Skip to main content
BSS/OSS Academy

Architecture · Rack · Reference

DGX: from server to SuperPOD

DGX is NVIDIA’s own line of complete AI systems: GPUs, CPUs, networking and software sold as one product. It started in 2016 as a single server, and the unit NVIDIA sells has since grown to a whole liquid-cooled rack. BasePOD and SuperPOD are NVIDIA’s designs for joining many of those systems into one cluster.

Part 1

The DGX systems

Every DGX NVIDIA has announced, oldest first. Select one to see what was in it, what changed and where the figures come from. Years are announcement dates; shipping usually followed months later.

April 2016 · Server · Pascal

DGX-1

The first DGX. NVIDIA called it “the first system built with … Tesla P100 GPU accelerators, interconnected with NVIDIA NVLink”: a complete server for deep learning, sold with its software, where teams had been building their own GPU machines.

GPUs
8 × Tesla P100
GPU memory
16 GB per GPU, 128 GB in total (derived)
CPU
2 × Intel Xeon E5-2698 v4, 20 cores each
GPU links
NVLink in a “hybrid cube mesh”: each GPU links directly to some of the others, not all
Network
4 × EDR InfiniBand, 2 × 10 GbE
Power
3,200 W maximum in the launch datasheet; NVIDIA’s later user guide gives 3,500 W
Cooling
Air
Form factor
3U rack server

Sources

GPUs per unit: one unit is what NVIDIA sells as one system. Log scale.

Part 2

The pattern

The GPUs got faster in every generation, but the bigger change is what NVIDIA sells as one unit, and so what a data centre has to be ready to receive.

  1. 1 · 2016–2018

    A complete AI server

    DGX-1 sold deep learning as one box: 8 GPUs, CPUs, networking and software, tested together. DGX-2 doubled it to 16 GPUs on NVSwitch and 10 kW, and became the first SuperPOD.

  2. 2 · 2020 onward

    The 8-GPU server, every generation

    From DGX A100 the server settles at 8 GPUs and is refreshed with each GPU: H100, H200, B200, B300, then Rubin NVL8. Power per server climbs from 6.5 kW to about 24 kW, and by Rubin it is liquid-cooled.

  3. 3 · 2023 onward

    NVIDIA’s own CPU

    With Grace, NVIDIA designs the CPU as well as the GPU and joins them with NVLink-C2C. DGX GH200 was the first; the desk systems (Spark, Station) are built on the same idea.

  4. 4 · 2024 onward

    The rack is the product

    With NVL72, 72 GPUs share one NVLink domain in a liquid-cooled rack. The unit a data centre receives is a whole rack of 120 kW or more, so power and cooling become the first design question.

The desk systems (DGX Station and DGX Spark) run alongside the data centre line. They are for developing and testing on one machine, not a smaller version of a cluster.

Part 3

DGX and HGX: most GPU servers are not DGX

DGX

NVIDIA’s own complete system: the GPU board, CPUs, memory, storage, network cards and NVIDIA’s software, sold and supported by NVIDIA as one product.

HGX

The GPU baseboard on its own: typically eight GPUs joined by NVLink. NVIDIA supplies it to server makers, who build their own servers around it with their own CPUs, cooling and support.

  • Most GPU servers in data centres are HGX-based servers from other manufacturers, not DGX. Saying “DGX” for all of them is a common mistake.
  • DGX trades choice for a single tested design and one vendor to call. The cost is a higher price and less freedom over CPUs, storage and support.
  • At rack scale the same split applies: NVIDIA’s DGX GB200 and GB300 racks sit beside NVL72 racks built by server makers on NVIDIA’s design.

Part 4

DGX BasePOD and DGX SuperPOD

A POD is not a different machine. It is NVIDIA's reference design for joining many DGX systems into one cluster: the networks, storage, management and software around them, and how they are laid out.

What a POD adds around the DGX systems

DGX systems

The compute: DGX servers, or from Grace Blackwell onwards whole DGX racks. A POD fixes how many there are and how they are arranged.

Why it matters

Everything else in the POD is sized from this number. Add systems and the networks, storage and power all have to grow with them.

DGX BasePOD compared with DGX SuperPOD
DGX BasePODDGX SuperPOD
What NVIDIA calls it“A prescriptive AI infrastructure for enterprises”, introduced in 2022 as “the evolution of DGX POD”.“The product that NVIDIA provides to the market … a complete turnkey solution with specific BOM, Installation services, support services, and guaranteed performance.”
Who shapes it“Primarily driven by NVIDIA-certified storage vendors”. NVIDIA provides the reference architecture; storage partners can customise it.NVIDIA, to a fixed bill of materials. NVIDIA’s FAQ says NVIDIA Professional Services installation is required.
SizeNVIDIA’s current reference architecture: “scaling from two nodes to dozens”, with 2 to 8 compute nodes in its reference configurations.Built from scalable units. Current designs go to thousands of GPUs (see below).
Compute fabricTwo NVIDIA Quantum QM9700 NDR 400 Gb/s InfiniBand switches in the reference configurations.A full rail-optimised fabric sized to the number of scalable units.
StoragePartner storage on an Ethernet fabric shared with management. NVIDIA’s page names nine storage partners.Certified storage on its own fabric, specified in the reference architecture.
Typical useAn enterprise’s first AI cluster, or a team-sized one.Large-scale training; NVIDIA builds its own internal clusters this way.
Trade-offMore choice of storage and integrator, but more of the integration risk stays with the buyer and partner.Less choice, but a tested design and one vendor answerable for performance.

Scalable units: how a SuperPOD grows

A SuperPOD is built from identical blocks called scalable units (SUs). Each SU has its own systems and its own section of the network, so the cluster grows by adding blocks rather than by redesigning it. The size of a block depends on the generation.

Year
2026
Scalable unit
“Each SU contains 8 DGX Vera Rubin NVL72 systems” in the InfiniBand design: 576 GPUs. The Spectrum-X Ethernet design uses 3 racks per SU.
Largest
“Up to 128 rack configurations with 9216 GPUs”; the same document also describes 18 SUs (10,368 GPUs) on a two-layer fabric.
Compute fabric
Quantum-X800 (800 Gb/s) InfiniBand, or Spectrum-X Ethernet in the second design; NVLink 6 inside each rack
Storage
Spectrum-4 Ethernet, shared with in-band management
Software
Mission Control, Slurm, Run:ai

NVIDIA gives about 2 MW per SU, up from 1.2 MW for GB200 and GB300. Announced in January 2026 for the second half of the year, with a DGX Rubin NVL8 server option alongside.

NVIDIA’s figures disagree: 9,216 GPUs and 18 SUs in one document, and a 14-rack example in the launch blog that fits neither SU size. Quoted, not reconciled.

Sources

1

1 × 8 racks = 8 racks, 576 GPUs

GPU counts only. A real SuperPOD also has management, storage and network racks, and NVIDIA sets the largest size it supports.

How the PODs evolved

  1. 2018–19 DGX POD. NVIDIA’s first reference designs for a rack of DGX systems with partner storage. The 2019 SuperPOD was “built using the NVIDIA DGX POD reference architecture”.

  2. 2019 The first DGX SuperPOD. Built from DGX-2 and DGX-2H systems on EDR InfiniBand. NVIDIA says its first DGX-2H cluster, 36 systems, went “from concept to results in just three weeks”.

  3. 2020 SuperPOD becomes a product. With DGX A100, NVIDIA sells SuperPODs through certified partners, built from scalable units of 20 systems.

  4. 2022 DGX BasePOD replaces DGX POD. The smaller, partner-led design gets its own name, alongside the DGX H100 SuperPOD with 32-system units.

  5. 2023 DGX GH200. 256 Grace Hopper superchips joined by NVLink into one system. NVIDIA did not publish it as a SuperPOD design.

  6. 2024 The rack becomes the block. The GB200 SuperPOD is built from 8-rack units, about 1.2 MW each. 8-GPU server designs (B200) carry on alongside.

  7. 2025 GB300 and Mission Control. Blackwell Ultra racks with an 800 Gb/s compute fabric; Mission Control becomes NVIDIA’s operations software for the cluster.

  8. 2026 Vera Rubin. About 2 MW per 8-rack unit, and an Ethernet (Spectrum-X) design alongside InfiniBand.

Sources

Part 5

Where to go next