HGX is NVIDIA's eight-GPU baseboard that server makers build into their own systems, DGX is NVIDIA's own branded, fully supported system built around that kind of board, and NVL72 is a different shape entirely: a liquid-cooled rack that links 72 GPUs in one NVLink domain. The three names describe a board, a product and a rack, so they are not strict alternatives to each other.
TL;DR
- HGX is a baseboard: NVIDIA says it comes with "eight NVIDIA Rubin, NVIDIA Blackwell, or NVIDIA Blackwell Ultra SXMs." CPU and memory specs "are defined by OEM offerings."
- DGX is NVIDIA's own complete system, sold with NVIDIA software, Mission Control and three years of business-standard support.
- NVL72 is a rack of 72 GPUs and 36 Grace CPUs (GB200 or GB300). DGX GB200 is a DGX-branded version of that rack.
- Pick HGX for flexibility and the broadest range of vendors, DGX for a turnkey NVIDIA-supported system, NVL72 when one model has to span far more than 8 GPUs.
- Verdict: for most teams the practical unit is an 8-GPU HGX or DGX node. Rack-scale is for the largest models and needs a liquid-cooled facility.
The three names in one table
| HGX | DGX | NVL72 | |
|---|---|---|---|
| What it is | GPU baseboard (8 SXM GPUs) built into OEM servers | NVIDIA-branded complete system | Rack of 72 GPUs and 36 Grace CPUs |
| Who sells it | OEM and ODM partners | NVIDIA | NVIDIA (as DGX GB200) and OEMs |
| GPU count | 8 | 8 per node (B200, B300) or 72 per rack (GB200) | 72 |
| NVLink domain | 8 GPUs | 8 GPUs, or 72 for DGX GB200 | 72 GPUs |
| CPUs and memory | Chosen by the OEM | Fixed by NVIDIA (dual Intel Xeon in DGX B200 and B300) | Grace CPUs, 17 TB LPDDR5X per rack |
| Cooling | Depends on the OEM system | DGX B200 and B300 are air-cooled servers per the pages we read | Liquid-cooled |
| Support | From the OEM | Three years of business-standard hardware and software support | From NVIDIA or the OEM |
The DGX row says "air-cooled servers per the pages we read" because NVIDIA's DGX B200 and DGX B300 pages list an operating temperature range and a 10U server form factor, and do not describe liquid cooling for them.
HGX: the baseboard
NVIDIA's HGX page describes a single baseboard carrying eight SXM GPUs connected by NVLink and NVSwitch. Server vendors take that board, add their own CPUs, memory, storage, networking and chassis, and sell the result under their own name. NVIDIA's page notes that "CPU and memory specs are defined by OEM offerings" and points to a Qualified System Catalog and NVIDIA-Certified Systems for where to buy.
NVIDIA's published HGX figures on that page:
| Spec | HGX B300 | HGX B200 |
|---|---|---|
| GPUs | 8 Blackwell Ultra SXM | 8 Blackwell SXM |
| Total GPU memory | 2.1 TB | 1.4 TB |
| FP4 Tensor Core | 144 PFLOPS sparse, 108 dense | 144 PFLOPS sparse, 72 dense |
| FP8/FP6 Tensor Core | 72 PFLOPS (sparse) | 72 PFLOPS (sparse) |
| NVLink per GPU | 1.8 TB/s | 1.8 TB/s |
| Total NVLink | 14.4 TB/s | 14.4 TB/s |
| Network bandwidth | 1.6 TB/s | 0.8 TB/s |
The page lists HGX B300 and HGX B200 as "shipping now," and also shows HGX Rubin NVL8 and HGX Vera Rubin NVL8 variants. It does not list Hopper baseboards; for the Hopper GPUs themselves see H200 vs B200 vs GB200.
Because the board is the same, two HGX B200 servers from different vendors share GPU performance but can differ in CPU, system memory, NIC choice, cooling and price. That is the main reason HGX exists: a common GPU platform with competition around it.
For the GPU pages, see /gpu/nvidia-b200, and the B200 guide for the chip itself.
DGX: NVIDIA's own system
DGX is NVIDIA's complete, supported product. NVIDIA describes the DGX platform as "built from the ground up for enterprise AI," and DGX SuperPOD as infrastructure "configurable with any NVIDIA DGX system."
Two examples from NVIDIA's pages:
| Spec | DGX B200 | DGX B300 |
|---|---|---|
| GPUs | 8 Blackwell | 8 Blackwell Ultra SXM |
| GPU memory | 1,440 GB | 2.1 TB |
| CPUs | 2x Intel Xeon Platinum 8570 (112 cores) | Intel Xeon 6776P |
| System memory | 2 TB, configurable to 4 TB | not listed on the page we read |
| Power | about 14.3 kW max | about 14 kW |
| Rack units | 10U | 10U |
| Networking | 8x ConnectX-7 at up to 400 Gb/s, 2 BlueField-3 DPUs | 8x ConnectX-8 at up to 800 Gb/s, 2 BlueField-3 DPUs |
NVIDIA says DGX B200 offers "3X the training performance and 15X the inference performance of previous-generation systems" (compared with DGX H100; a vendor projection), and says DGX B300 raises dense FP4 performance by 1.5x and attention performance by 2x over DGX B200.
DGX also includes the software layer: Mission Control, NVIDIA AI Enterprise and DGX OS, plus three years of business-standard support. If you want one vendor to call for hardware and software, that is what DGX sells.
NVIDIA also uses the DGX name for small systems: DGX Spark (GB10 Superchip, 128 GB unified memory, for developer desktops) and DGX Station (GB300 Grace Blackwell Ultra Desktop Superchip, 748 GB coherent memory). See DGX Station GB300. Those are desk-side machines, not data-center nodes.
NVL72: the rack
NVL72 breaks the eight-GPU limit. GB200 NVL72 and GB300 NVL72 put 72 Blackwell or Blackwell Ultra GPUs and 36 Grace CPUs in one liquid-cooled rack where all GPUs share 130 TB/s of NVLink bandwidth.
- NVIDIA's technical blog says each rack has 18 compute trays with four GPUs and two Grace CPUs apiece, and nine NVLink switch trays.
- NVIDIA lists 13.4 TB of HBM3E for GB200 NVL72 and 20 TB for GB300 NVL72.
- NVIDIA's product pages do not state rack power. Supermicro lists GB200 NVL72 at 132 kW, with in-rack coolant distribution units.
- NVIDIA also sells it as DGX GB200, which has the same 36-Superchip, 72-GPU rack layout, with DGX software and support. NVIDIA's DGX GB200 page links to the NVL72 page but does not use the NVL72 name.
Read the full details in the GB200 NVL72 guide and GB300 vs GB200.
Where MGX fits
You may also see MGX. NVIDIA calls it "an open modular reference architecture" for OEMs, ODMs and ecosystem partners to build accelerated servers from single nodes to rack-scale designs. It is a design framework, not a product you buy. NVIDIA's MGX page does not say how it relates to HGX or NVL72, so we do not claim a relationship beyond that.
Power and cooling compared
| System | Power (source) | Cooling |
|---|---|---|
| DGX B200 | about 14.3 kW max (NVIDIA) | 10 to 35 degrees C operating range |
| DGX B300 | about 14 kW (NVIDIA) | not detailed on the page |
| GB200 NVL72 | 132 kW (Supermicro listing) | direct liquid cooling with a coolant distribution unit |
| HGX servers | set by the OEM | set by the OEM |
An eight-GPU node at about 14 kW is a different world from a 130 kW rack. That ten-fold difference is why most teams start with nodes. For networking between nodes, read NVLink vs InfiniBand and what is NVLink.
Which to choose
- You want the widest vendor choice and price competition: HGX systems from OEMs.
- You want NVIDIA to be accountable for hardware and software together: DGX.
- Your model needs a 72-GPU NVLink domain: NVL72, in DGX GB200 form or from an OEM, with a facility that can cool it.
- You only need to run workloads, not own hardware: rent the GPUs instead. Most teams reach 8-GPU Blackwell and Hopper nodes this way. The datacenter GPU guide maps the whole range, and the GPU pages are /gpu/nvidia-b200 and /gpu/nvidia-gb200.
Cost
We do not type prices here. Purchase prices for HGX, DGX and NVL72 systems are set by sellers and not published by NVIDIA, and rental prices change daily, so use the live box below and multiply by your GPU count and hours. The cost comparison that matters is cost per token or per training run on your workload, not the sticker price of the system.
Rent today
Aquanode manages and optimizes GPUs for training and inference workloads, and you can rent the GPUs below on demand. The box shows live availability; "None right now" means there is no offer at the moment.
What's next
NVIDIA's HGX page already lists HGX Rubin NVL8 and HGX Vera Rubin NVL8, and its Vera Rubin NVL72 page lists a 72-GPU rack with 216 TB/s of NVLink bandwidth (NVIDIA's Rubin technical blog says 260 TB/s, and its HGX page lists 3.6 TB/s of NVLink per GPU). NVIDIA describes these as ramping into production. See Vera Rubin NVL72.
FAQ
Is HGX the same as DGX?
No. HGX is a baseboard that partners build into servers. DGX is NVIDIA's complete branded system with NVIDIA software and support.
Is DGX GB200 the same as GB200 NVL72?
They share the same 36-Superchip, 72-GPU rack layout. DGX GB200 is NVIDIA's branded, supported version; NVL72 is the rack architecture that OEMs also build.
What does NVL72 mean?
The 72 is the number of GPUs in one NVLink domain: 72 Blackwell GPUs across 18 compute trays, with 36 Grace CPUs.
Does HGX B200 include CPUs?
NVIDIA's HGX page describes a GPU baseboard and says CPU and memory specs are defined by OEM offerings.
How much power does a DGX B200 use?
NVIDIA lists about 14.3 kW maximum for DGX B200 in 10 rack units.
Sources
- NVIDIA HGX platform page: https://www.nvidia.com/en-us/data-center/hgx/
- NVIDIA DGX platform page: https://www.nvidia.com/en-us/data-center/dgx-platform/
- NVIDIA DGX B200 page: https://www.nvidia.com/en-us/data-center/dgx-b200/
- NVIDIA DGX B300 page: https://www.nvidia.com/en-us/data-center/dgx-b300/
- NVIDIA DGX GB200 page: https://www.nvidia.com/en-us/data-center/dgx-gb200/
- NVIDIA GB200 NVL72 page: https://www.nvidia.com/en-us/data-center/gb200-nvl72/
- NVIDIA GB300 NVL72 page: https://www.nvidia.com/en-us/data-center/gb300-nvl72/
- NVIDIA technical blog, GB200 NVL72: https://developer.nvidia.com/blog/nvidia-gb200-nvl72-delivers-trillion-parameter-llm-training-and-real-time-inference/
- NVIDIA MGX page: https://www.nvidia.com/en-us/data-center/products/mgx/
- NVIDIA Vera Rubin NVL72 page: https://www.nvidia.com/en-us/data-center/vera-rubin-nvl72/
- Supermicro SRS-GB200-NVL72 rack listing: https://www.supermicro.com/en/products/rack/srs-gb200-nvl72
