US What SXM and PCIe mean
SXM vs PCIe is a choice of form factor, not of GPU architecture. NVIDIA sells its top data centre GPUs, such as the A100, H100 and H200, in both forms, and the chip inside is the same. What changes is how the GPU is mounted, how much power it can draw and how it talks to the other GPUs in the server.
SXM is NVIDIA’s socketed module. The GPU sits on a small board that clips flat onto a baseboard, which supplies power and carries NVLink signals between the GPUs. NVIDIA does not publish an official expansion of the name, although it is often given as Server PCI Express Module. Generations have tracked the GPUs: SXM2 for Volta, SXM4 for Ampere and SXM5 for Hopper.
A PCIe GPU is a conventional add-in card that plugs into an x16 slot, with its own heat sink and power connector, like any other expansion card.
How the two compare
Using the H100 as the example, the differences look like this. The NVL column covers PCIe cards sold with an NVLink bridge, which sit between the two.
| SXM | PCIe card | PCIe NVL card | |
|---|---|---|---|
| Mounting | Module on HGX baseboard | Card in x16 slot | Card in x16 slot |
| GPUs per server | 4 or 8 on one board | 1–8 depending on chassis | Pairs, or groups of four on H200 NVL |
| GPU-to-GPU link | NVSwitch, all-to-all, 900 GB/s | PCIe 5.0, around 128 GB/s | NVLink bridge, 600–900 GB/s |
| Max power (H100) | Up to 700 W | 300–350 W | 350–400 W (H100 NVL) |
| Memory (H100) | 80 GB HBM3 | 80 GB HBM2e | 94 GB HBM3 |
| Cooling | Large heat sinks or liquid cold plates | Passive card, server fans | Passive card, server fans |
| Bought as | Complete HGX server only | Individual card | Individual card plus bridge |
| Best suited to | Large-scale training, biggest models | Inference, mixed workloads | Large-model inference, fine-tuning |
Because an SXM module can draw roughly twice the power, it runs at higher sustained clocks, and its memory is faster on most versions. On the H100, NVIDIA’s peak Tensor Core figures for SXM are around 30 per cent higher than for the PCIe card.
When SXM fits
SXM is built for jobs that need many GPUs to behave like one.
-
//
Training large models. Every GPU on an HGX board can exchange data with every other at full NVLink speed, so training scales cleanly across eight GPUs.
-
//
Very large models in inference. Models split across eight GPUs keep their per-token traffic on NVSwitch instead of PCIe.
-
//
Cluster building blocks. HGX servers pair the GPU board with fast network adapters, usually InfiniBand or RDMA Ethernet, to join many servers into one cluster.
-
//
Facilities to match. An eight-GPU HGX server can draw around 10 kW, so it needs high-density rack power and, increasingly, liquid cooling.
When PCIe fits
PCIe cards suit the much larger number of sites that run inference, analytics and graphics in standard rack servers.
-
//
Standard servers and racks. Cards fit air-cooled 2U and 4U GPU servers on normal rack power, and can be added or swapped individually.
-
//
Inference and per-GPU jobs. When each GPU serves its own model or streams, PCIe bandwidth between cards is rarely the limit.
-
//
Choice of card. PCIe covers the whole range, from the H100 PCIe card to lower-power L4 and L40S cards and RTX PRO Blackwell Server Edition GPUs.
- //
Upgrades, servicing and lifecycle
The two formats also age differently. A PCIe card can be removed and replaced on its own, and a server with enough slot power and airflow can often take a newer generation of card later, or a mix of cards for different jobs. Spares are simple to hold, because one card fits many servers.
SXM modules belong to their baseboard. Replacing a failed module is a specialist job, and moving to a new GPU generation usually means a new baseboard and often a new server, although NVIDIA designed HGX H200 boards to fit systems built for HGX H100. For long-running industrial and edge deployments, where the same server may stay in service for many years, the flexibility of PCIe cards is often as important as raw speed.
Quick answers on SXM vs PCIe
These are the questions people ask most when comparing the two formats. You can browse our data centre GPUs to see which cards come in PCIe and NVL form.
SXM and PCIe GPUs: common questions
What is SXM?
NVIDIA’s socketed GPU module. It mounts a data centre GPU flat on a server baseboard that supplies power and connects the GPUs with NVLink, instead of using a slot-in card.
Is SXM faster than PCIe?
For the same GPU, usually yes. SXM versions run at higher power and clocks and link GPUs through NVSwitch. For a single GPU running its own job, the gap is smaller than the power figures suggest.
Can an SXM GPU go in a PCIe slot?
Not as a supported option. SXM modules are sold on NVIDIA HGX baseboards inside complete servers. Third-party adapters exist, but they fall outside NVIDIA’s validated designs.
Is the H200 available as a PCIe card?
Yes, as the H200 NVL. It is a dual-slot PCIe card with 141 GB of HBM3e that links to one or three other cards through NVLink bridges.
Which is better for inference?
Usually PCIe. Most inference runs a model per GPU or per pair, which PCIe and NVL cards handle in standard air-cooled servers. SXM pays off for the very largest models and for training.
We build PCIe and NVL GPU servers as standard, with the cards, bridges, power and airflow matched to the chassis, and we can help you decide whether your workload needs NVLink at all.