PCIe, SXM, HGX, DGX and NVL72: NVIDIA GPU form factors explained
Short answer: the same NVIDIA GPU can come in very different packages, and the package decides how fast GPUs talk to each other, how much power they draw and what kind of data centre they need. PCIe is a card that plugs into an ordinary server. SXM is a module that sits on an NVIDIA baseboard and links GPUs over NVLink. HGX is that baseboard with eight SXM GPUs, sold to server makers. DGX is NVIDIA's own complete server built around the same board. NVL72 is a whole rack in which 72 GPUs share one NVLink fabric.
Below, each one in plain language, then a table of which form each GPU comes in, from the H100 to Vera Rubin.
PCIe cards: GPUs for ordinary servers
A PCIe GPU is an add-in card, much like a graphics card in a desktop, that plugs into a standard server slot. It is the most flexible form: almost any server can take one, most are air-cooled, and you can put one, two or four in a machine.
The catch is the connection between cards. PCIe Gen 5, the fastest slot in today's servers, moves about 128 GB/s between a card and the rest of the system. That is fine when each job fits on one GPU, such as inference on a mid-size model, video work or rendering. It becomes the bottleneck when several GPUs have to share one large model, which is why training clusters do not use PCIe cards.
Examples: the RTX PRO 6000 Blackwell Server Edition (96 GB) is PCIe only, and the H100 and H200 also come as "NVL" PCIe cards (more on that name below).
SXM modules: GPUs built to work together
An SXM GPU is not a card. It is a module that mounts directly onto an NVIDIA baseboard, which lets it draw far more power (700 W for the H100 and H200, around 1,000 W and more for Blackwell) and connect to its neighbours over NVLink, NVIDIA's direct GPU-to-GPU link.
NVLink is what matters. On the H100 and H200 it runs at 900 GB/s per GPU, about seven times PCIe Gen 5. On Blackwell (B200 and B300) it runs at 1.8 TB/s per GPU, about fourteen times. With an NVLink Switch on the board, every GPU can reach every other at that speed, so eight GPUs behave like one large accelerator. That is why serious training and large-model serving run on SXM. Our NVLink, NVSwitch and InfiniBand guide explains the fabric in detail.
HGX: the eight-GPU board inside most AI servers
HGX is NVIDIA's baseboard carrying eight SXM GPUs joined by NVLink Switch chips. NVIDIA sells it to server makers, who build the rest of the server around it: the CPUs, memory, storage, networking, chassis and cooling. Most 8-GPU AI servers you can rent or buy, whoever's name is on the front, have an HGX board inside.
Current boards, from NVIDIA's specifications:
- HGX B300: 8 Blackwell Ultra GPUs, 288 GB each, 1.8 TB/s NVLink per GPU and 14.4 TB/s in total.
- HGX B200: 8 Blackwell GPUs, 1.4 TB in total, which is 180 GB per GPU (the B200 chip carries 192 GB, of which 180 GB is usable in these systems), with the same NVLink speeds.
- HGX H200 and H100: 8 Hopper GPUs at 141 GB and 80 GB each, 900 GB/s NVLink per GPU.
A naming quirk: at its March 2025 conference NVIDIA called the B300 board "HGX B300 NVL16", because each Blackwell GPU package contains two compute dies and NVIDIA briefly counted dies rather than packages. It is still eight GPUs as you and your software see them.
DGX: NVIDIA's own complete server
DGX is NVIDIA's own branded system, built around the same eight-GPU board. The GPUs and NVLink are the same as in an HGX server; what DGX adds is NVIDIA's complete, validated build and its software and support. For example, the DGX B200 has 8 Blackwell GPUs with 1,440 GB of GPU memory in total, two Intel Xeon processors, NVIDIA ConnectX networking and BlueField DPUs, draws about 14.3 kW at most, and comes with DGX OS and three years of NVIDIA support, with NVIDIA AI Enterprise and Mission Control software recommended. The DGX B300 draws about 14 kW.
So the practical difference is not speed. It is who built and supports the box, the software stack, and the price, which NVIDIA does not publish. Related names:
- DGX SuperPOD: NVIDIA's reference design for a whole AI supercomputer built from many DGX systems, with storage, networking and management.
- DGX Cloud: NVIDIA-managed GPU clusters sold through the large public clouds.
- DGX Spark and DGX Station: NVIDIA's desktop AI computers for developers. Despite the name, they are not data-centre servers.
NVL: two different things with the same three letters
NVL on a card means a PCIe card that can be bridged to its neighbours with NVLink. The H100 NVL has 94 GB and a 600 GB/s NVLink bridge. The H200 NVL has 141 GB, draws up to 600 W, and links two or four cards at 900 GB/s per GPU. It is a middle ground: faster than plain PCIe between the bridged cards, but no NVLink Switch, so not the all-to-all fabric of an SXM server.
NVL with a number means a rack-scale NVLink domain: the number is how many GPUs share one NVLink fabric. The GB200 NVL72 and GB300 NVL72 each put 72 Blackwell GPUs and 36 Grace CPUs in a single liquid-cooled rack, with 130 TB/s of NVLink bandwidth across all 72. Such a rack draws in the region of 120 to 140 kW, far beyond what most data-centre halls were built for. NVIDIA's next generation, Vera Rubin, ships as Vera Rubin NVL72: it was first announced as "NVL144" while NVIDIA was counting dies, then renamed back to counting packages. Our GB200 and NVL72 explainer goes deeper.
Superchips: the "G" in GB200 and GB300
A Grace superchip puts NVIDIA's own Arm-based Grace CPU on the same module as the GPUs, joined by a fast link called NVLink-C2C. The GH200 pairs one Grace with one Hopper GPU; the GB200 pairs one Grace with two Blackwell GPUs; the GB300 one Grace with two Blackwell Ultra GPUs. In practice, GB200 and GB300 are bought as NVL72 racks, not as single servers.
MGX: the modular design behind many other servers
MGX is NVIDIA's modular reference design that lets server makers build many different machines, from single nodes to whole racks, from shared building blocks. NVIDIA's RTX PRO Servers, which carry the RTX PRO 6000 Server Edition, are one example.
Which form does each GPU come in?
GPU | Forms | Memory per GPU | Link between GPUs | Typical system |
|---|---|---|---|---|
SXM; NVL PCIe card | 80 GB (SXM); 94 GB (NVL) | NVLink 900 GB/s (SXM); 600 GB/s bridge (NVL) | 8-GPU HGX H100 or DGX H100 | |
SXM; NVL PCIe card | 141 GB | NVLink 900 GB/s; NVL bridges 2 or 4 cards | 8-GPU HGX H200 or DGX H200 | |
SXM | 180 GB in HGX and DGX (192 GB on the chip) | NVLink 1.8 TB/s | 8-GPU HGX B200 or DGX B200 | |
SXM | 288 GB | NVLink 1.8 TB/s | 8-GPU HGX B300 or DGX B300 | |
GB200 | Grace superchip (1 CPU + 2 GPUs) | About 186 GB | NVLink 1.8 TB/s, 72 GPUs in one domain | GB200 NVL72 rack, liquid-cooled |
GB300 | Grace superchip (1 CPU + 2 GPUs) | Up to 288 GB | NVLink 1.8 TB/s, 72 GPUs in one domain | GB300 NVL72 rack, liquid-cooled |
PCIe (Server Edition) | 96 GB GDDR7 | PCIe Gen 5 only, no NVLink | RTX PRO Servers (MGX) | |
Rack (NVL72); 8-GPU HGX Rubin NVL8 | 288 GB HBM4 | Sixth-generation NVLink | Vera Rubin NVL72 rack; 8-GPU servers |
Figures are NVIDIA's published specifications as of October 2026. Memory is HBM unless stated. GB200 memory per GPU is derived from NVIDIA's 372 GB per superchip.
Which one should you rent?
For most AI teams the answer is an 8-GPU SXM server on an HGX board. It gives you the full NVLink fabric for training and large-model serving, it runs in ordinary high-density data-centre space, and it scales out to clusters over InfiniBand. That is exactly what FlexiCloud rents: dedicated 8-GPU H100, H200, B200 and B300 SXM servers, single-tenant bare metal, in India, the EU and the US. If you need NVIDIA's DGX software and support stack specifically, tell us when you ask for a quote.
PCIe cards make sense when each job fits on one GPU, such as inference, video and graphics, which is where the RTX PRO 6000 fits. NVL72 racks are built for the largest frontier-model training and serving, by teams with the liquid-cooled, very high-power data-centre space to match. We do not rent NVL72 racks.
See all GPU servers and prices, or tell us what you are running and we will match the GPU and the server to the job.