | Up to 12 PCIe slots, at most 12 NICs and DPUs |
| Storage | 2 x 1.9 TB M.2 and 8 x 3.84 TB U.2 | 10 NVMe bays and 2 M.2 slots | Up to 16 E3.S or 10 U.2 drives |
| Power supplies | Six 3.3 kW, 5+1 redundant | Six 5,250 W, 3+3 redundant | Twelve 3,200 W |
| Software named | DGX OS included. AI Enterprise and Mission Control recommended | RHEL | Ubuntu (Dell lists RHEL for B300 only) |
| Support or warranty | Three years, hardware and software | No term on the product page | Three-year hardware warranty, no software support |
The B300 generation moves the cluster network ports. DGX B300 has eight 800 Gb/s ConnectX-8 ports on its GPU tray, where earlier DGX systems had them on the motherboard tray [12]. Dell's XE9780 with B300 has eight embedded OSFP ports, and Lenovo's SR680a V4 wires its eight OSFP ports to ConnectX-8 switches built into the GPU system board [29, 30]. B200 servers take network cards in PCIe slots [28, 29]. A B200 quote and a B300 quote can differ in network scope for that reason alone.
NVIDIA MGX vs HGX
NVIDIA drew the line itself when it launched MGX in May 2023. MGX "offers flexible, multi-generational compatibility with NVIDIA products" so that builders can reuse designs, while "HGX is based on an NVLink-connected, multi-GPU baseboard" [19]. With MGX, "manufacturers start with a basic system architecture optimized for accelerated computing for their server chassis, and then select their GPU, DPU and CPU" [19]. The current MGX page covers single-node servers to rack scale, with Arm or x86 CPUs [18].
MGX also names racks. DGX B300 is "deployable in NVIDIA MGX racks for the first time" [10], taking 54 V DC power from busbars in MGX-compliant racks or using its own AC power supplies [12]. DGX GB300, a liquid-cooled rack-scale system, is designed for MGX racks [15]. Rack power and liquid cooling covers the power side of that choice.
An MGX label describes a chassis or rack design. It says nothing about which GPUs are inside. The H200 page shows both paths: H200 SXM is offered in HGX systems, and the PCIe H200 NVL in MGX systems [5].
Does DGX include NVIDIA AI Enterprise?
It depends on the GPU generation. NVIDIA's licensing guide, last updated 2 September 2026, says AI Enterprise "is included in the NVIDIA DGX software bundle for NVIDIA DGX systems using the NVIDIA Hopper architecture" and that licenses "need to be purchased separately for NVIDIA DGX systems using the NVIDIA Blackwell architecture" [24]. It is licensed per GPU [24].
The DGX B200 and DGX B300 web pages mark DGX OS "Included" and AI Enterprise and Mission Control "Recommended" [7, 10]. Their datasheets list the same three products with no label [8, 11], so a datasheet alone does not answer the question. NVIDIA's DGX Resources page says AI Enterprise is "included in the DGX Software Bundle for DGX systems" without naming a generation, so the licensing guide's rule is the one to check [24, 25].
On the HGX side, NVIDIA's H200 page marks AI Enterprise as an "Add-on" for H200 SXM and "Included" for H200 NVL [5]. None of the three OEM B200 and B300 documents reviewed mentions it [28, 29, 30].
Support and warranty differ in scope
The DGX B200 and B300 datasheets give three years of business-standard support for hardware and software [8, 11]. At that level, severity 1 cases get an initial response within four business hours [27]. Dell's XE9780 carries a three-year hardware warranty "limited to hardware-related issues" that "does not include software support" [29]. Lenovo gives a three-year customer-replaceable-unit and onsite warranty, 9x5 next business day [30].
NVIDIA support entitlements "are non-transferable, are non-assignable and may be terminated when the product is transferred to another party" [27], so support on a used DGX may not pass to its buyer.
What NVIDIA-Certified Systems means
NVIDIA uses two levels. A qualified server has passed "thermal, mechanical, power, and signal integrity qualification" [21]. That is a prerequisite for certification, which adds functional, security and performance tests "including single-GPU, multi-GPU, and multi-node tests", along with network and storage offload, TPM and remote management checks [21]. NVIDIA states that more than 400 certified systems are available [21].
The Supermicro, Dell and Lenovo servers named here are on the list updated 30 October 2026 [22]. DGX B200 and B300 do not appear in its data center tables. The list names the GPU board and the network devices tested and gives no CPU or memory configuration [22]. Hardware support for certified systems comes from the system vendors [21].
Checklist: a DGX quote beside an HGX quote
- GPU board: generation, GPU count and GPU memory. NVIDIA gives B300 memory as 2.1 TB (270 GB per GPU) on its HGX page and datasheet, and as "8 x 288 GB = 2.3 TB total" in its DGX B300 user guide [2, 6, 13].
- Host CPUs and memory: model, cores, capacity and DIMM count. Dell relays NVIDIA design-guide minimums of more than 1 TB of host memory for B200 servers and more than 2 TB for B300 [29].
- Networking: card model, count, speed and where the cards attach. Dell's guide states: "For AI and HPC deployments, a 1:1 NIC-to-GPU ratio is recommended" [29].
- Software: which licenses the price includes, for how many GPUs and for how long [24, 25].
- Support: term, hardware or hardware and software, response times, and whether it survives a resale [27, 29, 30].
- Power and rack: AC power supplies or DC busbar, their count and rating, and the rack type [12, 13].
- Certification: the list entry and its reference configuration [22].
Dated HGX B200 and B300 listings, and what single line items add, are in what an HGX server price includes. Evidence classes in GPU pricing explains why a listing and a quote differ.
Rillor specifies, sources and delivers GPU systems, from a single 8-GPU server to full racks, and checks each part of the build against the workload and the site before it quotes. Each quote states the exact build, its condition and the tests run before delivery.
Read the full report
The full report adds the DGX and HGX pairings across four GPU generations, a five-system parts table for B200 and B300, the software and support terms document by document, and the numbered sources.